RAMP: Toward Ubiquitous Quantum-Safe Key Distribution as a Service

RAMP: Toward Ubiquitous Quantum-Safe Key Distribution as a Service

arXiv:2610.22879v1 [cs.CR] 16 Oct 2026

Jiafan Wang CSIRO Technology

Sebastian Kish University of Technology Sydney

Abstract

Dongxi Liu CSIRO Technology

key establishment. Post-quantum cryptography (PQC) [4] provides a software-based migration path1 for replacing publickey primitives vulnerable to quantum attacks, while quantum key distribution (QKD) [3] offers a fundamentally distinct approach of secret generation based on quantum physics [50]. Appropriately parameterized symmetric cryptography [33] also remains a practical building block against quantum adversaries once shared secret state has been established. The deployment of PQC does not eliminate the value of QKD. PQC remains computational, i.e., its security relies on the hardness of mathematical problems against classical and quantum adversaries. Instead, QKD derives its secrecy from the properties of quantum communication under a specified physical model. The two paradigms thus provide complementary trust foundations for quantum-safe key establishment. This distinction has motivated substantial investment in QKD infrastructure and high-assurance communication systems2 . However, its deployment model remains difficult to scale. QKD requires specialized optical hardware, calibrated quantum links, and protocol-specific processing at dedicated nodes. Thus, it is more desirable to make this scarce infrastructure accessible to ordinary endpoints via a service-oriented model rather than simply deploying more QKD devices.

Quantum key distribution (QKD) offers a security foundation independent of computational hardness assumptions. However, its reliance on specialized hardware limits access by ordinary users. Turning QKD into a service poses a fundamental trust challenge, as the providers that supply QKD-derived material themselves could retain knowledge of the keying inputs of users. This makes end-to-end confidentiality solely depend on the trustworthiness of the service infrastructure. We present RAMP, a service-oriented framework that decouples QKD material generation from key consumption. Providers operate QKD infrastructure and supply raw-key materials, while users securely acquire, mask, and locally process them into final keys without operating QKD devices. RAMP employs reconciliation-preserving masking compatible with both DV- and CV-QKD. It mitigates the trust dependence by separating the QKD-service and user-side masking domains. We formalize complementary confidentiality guarantees: RAMP protects the final key even if colluding providers expose their complete raw-key views, provided the masking domain remains secure; conversely, exposing the complete masking state does not break final-key secrecy as long as the QKD source and acquisition paths remain protected. The results establish the dual-domain security boundary of RAMP. We implement RAMP using experimentally acquired QKD data and standardized quantum-safe cryptography. After acquiring service-provisioned keying material, commodity endpoints complete user-side QKD post-processing with an average key-distillation throughput of 2.1Kb/s on an M4 Pro CPU. To our knowledge, we demonstrate the first experimental realization of such service-provisioned, user-side QKD processing, presenting a viable path toward QKD as a service.

1.1

Challenges in More Accessible QKD

Existing QKD systems [3, 20, 39, 51] primarily operate in a point-to-point manner. The two-party key establishment requires dedicated optical links and a pair of QKD devices. Both parties participate in quantum-state transmission to generate raw keys, and then perform classical post-processing (including sifting, parameter estimation, information reconciliation, and privacy amplification) to distill the final secret key 1 https://www.nccoe.nist.gov/applied-cryptography/migrat

1

Introduction

ion-to-pqc 2 For example, secure quantum communication backbone across the European Union (https://digital-strategy.ec.europa.eu/en/policies /european- quantum- communication- infrastructure- euroqci) and quantum information science as a strategic national priority in US (https://www.quantum.gov/).

Large-scale quantum computers would compromise widely deployed public-key mechanisms such as RSA and Diffie– Hellman [44], motivating the transition toward quantum-safe 1

Figure 1: The RAMP Framework (Blue Arrow for Quantum Channel)

1.2

from raw keys. It is physically and economically infeasible to deploy specialized hardware at numerous endpoints.

From QKD to Quantum-Safe Service

We answer the question affirmatively with a service-oriented framework called RAMP. RAMP enables users to obtain QKDderived keys without owning specialized devices, while limiting the trust placed in the service providers that supply the underlying QKD materials. RAMP denotes the user-side operational cycle, i.e., Register–Acquire–Mask–Process. RAMP users obtain and leverage the entropy from QKD by interacting with their service providers. We strategically separate the quantum and the classical stages of QKD: the service providers handle the transmission of quantum states and the maintenance of fresh raw-key materials, while the users perform the cryptographic protection and QKD postprocessing locally using general-purpose computing hardware. This separation allows a shared QKD backend to supply raw-key materials to multiple users without requiring QKD hardware at the user endpoints, while decoupling continuous QKD generation from on-demand key consumption.

Device-Centric QKD Networks. The proposals of QKD network [11,29] aim to scale point-to-point QKD by multiplexing quantum channels and coordinating multi-hop relaying for broader coverage. National initiatives, such as SECOQC [36], SwissQuantum [47], and Tokyo QKD Network [43], showcase the viability of linking multiple QKD nodes. While the advancement shows the potential to extend QKD networks beyond laboratory environments, these systems remain resource-intensive, imposing large overhead. Each node must host modules for quantum state transmission, maintain stable quantum channels, and handle post-processing. Note that the model of QKD network is inherently device-centric. It interconnects a pre-configured set of nodes, instead of offering on-demand key acquisition for regular users. Trust Issues in QKD as a Service. Recent works [9, 10, 23] attempt to make QKD a service that is available to multiple users. Existing designs mainly focus on resource and service provisioning over deployed infrastructures of QKD networks, including secret-key assignment, network orchestration, and the integration of QKD-derived keys into classical protocols. These efforts improve the operational flexibility of QKD networks, but their service boundary typically remains trusted: the infrastructure that provisions a QKD-derived key may learn or reconstruct the key delivered to the user. Consequently, compromise or collusion within the service infrastructure can directly devastate end-to-end key confidentiality. We comprehensively review related work in Section B. The literature leaves open a fundamental question:

Technical Challenges and Solutions. The design of RAMP addresses several challenges in integrating QKD with standardized quantum-safe building blocks, while preserving correctness, effectiveness, and security in a service-oriented model. • Complementary Compromise Resilience. Moving QKD behind a service boundary creates an unusual security problem, as the providers by definition know the complete QKD material supplied to the users for secret derivation. RAMP therefore introduces an independent user-side masking domain, such that provider compromise is tolerated through the masking domain, while exposure of the masking state falls back to the secrecy of the protected QKD source and acquisition paths.

Can QKD-derived keying material be delivered as a convenient service to users without QKD hardware, while avoiding sole reliance on the secrecy and trustworthiness of the service infrastructure for endto-end final-key confidentiality?

• Reconciliation-Preserving Masking. As QKD raw-key materials are inherently noisy, conventional encryption cannot simply be inserted into the processing pipeline without considering its effect on reconciliation. RAMP 2

• Dual-Domain Resilient Key Establishment. We design RAMP around two cryptographically separated security domains, namely, the QKD service domain and the user-side post-quantum masking domain. The raw-key acquisition is protected by independently provisioned credentials, while a pair of users establishes the masking secret through authenticated post-quantum key establishment. RAMP thus prevents a sole assumption from protecting both the QKD-derived material and its mask.

applies customized relabeling to raw keys via a masking mechanism. This transformation preserves the underlying discrepancy or likelihood structure of general QKD modes, allowing standard QKD post-processing to operate directly on the masked representation. As a byproduct, RAMP retains the underlying QKD soundness for detecting unilateral inconsistent or low-quality material. • Service-Oriented Key Acquisition. The independently stored raw-key materials should be bound to the intended user session without requiring specific coordination between the service providers, in order to prevent unintended reuse. RAMP therefore deploys user-mediated acquisition, enforcing identifier binding and raw-key lifecycle management before masking and post-processing.

• Reconciliation-Preserving Post-Processing Pipeline. By employing masking with no error propagation, RAMP supports the standard post-processing of general QKD modes and preserves its inherent error tolerance. RAMP further utilizes this property to enable the detectability of inconsistent materials from any one of the providers.

• Interoperability with Standardized Suites. Compatibility with standardized implementations of QKD and PQC poses a substantial challenge to practical deployment. RAMP is designed for generic QKD protocols. The construction incorporates standard cryptographic interfaces, which can be instantiated with well-developed post-quantum and symmetric primitives.

• Security under Complementary Compromise Modes. We formalize two complementary security notions for RAMP. We prove final-key confidentiality against colluding service providers that expose their complete raw-key views, assuming the user-side post-quantum masking domain remains secure; meanwhile, we show that exposing the masking secret does not invalidate final-key secrecy when the underlying QKD execution and acquisition path remains uncompromised. These results establish the dual-domain compromise resilience of RAMP.

Distinction against Key Combiner. RAMP is not designed as a replacement for generic hybrid key combiners [7, 13], which combine independently protected secrets at final-key derivation. These works, when applied to QKD-as-a-service, could let the service providers complete QKD post-processing, deliver a QKD key, and combine it at the users with independently established pre- or post-quantum keys as the final key. RAMP deliberately advances the service boundary from final-key delivery to key processing. After receiving the material, the users themselves perform parameter estimation, reconciliation, and privacy amplification. Our reconciliationpreserving masking makes this user-side QKD processing possible without allowing provider knowledge of the supplied material to determine the final key. RAMP therefore proposes a distinct trust–cost tradeoff by removing security-critical processing from the provider-side trusted execution path.

Organization. Section 2 provides preliminaries on QKD and involved cryptographic primitives. Section 3 presents our system model, including the interfaces, threat model, and security goals. We construct RAMP in Section 4 and formalize its security in Section 5. Section 6 reports our experimental evaluation of RAMP and Section 7 concludes this work.

1.3

2

• Evaluation with Real-World QKD. We implement the RAMP processing pipeline on realistic user devices, using experimentally acquired QKD data and standardized quantum-safe cryptographic primitives. We demonstrate practical secret-key throughput with modest RAMPspecific cost over the underlying QKD processing.

Contributions

Our contributions span system design, service-oriented scheme, dedicated security analysis, and empirical evaluation.

2.1

Preliminaries Quantum Key Distribution (QKD)

QKD [3] enables two distant parties, conventionally referred to as Alice and Bob, to establish a shared secret key using quantum communication. The secret is encoded into quantum states, and QKD security relies on fundamental properties of quantum mechanics, e.g., the no-cloning theorem [50]. Any attempt by an adversary to intercept or measure the transmitted quantum states, regardless of its computational power, inevitably introduces disturbances, which can be detected by

• QKD-Derived Key Distribution as a Service. We propose RAMP, a service-oriented framework that decouples QKD material generation from key consumption. Specialized providers operate the quantum infrastructure and continuously produce correlated raw-key materials, while ordinary users acquire and locally process these materials on demand without operating QKD devices. 3

the legitimate parties. As a result, QKD provides informationtheoretic security of the generated keys [41]. We divide the QKD procedure into two stages: the quantum stage, where quantum states are exchanged and measured, and the classical stage, where post-processing operations are performed over an authenticated classical channel.

apply universal hash functions to compress the reconciled key into a shorter, uniformly random secret key. The final key length depends on the estimated leakage during reconciliation and the security parameters. Alice and Bob end with an identical shared secret key that is provably secure under the chosen QKD protocol. Composable QKD Secrecy. We adopt the standard composable secrecy notion of QKD [37, 49]. Let K ∈ {0, 1}ℓ denote the final key produced by an accepted QKD execution, and let E denote the adversary’s complete side-information system allowed by the instantiated QKD security model. In particular, E may contain quantum information retained from the quantum-channel interaction together with the public classical post-processing transcript. Let ρ̄acc KE denote the normalized classical–quantum state of K and E conditioned on protocol acceptance. Let ρ̄acc E = trK ρ̄acc , where tr denotes the partial trace over K, i.e., ρ̄acc K KE E is the marginal state of E obtained by discarding K. The ideal key state is τK = 2−ℓ ∑ |k⟩⟨k|,

Quantum Stage. This stage consists of preparation, transmission, and measurement of quantum states over a quantum channel. Depending on how information is encoded and measured, QKD protocols are categorized as discretevariable QKD (DV-QKD) [3] and continuous-variable QKD (CV-QKD) [39]. DV-QKD encodes with discrete quantum degrees of freedom (e.g., photon polarization), while CV-QKD uses continuous modulation of optical field quadratures. At the end of the quantum stage, Alice and Bob obtain correlated classical data derived from their respective preparations and measurements. These data, referred to as raw keys, are generally noisy and only partially correlated due to channel loss, noise, and potential eavesdropping. In addition to the raw keys, the devices record metadata such as detection efficiency, detector electronic noise, shot-noise calibration, timing statistics, and other implementation parameters used for subsequent parameter estimation and security evaluation.

k∈{0,1}ℓ

which represents a uniformly distributed ℓ-bit key. For quantum states ρ and σ, we write

Classical Stage. After quantum transmission, Alice and Bob perform classical post-processing over an authenticated channel to distill a shared secret key. Although specific implementations differ between DV-QKD and CV-QKD, the classical stage generally consists of the following steps:

1 D(ρ, σ) = ∥ρ − σ∥1 2 for their trace distance. When the systems are entirely classical, D reduces to the usual total-variation (statistical) distance. Let pacc denote the probability that the QKD execution accepts. The protocol provides εqkd -secrecy if, for every attack covered by the instantiated QKD security model,

• Data Filtering (Sifting): Alice and Bob publicly exchange protocol-dependent information (e.g., basis choices in DV-QKD or measurement/configuration information in CV-QKD) to identify and retain subsets of data that are sufficiently correlated for key generation. This step may be implicit or combined with other processing.

acc pacc D(ρ̄acc KE , τK ⊗ ρ̄E ) ≤ εqkd .

Operationally, this bounds the distinguishability between the real QKD key and an ideal uniform key that is independent of all adversarial side information.

• Parameter Estimation: A subset of the data is disclosed to estimate channel characteristics and noise levels. DVQKD typically quantifies it by the quantum bit error rate (QBER), while CV-QKD involves estimating quantities such as transmission, excess noise, and signal-to-noise ratio (SNR). If the estimated parameters exceed security thresholds, the protocol is aborted.

2.2

Cryptographic Building Blocks

We review the syntax of cryptographic building blocks used in RAMP. Section C recalls their security notions and discusses their quantum resistance, respectively. 2.2.1

• Information Reconciliation: Due to noise and imperfections, Alice and Bob’s raw keys are not identical. Error correction techniques are applied to align their data. Interactive protocols such as Cascade [6] are commonly used in DV-QKD, while forward error correction using low-density parity-check (LDPC) codes [14] is widely used in CV-QKD and high-rate systems.

Digital Signature Scheme (SIG)

A digital signature scheme [22] enables a party to authenticate messages and prove their origin. A signer generates a signature on a message using a secret signing key, and anyone could use the corresponding public key to verify the authenticity and integrity of the message. Definition 1 (Digital Signature). A digital signature scheme SIG = (KeyGen, Sign, Vf) with message space M and signature space S is defined as follows:

• Privacy Amplification: To remove any partial information potentially available to an adversary, Alice and Bob 4

• (pk, sk) ← KeyGen(1λ ): The algorithm takes as input a security parameter 1λ . It outputs a secret signing key sk and its corresponding public key pk.

Definition 3 (XOR-Based Synchronous Stream Cipher). An XOR-based synchronous stream cipher SSC = (KeyGen, Keystream, Enc, Dec) with key space K is defined as follows:

• σ ← Sign(sk, m): The signing algorithm outputs a signature σ ∈ S with a secret key sk and a message m ∈ M .

• k ← KeyGen(1λ ): The key generation algorithm inputs a security parameter 1λ and outputs a secret key k ∈ K .

• b ← Vf(pk, m, σ): The verification algorithm takes as input a public key pk, a message m ∈ M , and a signature σ ∈ S . It outputs b ∈ {0, 1}, where b = 1 indicates σ is a valid signature on m; b = 0, otherwise. 2.2.2

• z ← Keystream(k, n): The keystream generation algorithm takes as input a secret key k and an output length n. It outputs a n-bit keystream z ∈ {0, 1}n .

Authenticated Key Exchange (AKE)

• c ← Enc(m, z): The encryption algorithm takes as input a message m ∈ {0, 1}n and a keystream z ∈ {0, 1}n . It outputs the ciphertext c = m ⊕ z, where ⊕ denotes the bit-wise XOR operation.

An authenticated key-exchange protocol [2, 8] allows two parties to establish a shared session key while authenticating the intended communicating peer. Definition 2 (Authenticated Key Exchange). An authenticated key-exchange protocol AKE is a two-party, sessionbased protocol with key space K executed over an adversarially controlled network. A local session at party U with intended peer U ′ is identified by π = (U, U′ , sid, role), where role ∈ {init, resp} and sid is a session identifier. Upon successful completion on two matching sessions π = (U, U′ , sid, init) and π′ = (U′ , U, sid, resp), AKE(U, U′ , sid) outputs a session key k ∈ K ; otherwise, it outputs ⊥. 2.2.3

• m ← Dec(c, z): The decryption algorithm takes as input a ciphertext c ∈ {0, 1}n and the keystream z ∈ {0, 1}n . It outputs the plaintext m = c ⊕ z.

3

We aim to build a service-oriented system that enables end users or applications to establish secret keys from QKDderived entropy without deploying or operating specialized QKD hardware locally. Instead, users interact with their service providers that maintain the QKD infrastructure and continuously generate raw-key materials. The design follows a four-step workflow for any user:

Secure Channel from Provisioned Credential

A secure channel (SC) protects communication between two parties over an adversarially controlled network by providing confidentiality and integrity of the transmitted messages [24]. A secure channel from provisioned credentials lets the communicating parties establish a long-term symmetric credential via a prior authenticated enrollment or provisioning phase. The provisioned credential serves as the trust root for subsequent protected sessions. In particular, the online channel protection does not require an additional public-key trust assumption. The provisioned credential is subsequently used to authenticate the peers and derive session-specific traffic keys for protected communication. This model is supported by standardized protocols. In particular, TLS 1.3 enables such a channel based on symmetric pre-shared keys (PSKs) [19, 42]. 2.2.4

System Design

• Register with a local service provider for the key distribution service and establish the necessary credentials; • Acquire raw keys from the service provider, which are continuously generated through QKD devices; • Mask the acquired raw keys to protect them from provider-side exposure and enhance security; and • Process the masked raw keys through post-processing to derive the final distilled secret key.

Synchronous stream ciphers (SSC)

We assume that the service providers manage their own physical QKD devices. These devices feature the basic functionality of regular point-to-point QKD protocols (either DVor CV-QKD) when being initialized and paired. As Fig. 1 depicts, our description focuses on the scenario where two users subscribe to the service from their local providers. The interfaces naturally admit deployments with multiple providers and subscribers, which is not evaluated in this work due to the resource limit.

Synchronous stream ciphers (SSC) [30] generate the keystream independently of the plaintext and ciphertext, using only a secret key (and possibly an initialization vector). Encryption (resp. Decryption) simply executes bit-wise XOR between the plaintext (resp. ciphertext) and the keystream. It implies that any transmission errors in the ciphertext (resp. plaintext) affect only the corresponding symbols in the plaintext (resp. ciphertext) during operations. This property is also referred to as no error propagation. 5

3.1

Interfaces

eavesdropper by the security model of QKD. In particular, A may intercept, delay, replay, drop, or modify classical messages exchanged among service providers and users, while interacting with the quantum channel as considered in QKD. We expand the threat model to cover two forms of compromise that affect distinct parts of the QKD service system.

Notations. We denote the set of all service providers and the set of potential users/subscribers by P and U , respectively. R K is a collection of raw-key materials and M D records the corresponding metadata. Both R K and M D are indexed by a set of key identifiers K I D . For example, we can obtain a raw key rk from its collection via rk ← R K [kid] with its key identifier kid ∈ K I D . Without loss of generality, we use P, P′ ∈ P to denote two service providers that serve two users U, U′ ∈ U , respectively. We denote the holder of some data by its subscript, e.g., K I D P is the key identifiers by P, and omit it when it is clear. We define the interfaces of our system as below, each of which invokes a protocol by specified parties.

• Service-provider corruption. A may corrupt one or both service providers. Upon corruption, it obtains the provider’s state associated with the target session, including raw-key materials, metadata, and acquisition credentials. A may deviate the provider from the prescribed protocol. This captures both curious providers that expose their local views and actively misbehaving providers that return manipulated or low-quality materials. • Masking-state exposure. A may obtain the user-side masking state of the target session, including the cryptographic session secret, the derived masking key, and the resulting masking keystream. This models compromise of the masking mechanism, such as exposure of its session secret or failure of its underlying public-key assumptions, either post-quantum or classical ones.

• Init(P ): All service providers from P jointly initialize the system. They build connections among pairs of QKD devices and confirm the specification. Each party sets up and announces public parameters (to potential users), e.g., information on certificates, public keys, available service, and the data format. • RKGen(P, P′ ): It invokes continuous QKD sessions between service providers P, P′ ∈ P for raw-key generation. These materials are stored as R K P and R K P′ , indexed by the corresponding key identifiers, and attached with the necessary metadata M D P and M D P′ , by both parties, respectively.

Users are considered honest, as our target is to establish secret keys for them; the keys are otherwise trivially exposed. Security Conditions. We determine three security conditions that are associated with a target session of our system. • Masking security (M). The secret state used to generate the user-side masking key remains uncompromised. Concretely, A does not obtain the masking secret, any derived masking keys, or the resulting keystream, and the involved cryptographic mechanisms remain secure.

• Regist(U, P): User U subscribes to the service from the local provider P. It provisions an independent symmetric acquisition credential kU,P , which is used exclusively to protect the raw-key acquisition between U and P. • Acq(U, P, kU,P , K I D P , R K P , M D P ): User U acquires raw keys from the service provider P where U has registered. The returned raw key is denoted as rk.

• QKD-service secrecy (Q). The underlying QKD execution satisfies its assumptions of secrecy and authenticity, and the provider-side state containing the target QKD raw-key materials is not exposed to A . Q captures the authenticity of the user-handled QKD-related information and the secrecy of the QKD-derived source.

• Mask(U, U′ , rk, rk′ ): Two users U, U′ ∈ U input rk and rk′ , respectively. rk and rk′ are obtained via Acq() from P and P′ with the same kid, but may be different from each other due to channel noise or misbehavior. U and U′ mask rk and rk′ , respectively, using user-derived keystreams. The masked raw keys are denoted as rkmask and rk′mask , respectively.

• Acquisition security (A). The acquisition paths for both users preserve the confidentiality and integrity of the rawkey materials and their metadata, and the corresponding acquisition credentials remain uncompromised.

• Proc(U, U′ , rkmask , rk′mask ): Two users U, U′ ∈ U conduct post-processing on rkmask and rk′mask , yielding a final secret key kfinal shared only between them.

3.2

The conditions M, Q, and A do not assume that compromise events are statistically independent in a real deployment. They intend to identify distinct cryptographic assets and trust roots whose exposure has different consequences. Specifically, the acquisition credentials can be provisioned independently for each user–provider relationship and are not derived from the user-to-user state. The user-side masking keystream is not derived from provider-side QKD or acquisition state. A complete compromise of user endpoints may

Threat Model

Adversary Capabilities. We consider a probabilistic or quantum polynomial-time adversary A that may control the classical network and exercise the capabilities permitted to an 6

4.1

expose several of these assets simultaneously, which is outside the complementary-compromise guarantees in this work. Security goals. We target final-key confidentiality whenever either the masking condition remains secure or both the QKDservice and acquisition conditions remain secure, namely,

System Setup and Raw-Key Generation

Init(). All service providers from P jointly bootstrap the global environment by registering service-provider identities, agreeing on parameters, setting up cryptographic suites, and establishing connections (provisioning physical QKD links).

Conf = M ∨ (Q ∧ A). 1. Storage Initialization. Each service provider initializes its local storage, including the collection of raw keys R K , metadata M D , and their index set K I D . It also maintains a collection ST to record the state of each raw-key block. Depending on the usage of the material regarding kid ∈ K I D , ST [kid] ∈

Upon protocol acceptance, we capture dual-domain resilience and inconsistency detectability with the following properties. • Provider-Compromise Final-Key Confidentiality. Even if both service providers are corrupted and A obtains their complete views of the target session, the final key remains computationally indistinguishable from uniform, provided the user-side masking condition M holds.

{AVAILABLE, PENDING, CONSUMED, EXPIRED}. The state of a pending raw-key block will eventually become CONSUMED or EXPIRED. Once a raw-key block has been consumed or conservatively expired, RAMP never makes it available to any user session.

• Masking-Compromise Final-Key Confidentiality. Even if A obtains the complete masking state of the target session, the final key remains secret, provided that the QKD-service secrecy condition Q and the independent acquisition security condition A both hold.

2. Digital Signature Setup. Each service provider executes SIG.KeyGen(1λ ) to obtain its own private signing key and public verification key of SIG.

• Detectability of Inconsistent Key Material. If one service provider supplies material that is inconsistent or otherwise inadmissible under the QKD post-processing procedure, RAMP preserves the underlying QKD bound on the probability that such material would be accepted.

3. QKD Connection. Each service provider publishes a QKD capabilities descriptor, including the supported QKD mode (DV/CV), the structure of raw-key materials, the maximum tolerated error, and the metadata format.

Section 5 formalizes these properties for our construction. We assume that paired QKD devices could produce correlated raw keys with a statistical error rate that is observable via parameter estimation. The authenticity regarding (the public keys of) any party can be established via standard PKI of the TLS protocol [5, 45]. These assumptions concern operational requirements, orthogonal to the new challenges of our system.

4

4. Announcement. The system publishes identities and public keys of all service providers, along with the routing data, as a reference for subscription. RKGen(). Assume that P, P′ ∈ P operate a paired QKD link. 1. QKD Execution (Quantum Stage). P and P′ continuously execute the quantum stage of the selected QKD protocol to obtain correlated raw keys. Both providers assign each pair of correlated raw-key blocks a common identifier kid by hashing their own identities and a unique block identifier from the underlying QKD execution.

Construction

RAMP follows the four-step workflow, i.e., Register–Acquire– Mask–Process. As per the interfaces defined in Section 3.1, we separate system initialization Init(), continuous raw-key generation by providers RKGen(), user registration Regist(), raw-key acquisition from providers Acq(), keystream-based masking Mask(), and post-processing between users Proc(). RAMP utilizes the following cryptographic functionalities. A digital signature scheme SIG authenticates provideroriginated service information. A secure acquisition channel between each user and its local provider is established via a provisioned symmetric credential. Users establish a session secret through an authenticated key exchange protocol AKE. The synchronous stream cipher SSC expands this secret into a masking keystream. The acquisition credential and AKE secret are generated independently. The key sizes and/or output lengths of involved primitives follow the suggestions of NIST [33] for security against quantum adversaries.

2. Storage Update. P denotes the raw-key material and the metadata regarding kid by rkP and md P , respectively. P updates R K P [kid] as rkP , M D P [kid] as md P , and ST P [kid] as AVAILABLE. P′ updates on rkP′ and md P′ . Fig. 2 shows the system initialization and raw-key generation with P = {P, P′ }. The service providers could continuously generate and store raw-key material via RKGen().

4.2

Registration and Raw-Key Acquisition

Regist(). Enroll user U ∈ U with a local service provider P ∈ P and establish a credential used exclusively by them. 7

6. Upon receiving (kid, σP,kid ), P′ verifies σP,kid with the public key pkP of P, and then checks whether ST P′ [kid] =AVAILABLE. If either check fails, the acquisition aborts. Otherwise, P′ updates ST P′ [kid] as PENDING and retrieves from its local storage

Service Provider P'

Service Provider P Init()

𝒦ℐ𝒟P′ ← ∅ ℛ𝒦P′ , 𝒮𝒯P′ , ℳ𝒟P′ ← empty 𝑝𝑘P′ , 𝑠𝑘P′ ← SIG. KeyGen()

𝒦ℐ𝒟P ← ∅ ℛ𝒦P , 𝒮𝒯P , ℳ𝒟P ← empty 𝑝𝑘P , 𝑠𝑘P ← SIG. KeyGen() Publishes QKD Capabilities

Establish QKD Links

rkP′ := R K P′ [kid],

Announce Identities, Public Key, and Connection Info.

md P′ := M D P′ [kid].

7. P′ → U′ : P′ returns (rkP′ , md P′ ) to U′ .

RKGen()

Before masking, U and U′ respectively validate (rkP , md P ) and (rkP′ , md P′ ) against the QKD profile agreed during initialization. Invalid or unsupported metadata causes abort.

Quantum Stage of QKD for Raw Key 𝑟𝑘 Derive 𝑘𝑖𝑑 for 𝑟𝑘p ℛ𝒦P 𝑘𝑖𝑑 ≔ 𝑟𝑘p ℳ𝒟P 𝑘𝑖𝑑 ≔ 𝑚𝑑p 𝒮𝒯P [𝑘𝑖𝑑] ≔ AVAILABLE

Derive 𝑘𝑖𝑑 for 𝑟𝑘P′ ℛ𝒦P′ 𝑘𝑖𝑑 ≔ 𝑟𝑘P′ ℳ𝒟P′ 𝑘𝑖𝑑 ≔ 𝑚𝑑P′ 𝒮𝒯P′ [𝑘𝑖𝑑] ≔ AVAILABLE

4.3

Figure 2: System Setup and Raw-Key Generation

Mask(). U and U′ now hold the raw-key materials (rkP , mdP ) and (rkP′ , mdP′ ), respectively, associated with the same identifier kid. They then derive a common secret mask and apply it to the protocol-specific reconciliation inputs. No communication with P and P′ would be required during this phase.

More specifically, U and P provision an independent symU,P metric acquisition credential kacq ←$ {0, 1}λ through an authenticated enrollment mechanism, which could be conducted out-of-band via provisioned or device-bound credentials [12]. U,P U and P utilize kacq to instantiate a secure communication channel for subsequent raw-key acquisition between them.

1. Session Binding. U and U′ determine the session identifier sid by hashing (U,U ′ , P, P′ , kid). sid is used throughout the remainder of the user-side execution, binding the AKE session and subsequent QKD processing.

Acq(). Consider that U, served by P, intends to obtain a QKDderived key with U′ , served by P′ . The high-level idea of Acq() is to let U acquire rkP from local P and forward kid of rkP to U′ , such that rkP′ regarding kid from P′ can be retrieved. The following design aligns with standard QKD interfaces formalized by ETSI [15]. After raw-key acquisition, U and U′ hold correlated raw-key materials rkP and rkP′ indexed by the same kid, respectively.

2. Pre-Mask QKD Preparation. The representation delivered by the QKD devices is protocol-dependent. Before masking, U and U′ perform the public QKD operations that must precede reconciliation, including the applicable sifting and parameter-estimation steps. U and U′ authenticate the classical messages required by QKD. Any samples disclosed for parameter estimation are removed.

1. U → P: U requests raw-key material from P through the acquisition channel between them and identifies (U′ , P′ ) as the intended peer user and its service provider.

Data Filtering (Sifting). U and U′ obtain metadata (e.g., basis information in DV-QKD or measurement/configuration data in CV-QKD) from md P and md P′ , respectively. U and U′ retain only the subset of data that is expected to be correlated.

2. When receiving a valid request from a registered user, P picks a key identifier kid ← K I D such that ST P [kid] =AVAILABLE. P retrieves the corresponding raw key and its associated metadata as rkP := R K P [kid],

Raw-Key Masking

Parameter Estimation. U and U′ disclose publicly a random subset of filtered rkP and rkP′ to estimate channel characteristics (e.g., QBER in DV-QKD or SNR in CVQKD). The step aborts if the result exceeds protocoldefined thresholds, indicating excessive noise or interference. Since parameter estimation is performed only after both raw-key blocks have been acquired by the users, a unilateral provider cannot subsequently adapt its supplied block to the other user’s disclosed samples.

md P := M D P [kid].

P updates ST P [kid] as PENDING to avoid race conditions. To bind the selected identifier to the intended user and provider pair, P produces a signature as
σP,kid ← SIG.Sign skP , Acq ∥ U ∥ U′ ∥ P ∥ P′ ∥ kid .

We denote the remaining portions on U (resp. U′ ) after the above two operations by x (resp. x′ ). (x, x′ ) is the input for subsequent reconciliation based on the original acquired (rkP , rkP′ ). For DV-QKD, x and x′ are binary strings retained for reconciliation. For CV-QKD,

3. P → U: P returns (rkP , md P , kid, σP,kid ) to U. 4. U → U′ : U stores (rkP , md P ) and sends U′ (kid, σP,kid ). 5. U′ → P′ : U′ submits (kid, σP,kid ) to P′ . 8

4.4

the preparation includes the protocol-specific conversion needed to obtain the binary reference string and the associated soft information used by reconciliation.

Proc(). This stage performs QKD post-processing on the masked raw keys rkmask and rk′mask to eliminate discrepancies and distill the final key kfinal for U and U′ .

3. Masking-Key Establishment. After the reconciliation inputs (x, x′ ) have been fixed, U and U′ establish a session-specific masking key through

1. Information Reconciliation. Due to noise and imperfections, rkmask is not identical to rk′mask . Users apply error correction (e.g., Cascade [6] in DV-QKD or LDPC [14] in CV-QKD) to align them with minimal leakage. Users conduct correctness verification on reconciled binary strings with the verification procedure of the instantiated QKD. A verification failure causes the session to abort.

kAKE ← AKE(U, U′ , sid). Masking proceeds only if the AKE execution accepts for the intended peer and the target sid [8]. 4. Keystream Derivation. Let n denote the length of the reconciliation input. U and U′ respectively derive an n-bit masking keystream bound to the RAMP session as

2. Privacy Amplification. Users compress the reconciled key with a universal hash function to remove information that is potentially leaked. The final key is kfinal ∈ {0, 1}ℓ .

z ← SSC.Keystream(kAKE , n). 5. Masking. The operation depends on the QKD modes. DV-QKD. The reconciliation inputs retained by U and U ′ after the previous steps are binary strings x, x′ ∈ {0, 1}n .

The output length ℓ needs to satisfy both confidentiality branches analyzed in Section 5. In particular, it shall not exceed the secure output length justified by the underlying QKD security analysis, and it also needs to account for all key-dependent public leakage.

Given a keystream z of length n, both users mask by rkmask = x ⊕ z,

rk′mask = x′ ⊕ z.

It preserves the discrepancy between the paired inputs: ′

′

After confirmation, U (resp. U′ ) instructs P (resp. P′ ) to update ST P [kid] (resp. ST P′ [kid]) as CONSUMED. If no instruction is received after a predefined period, P (resp. P′ ) updates ST P [kid] (resp. ST P′ [kid]) as EXPIRED on its own.

′

rkmask ⊕ rkmask = (x ⊕ z) ⊕ (x ⊕ z) = x ⊕ x . Therefore, the masked results (rkmask , rk′mask ) neither create new bit errors nor change the locations of existing errors. It allows the masked strings to be processed by the same reconciliation procedure as the original inputs. The masking in DV-QKD can also be directly modeled as an encryption of XOR-Based SSC as rkmask ← SSC.Enc(x, z),

Post-Processing and Key Distillation

Correctness and Robustness. The only modification on raw keys in RAMP is Mask(), which is realized with SSC that incurs no error propagation. Consequently, any discrepancies, due to either channel noise or active interference, remain localized. It ensures that the standard QKD post-processing can be conducted identically on the masked raw keys. As long as the underlying raw keys satisfy the correlation requirements assumed by QKD (i.e., the observed error rate remains below the prescribed threshold), both users derive identical final keys correctly. The property underpins robustness, since the localized errors remain effectively correctable within the tolerance of QKD post-processing. Therefore, RAMP preserves both correctness and robustness under the same conditions required by conventional QKD.

rk′mask ← SSC.Enc(x′ , z)

CV-QKD. In a typical reverse-reconciliation pipeline, U′ holds a binary reconciliation string as x′ ∈ {0, 1}n , while U holds the corresponding log-likelihood ratios (LLRs) as x = (L1 , . . . , Ln ) for these bits. RAMP masks this bit-level reconciliation input as rk′mask = x′ ⊕ z and ei = (−1)zi Li , ∀i ∈ {1, · · · , n}. L Specifically, when zi = 0, both the bit label and its LLR remain unchanged. When zi = 1, the bit label is flipped and the corresponding LLR sign is reversed. Its magnitude, and therefore its reliability, remains unchanged. We denote these LLRs as rkmask .

Key-Allocation Integrity. The provider signature binds the allocation identifier kid to the intended user–provider context, while the provider-side state collection ST prevents reuse of terminal identifiers. Under EUF-CMA security and correct state enforcement, a network adversary cannot forge a fresh authorized allocation or reactivate a consumed identifier. This mechanism authenticates the allocation context rather than the contents or physical provenance of the corresponding raw-key material. Content integrity during acquisition is supplied by the secure channel for an honest provider endpoint.

(rkmask , rk′mask ) thus represents the same soft-decision decoding problem with a relabeling of binary alphabet. Thus, for both DV- and CV-QKD, RAMP hides the effective reconciliation inputs from service providers while preserving the structure required by subsequent QKD operations. 9

5

Security Formalization

Theorem 1 asserts provider-compromise confidentiality of RAMP, with its proof and quantitative bound in Section D.1. No QKD secrecy or acquisition security is assumed, as the providers may reveal the unmasked QKD-derived input.

We formalize the security goals introduced in Section 3.2. Our analysis considers a target RAMP session between two honest users. The adversary controls the classical network and has the quantum-channel capabilities allowed by QKD. We target confidentiality with dual-domain resilience, which is captured by the security conditions Conf = M ∨ (Q ∧ A). Accordingly, we analyze two complementary compromise modes. One considers that both service providers are corrupted while the user-side masking state remains secure. The other allows exposure of the masking state while QKDservice secrecy and independent acquisition remain secure.

Theorem 1 (Provider-Compromise Confidentiality). Assuming that AKE provides session-key security, SSC is IND-CPAsecure, and the privacy amplification output length accounts for all key-dependent leakage released after masking, RAMP provides provider-compromise final-key confidentiality. Theorem 1 shows that exposing the provider-visible QKDderived inputs does not compromise the final RAMP key. The complementary value of QKD will be utilized in Theorem 2, where the complete masking state may be exposed while confidentiality is retained from the QKD execution. Hybrid key combiners [7,13] provide a related source-failure property. RAMP differs in the placement of the QKD service boundary, where users themselves execute security-critical processing.

5.1 Provider-Compromise Key Confidentiality We first consider compromise of the complete QKD-service side. Both service providers may be corrupted and their complete views of the target session may be exposed. In this case, the confidentiality relies on condition M, rather than Q or A.

5.2 Masking-Compromise Key Confidentiality

Definition 4 (Provider-Compromise Key Confidentiality). Let Π be a RAMP scheme. PCΠ,A (1λ ) proceeds as follows.

We now consider the complementary case in which the complete masking state is exposed. In this experiment, we make no assumption on AKE or SSC, but require that Q and A hold.

1. The challenger initializes RAMP and sets a target session between honest users U and U′ served by P and P′ .

Definition 5 (Masking-Compromise Key Confidentiality). Let Π be a RAMP scheme. MCΠ,A (1λ ) proceeds as follows.

2. A corrupts (P, P′ ) and obtains their complete targetsession state, including the QKD raw-key materials (rkP , rkP′ ), metadata (md P , md P′ ), acquisition credentials, and provider-side protocol state. The corrupted providers may deviate arbitrarily from the protocol.

1. The challenger initializes RAMP and sets a target session between honest users U and U′ served by P and P′ . 2. The target QKD execution satisfies Q. Namely, the underlying QKD secrecy and authenticity assumptions hold, while the provider-side state containing the target rawkey materials is not exposed to A . Whereas A could obtain the quantum and classical side information allowed by QKD.

3. The honest users execute RAMP. The masking-security condition M is required to hold. In particular, the target AKE session remains unexposed and A does not obtain kAKE or the resulting keystream z.

3. Both provider-to-user acquisition paths satisfy A, so the target raw-key materials and metadata are delivered confidentially and without adversarial modification. A may still control the public classical network.

$

4. The challenger samples b ← {0, 1}. If the target session aborts, the challenger returns ⊥, independently of b. Otherwise, it returns ( kfinal , b = 0, Kb = $ ℓ R ← {0, 1} , b = 1.

4. The challenger reveals the complete masking state of the target session to A , including kAKE and z. $

5. The challenger samples b ← {0, 1}. If the target session aborts, the challenger returns ⊥, independently of b. Otherwise, it returns ( kfinal , b = 0, Kb = $ ℓ R ← {0, 1} , b = 1.

together with all public messages of the target session. 5. The adversary outputs a guess b′ . We say Π provides provider-compromise final-key confidentiality if for any probabilistic or quantum polynomial-time adversary, the following advantage is negligible in λ:

together with all public messages of the target session. 1 AdvΠ,A (1 ) = Pr[b = b] − . 2 PC

λ

′

6. A outputs a guess b′ . 10

6

We say Π provides masking-compromise final-key confidentiality if for any probabilistic or quantum polynomial-time adversary, the following advantage is negligible in λ: λ ′ AdvMC Π,A (1 ) = Pr[b = b] −

6.1

1 . 2

Experimental Evaluation Experimental Setup and Scope

Our experiment evaluates RAMP from two aspects. Sections 6.2 and 6.3 evaluate whether real QKD material can be processed through the masked reconciliation and privacyamplification pipeline on general-purpose end-user hardware. Section 6.4 quantifies the system overhead introduced specifically by RAMP, including raw-key acquisition, session binding, post-quantum provider authentication, and protocol metadata. Throughout the evaluation, we employ experimentally acquired QKD data from qOptica 100 CV-QKD testbed [38]. Platforms. To characterize the overhead across hardware an end user could realistically own, we use two laptops of different capabilities without GPU acceleration. Specifically, we run experiments on a high-end laptop as Apple MacBook Pro (M4 Pro, 24 GB RAM) and an ordinary laptop, a Dell Latitude 7430 (Intel Core i7-1265U, 16 GB RAM), respectively. Post-Processing Implementation. We prototype this user-side pipeline in MATLAB, a common choice compatible with QKD implementation. Taking as input the raw-key materials from service providers and the masking key established by users, the pipeline conducts parameter estimation, masking, information reconciliation, and privacy amplification to distill the final key of the RAMP session. We instantiate SSC for keystream generation with AES-256-CTR. We employ forward error correction with LDPC codes of block length nH = 2×106 bits and rate R = 1/10. We aggregate 200 LDPC blocks, which correspond to 4 × 108 accepted bits, before privacy amplification. Section E.2 details the configuration. RAMP Control-Plane Implementation. We prototype RAMP’s setup, registration, raw-key acquisition, session binding, and authenticated masking-key establishment in Python. On both platforms, we evaluate a paired raw-key workload derived from CV-QKD with a size of approximately 3.65GB in total. We instantiate provider signatures SIG with NISTstandardized ML-DSA-65 [34]. The acquisition channels are instantiated with TLS 1.3 in PSK mode [19, 42], using independently provisioned acquisition credentials. The user-touser AKE is instantiated with mutually authenticated TLS 1.3 from OpenSSL [48] with NIST-standardized ML-KEM768 [35] for key establishment, after which a 32-byte session key kAKE is exported and bound to the RAMP session.

Theorem 2 asserts masking-compromise confidentiality of RAMP, with its proof and quantitative bound in Section D.2. It makes no assumption on the secrecy of AKE or SSC, as the target masking state, including kAKE and z, may be exposed. Theorem 2 (Masking-Compromise Confidentiality). Assuming that Q and A hold, the two acquisition channels provide secure-channel security, the underlying QKD protocol provides εqkd -secrecy, and the instantiated reconciliation procedure satisfies public-mask invariance (Lemma 2), RAMP provides masking-compromise final-key confidentiality. Based on Theorems 1 and 2, Section D.3 derives a corollary on dual-domain compromise resilience of RAMP.

5.3 Detectability of Inconsistent Key Material Detectability in RAMP refers to protocol abortion when the supplied raw-key materials violate the admissibility or correctness conditions of the instantiated QKD post-processing procedure. Such inconsistency may be detected during parameter estimation, reconciliation, or correctness verification. Theorem 3 (Detectability of Inconsistent Key Material). Assume that the underlying QKD post-processing procedure accepts inconsistent or otherwise inadmissible input material with probability at most εdet . If any one of the service providers supplies such material, then RAMP accepts the resulting session with probability at most εdet . RAMP performs parameter estimation before masking, so its acceptance behavior remains. For subsequent processing, the masking transformation preserves the discrepancy or softdecision structure of the reconciliation problem, and hence preserves its success or failure behavior. The same relabeling also preserves whether the reconciled binary strings agree, so correctness verification retains its underlying detection guarantee. Therefore, RAMP preserves the relevant acceptance and rejection behavior of the underlying QKD post-processing. Theorem 3 directly follows from the QKD soundness. The unilateral detectability also brings a useful selfchecking feature. Suppose the two providers are independently administered and at most one deviates. A deviating provider may degrade the correlation of its supplied material, e.g., by perturbing its local outcomes. However, it cannot deliberately improve its correlation with the honest-side undisclosed samples beyond the correlation already available from its own QKD view, because those samples are unknown when the raw-key blocks are acquired. In RAMP, the user-side parameter estimation over the paired material provides a direct consistency check on their correlation structure.

6.2

User-Side QKD Processing

Table 1 reports the computational cost when aggregating 200 LDPC blocks (4 × 108 accepted bits). It comfortably exceeds the finite-size threshold, which we evaluate in Section 6.3. Dominant Reconciliation vs. Lightweight Masking. LDPC decoding is the dominant computational stage on both platforms. It requires 717.0s on M4 Pro and 4165.2s on i7-1265U, corresponding to 3.6s and 20.8s per block, respectively. The 11

Table 1: User-Side Processing Time (s) on 200 LDPC Blocks of Length nH = 2 × 106 (4 × 108 accepted bits) Operation

M4 Pro

i7-1265U

127.276 3.723 2.600 0.318 46.759

34.110 14.257 19.588 0.561 248.425

6.3

Parameter estimation on the acquired QKD data yields an operating point with SNR = 0.161 and a channel capacity

Shared / Public (both U, U′ ) Load and clean data Rescale and sampling Key split Parameter estimation Masking keystream

CAB = log2 (1 + SNR) ≈ 0.215 bits/use. “bits/use” denotes secret bits per quantum channel use, or equivalently per transmitted signal state. For the LDPC code with a rate R = 1/10, the effective reconciliation efficiency is βeff =

U (soft info + decode + PA) Octonion map + LLRs Mask apply (LLR sign flip) LDPC decode (200 blocks) Verification tag Privacy amplification

142.514 0.198 717.011 14.664 43.047

616.830 0.421 4165.207 45.770 155.902

U′ (lightweight) Mask apply (bit XOR) Syndrome (200 blocks) Verification tag

0.080 2.829 15.070

0.137 18.043 45.755

1098.111 152.270 1123.514

5301.071 132.824 5387.427

U total U′ total Wall-clock total

Reconciliation Efficiency and Key Yield

2R ≈ 0.929, CAB

where the factor of 2 accounts for the two quadratures obtained by heterodyne detection. We observe zero frame errors and zero residual post-reconciliation bit errors over the evaluated blocks. Thus, at the evaluated operating point, masking in RAMP remains compatible with high-efficiency reconciliation without changing the observed decoding outcome. The estimated channel model gives a Holevo bound on the eavesdropper’s information of χE|B ≈ 0.182 bits/use, which is computed under the collective-Gaussian-attack model adopted by our CV-QKD testbed [38]. The corresponding asymptotic key rate is K∞ = βeffCAB − χE|B = 2R − χE|B ≈ 0.018 bits/use,

complete finite-size processing run finishes in approximately 18.7min on M4 Pro and 89.8min on i7-1265U.

confirming operation in a positive-key regime. The channel parameters are estimated jointly from both quadratures over the parameter-estimation sample budget.

Importantly, this dominant cost is inherited from standard QKD processing rather than introduced by the RAMP service. The RAMP-specific masking work consists of keystream generation and the two local mask applications. These operations take approximately 47.0s on M4 Pro and 249.0s on i7-1265U, accounting for only approximately 4.19% and 4.62% of the respective wall-clock runtimes. It confirms that masking for hiding keying inputs only incurs a small incremental cost.

Finite-Size Effects. Finite-size analysis introduces a privacyamplification penalty δ(N) that decreases with the aggregated raw-key length N, reducing the secret fraction to rsec (N) = 2R − χE|B − δ(N). A positive finite-size key therefore requires accumulating enough blocks to drive δ(N) below the asymptotic margin K∞ ≈ 0.018. With our security parameters this threshold is

Asymmetric Workload. RAMP also preserves the inherent asymmetry of syndrome-based reverse reconciliation. U′ performs only lightweight syndrome computation, essentially a sparse matrix-vector product, with 14.1ms (resp. 90.2ms) per block on the M4 Pro (resp. i7-1265U). By contrast, U bears the decode-heavy reconciliation. Consequently the U′ total time is nearly independent of platform capability (152.3s vs. 132.8s), while the U total time differs by almost 5×.

Nmin ≈ 2 × 108 bits (≈ 100 LDPC blocks). It is a modest requirement that follows directly from the keyrate margin with the both-quadrature parameter estimation. For the aggregated run of N = 4 × 108 bits (200 blocks) reported in Table 1, we obtain δ(N) ≈ 0.0126 and hence a finite-size secret fraction rsec ≈ 0.0057 bits/use, yielding a final key of ℓ ≈ 2.28 × 106 bits after privacy amplification. Privacy amplification over this 4 × 108 -bit input completes in 43.0s on M4 Pro (resp. 155.9s on i7-1265U) using the FFTbased Toeplitz extractor. It confirms that privacy amplification is not a bottleneck, as it is a small fraction of the decode time.

This distinction is useful in heterogeneous deployments. RAMP does not require both users to possess equivalent computational resources. More capable users handle the computationally intensive decoding, while resource-constrained devices work on the opposite side with minimal local overhead. Section F details its effect on deployment and optimization. 12

The effective key-distillation throughputs are around 2.1 × 103 bits/s (M4 Pro),

TLS 1.3 PSK protection under the independently provisioned acquisition credentials, local transport, deserialization, and persistence at the two users. The Mac completes acquisition in a median of 13.99s, corresponding to 260.6MB/s of application-level throughput. The Dell requires 22.95s, corresponding to 158.8MB/s. Across the 30 measured runs, the interquartile latency ranges are 13.45–14.32s and 21.46–24.52s, respectively. The result characterizes the protected acquisition path rather than a pure network or cryptographic primitive. We specifically evaluate the time of authenticating the binding between the selected kid, users, and providers using ML-DSA-65. Median signing and verification costs are 7.48ms and 6.82ms on M4 Pro, and 5.53ms and 4.39ms on i7-1265U, respectively. A signing-and-verification pair therefore accounts for only approximately 0.10% and 0.04% of the respective acquisition times. Thus, provider authentication is not the dominant cost of Acq(). The acquisition cost mainly comes from the transmission of the raw-key material.

4.2 × 102 bits/s (i7-1265U).

Viewed as 256-bit symmetric-key material, these rates correspond to approximately 8.2 and 1.6 AES-256 key-equivalents per second, which are derived from quantum entropy. The result demonstrates the computational throughput of the complete user-side post-processing pipeline at the evaluated finitesize operating point. The acquisition and masking-key establishment are measured separately in Section 6.4. Privacy amplification remains subdominant to LDPC decoding. On Aggregated Block Count. The 200-block result exhibits the complete user-side QKD post-processing workload at a finite-size-secure operating point (N > Nmin ). The asymptotic rate, the Holevo bound, and the finite-size threshold above are all properties of the estimated channel and are independent of how many blocks are processed. Certifying a fully independent finite-size key of a specific length additionally requires that the aggregated input consist of independent raw-key material of the corresponding size, which is a function of the acquired data volume rather than of the processing pipeline.

6.4

6.4.3

Session Binding. After acquisition, the two users bind the session by computing the hash of (U,U ′ , P, P′ , kid) as sid. The operation takes a median of only 19.1µs on M4 Pro and 24.3µs on i7-1265U, which is minor relative to both raw-key acquisition and subsequent QKD processing.

RAMP Service-Path Evaluation

The evaluation follows the execution order of RAMP. It allows us to clarify one-time provisioning costs and per-session operations, identifying where the system spends time along an actual execution. Online operations are measured over 30 executions after 5 warm-up runs, cryptographic primitives over 100 executions, and one-time setup over 10 executions. 6.4.1

Pre-Mask QKD Preparation. Consistent with the construction in Section 4.3, reconciliation inputs are fixed before the masking key is established. This operation belongs to the QKD-processing pipeline rather than the RAMP controlplane, therefore evaluated together with multidimensional reconciliation and post-processing in Section 6.2.

Setup and Raw-Key Provisioning

Masking-Key Establishment. Once the reconciliation inputs and sid are fixed, the users execute the authenticated keyestablishment step. Across all 60 measured sessions (30 per platform), the implementation successfully negotiates the MLKEM-768 TLS group and derives a fresh 32-byte kAKE . The current implementation completes the masking-keyestablishment wrapper in a median of 2.33s on M4 Pro and 2.28s on i7-1265U. It is executed once per RAMP session. However, the prototype invokes an external OpenSSL server process [48], and the measured wall-clock interval also includes process startup, synchronization, and teardown. The quantity reports an implementation-level masking-keyestablishment cost, rather than a primitive-level benchmark.

RAMP begins with provider initialization and QKD raw-key generation. We implement provider initialization and the storage update associated with importing already generated QKD material. The physical QKD stage itself is due to qOptica 100 CV-QKD testbed [38] and independent of our evaluation. The identity initialization of providers (and users) is complete in the tens-of-milliseconds range. Importing and indexing the prepared 3.65GB paired raw-key workload takes 1.32s on M4 Pro and 2.68s on i7-1265U, including integrity hashing, identifier derivation, and provider-side storage updates. 6.4.2

Session Binding and Masking-Key Establishment

Registration and Raw-Key Acquisition

Registration. Regist() provisions the independent symmetric acquisition credential kacq before raw-key acquisition. Its cost is only sub-millisecond to several milliseconds depending on both platforms, minor compared with subsequent acquisition.

6.5

Evaluation Summary

RAMP preserves practical QKD post-processing. On experimentally acquired CV-QKD data, the pipeline achieves a reconciliation efficiency of βeff ≈ 0.929, with no observed frame errors or residual post-reconciliation bit errors. The complete

Raw-Key Acquisition. We then execute Acq() over the 3.65GB paired workload. The measured path includes providerside raw-key retrieval and integrity checking, serialization, 13

Table 2: RAMP Service-Path Performance Protocol Stage

Measured Operation

M4 Pro

i7-1265U

Setup (Section 4.1)

Raw-Key provisioning

1.32s

2.68s

Acquisition (Section 4.2)

Raw-key acquisition Acquisition throughput ML-DSA-65 Sign ML-DSA-65 Verify

13.99s 260.6MB/s 7.48ms 6.82ms

22.95s 158.8MB/s 5.53ms 4.39ms

19.1µs 2.33s 30/30

24.3µs 2.28s 30/30

Session binding (Section 4.3) Masking-key establishment (Section 4.3)

sid computation ML-KEM-768 AKE† time ML-KEM-768 AKE success rate

† It is the wall-clock time of the current implementation wrapper, including OpenSSL subprocess management.

References

user-side prototype processes this workload at approximately 2.1Kb/s on the M4 Pro and 0.42Kb/s on the i7-1265U. RAMP-specific protection remains a small fraction of the overall cost. Mask generation and application account for only 4.19% and 4.62% of the measured post-processing wall-clock time on the two platforms, while session binding and ML-DSA authentication remain in the microsecond-tomillisecond regime. The current masking-key-establishment wrapper adds approximately 2.3s per session, including OpenSSL process-management overhead. The user-side QKD post-processing takes 1123.5s and 5387.4s. The measured cost is therefore dominated by the reconciliation of the QKD material rather than by the RAMP service path. The experiment reveals a valuable separation between service provisioning and user-side computation. The LDPC reconciliation is performed after acquisition at the users. As per our results, this workload is itself asymmetric, i.e., U executes the decode-heavy operation, whereas U′ performs lightweight syndrome generation. The protocol therefore executes more efficiently across two devices of substantially different capabilities, still requiring no QKD hardware at either endpoint. Overall, the experiments establish feasibility rather than a production-performance bound. Our evaluation is on one low-SNR CV-QKD operating point, one LDPC configuration, and a local acquisition path. Different QKD parameters, optimized decoders, networks, and hardware will affect performance. Despite the boundaries, our result shows that RAMP successfully moves QKD post-processing to commodity user devices with modest control costs. RAMP enables users without QKD hardware to locally distill QKD-derived keys.

7

[1] Mayuresh Vivekanand Anand, Ehsan Ebrahimi Targhi, Gelo Noel Tabia, and Dominique Unruh. Post-quantum security of the cbc, cfb, ofb, ctr, and XTS modes of operation. In PQCrypto, pages 44–63, 2016. [2] Mihir Bellare and Phillip Rogaway. Entity authentication and key distribution. In Douglas R. Stinson, editor, CRYPTO, volume 773 of Lecture Notes in Computer Science, pages 232–249, 1993. [3] Charles H. Bennett and Gilles Brassard. Quantum cryptography: Public key distribution and coin tossing. Theor. Comput. Sci., 560:7–11, 2014. (conference version published in IEEE Int. Conf. Comput., Syst. Signal, 1984). [4] Daniel J Bernstein and Tanja Lange. Post-quantum cryptography. Nature, 549(7671):188–194, 2017. [5] Nina Bindel, Udyani Herath, Matthew McKague, and Douglas Stebila. Transitioning to a quantum-resistant public key infrastructure. In PQCrypto, pages 384–405, 2017. [6] Gilles Brassard and Louis Salvail. Secret-key reconciliation by public discussion. In EUROCRYPT, pages 410–423, 1993. [7] Sonja Bruckner, Sebastian Ramacher, and Christoph Striecks. Muckle+: End-to-end hybrid authenticated key exchanges. In PQCrypto, pages 601–633, 2023.

Conclusion

RAMP elevates QKD from a device-centric model to a serviceoriented architecture. By decoupling quantum-state generation from user-side processing, RAMP enables ordinary users to obtain quantum-derived keys while providing final-key confidentiality under complementary compromise modes.

[8] Ran Canetti and Hugo Krawczyk. Analysis of keyexchange protocols and their use for building secure channels. In Birgit Pfitzmann, editor, EUROCRYPT, volume 2045 of Lecture Notes in Computer Science, pages 453–474. Springer, 2001. 14

[9] Yuan Cao, Yongli Zhao, Rui Lin, Xiaosong Yu, Jie Zhang, and Jiajia Chen. Multi-tenant secret-key assignment over quantum key distribution networks. Optics Express, 27(3):2544–2561, 2019.

[21] Paul Jouguet, Simon Kunz-Jacques, Anthony Leverrier, Philippe Grangier, and Eleni Diamanti. Long-distance continuous-variable quantum key distribution with a gaussian modulation. Physical Review A, 84(6):062317, 2011.

[10] Yuan Cao, Yongli Zhao, Jianquan Wang, Xiaosong Yu, Zhangchao Ma, and Jie Zhang. Sdqaas: Software defined networking for quantum key distribution as a service. Optics express, 27(5):6892–6909, 2019.

[22] Jonathan Katz and Yehuda Lindell. Introduction to Modern Cryptography, Second Edition. CRC Press, 2014. [23] Sergejs Kozlovics, Krisjanis Petrucena, Davis Larins, and Juris Viksna. Quantum key distribution as a service and its injection into TLS. In ISPEC, pages 527–545, 2023.

[11] Yuan Cao, Yongli Zhao, Qin Wang, Jie Zhang, Soon Xin Ng, and Lajos Hanzo. The evolution of quantum key distribution networks: On the road to the qinternet. IEEE Commun. Surv. Tutorials, 24(2):839–894, 2022.

[24] Hugo Krawczyk. The order of encryption and authentication for protecting communications (or: How secure is ssl?). In CRYPTO, volume 2139 of Lecture Notes in Computer Science, pages 310–331. Springer, 2001.

[12] A. Doherty, M. Pei, S. Machani, and M. Nystrom. Dynamic Symmetric Key Provisioning Protocol (DSKPP). RFC 6063, December 2010.

[25] Adeline Langlois and Damien Stehlé. Worst-case to average-case reductions for module lattices. Des. Codes Cryptogr., 75(3):565–599, 2015.

[13] Benjamin Dowling, Torben Brandt Hansen, and Kenneth G. Paterson. Many a mickle makes a muckle: A framework for provably quantum-secure hybrid key exchange. In PQCrypto, pages 483–502, 2020.

[26] Anthony Leverrier, Romain Alléaume, Joseph Boutros, Gilles Zémor, and Philippe Grangier. Multidimensional reconciliation for continuous-variable quantum key distribution. Physical Review A, 77(4):042325, 2008.

[14] David Elkouss, Jesus Martinez-mateo, and Vicente Martin. Information reconciliation for quantum key distribution. Quantum Info. Comput., 11(3):226–238, March 2011.

[27] Sheng-Kai Liao, Wen-Qi Cai, Johannes Handsteiner, Bo Liu, Juan Yin, Liang Zhang, Dominik Rauch, Matthias Fink, Ji-Gang Ren, Wei-Yue Liu, et al. Satelliterelayed intercontinental quantum network. Physical review letters, 120(3):030501, 2018.

[15] ETSI GS QKD 014. Quantum key distribution (qkd); protocol and data format of rest-based key delivery api, 2019. [16] ETSI TR 103 967. Quantum-safe cryptography (qsc); impact of quantum computing on symmetric cryptography, 2025.

[28] Thomas Lorünser, Stephan Krenn, Christoph Pacher, and Bernhard Schrenk. On the security of offloading post-processing for quantum key distribution. Entropy, 25(2):226, 2023.

[17] Carlos Rubio Garcia, Abraham Cano Aguilera, Catalina Ioana Stan, Juan José Vegas Olmos, Simon Rommel, and Idelfonso Tafur Monroy. Enhanced network security protocols for the quantum era: Combining classical and post-quantum cryptography, and quantum key distribution. IEEE J. Sel. Areas Commun., 43(8):2765–2781, 2025.

[29] Miralem Mehic, Marcin Niemiec, Stefan Rass, Jiajun Ma, Momtchil Peev, Alejandro Aguado, Vicente Martin, Stefan Schauer, Andreas Poppe, Christoph Pacher, et al. Quantum key distribution: a networking perspective. ACM Computing Surveys (CSUR), 53(5):1–41, 2020. [30] Alfred Menezes, Paul C. van Oorschot, and Scott A. Vanstone. Handbook of Applied Cryptography. CRC Press, 1996.

[18] Lov K. Grover. A fast quantum mechanical algorithm for database search. In STOC, pages 212–219. ACM, 1996.

[31] Mario Milicevic, Chen Feng, Lei M. Zhang, and P. Glenn Gulak. Quasi-cyclic multi-edge ldpc codes for longdistance quantum cryptography. npj Quantum Information, 4(1):21, 2018.

[19] R. Housley, J. Hoyland, M. Sethi, and C. A. Wood. Guidance for External Pre-Shared Key (PSK) Usage in TLS. RFC 9257, July 2022.

[32] Chanathip Namprempre. Secure channels based on authenticated encryption schemes: A simple characterization. In ASIACRYPT, volume 2501 of Lecture Notes in Computer Science, pages 515–532, 2002.

[20] Kyo Inoue, Edo Waks, and Yoshihisa Yamamoto. Differential phase shift quantum key distribution. Physical review letters, 89(3):037902, 2002. 15

[33] NIST. IR 8105, report on post-quantum cryptography. https://csrc.nist.gov/pubs/ir/8105/final, 2016.

[47] Damien Stucki, Matthieu Legre, Francois Buntschu, B Clausen, Nadine Felber, Nicolas Gisin, Luca Henzen, Pascal Junod, Gérald Litzistorf, Patrick Monbaron, et al. Long-term performance of the swissquantum quantum key distribution network in a field environment. New Journal of Physics, 13(12):123001, 2011.

[34] NIST. Module-lattice-based digital signature standard. https://doi.org/10.6028/NIST.FIPS.204, 2024. [35] NIST. Module-lattice-based key-encapsulation mechanism standard. https://doi.org/10.6028/NIST.F IPS.203, 2024.

[48] The OpenSSL Project. OpenSSL: The Open Source Toolkit for SSL/TLS. https://www.openssl.org/, 2025. Version 3.5.

[36] Momtchil Peev, Christoph Pacher, Romain Alléaume, Claudio Barreiro, Jan Bouda, Winfried Boxleitner, Thierry Debuisschert, Eleni Diamanti, Mehrdad Dianati, James F Dynes, et al. The secoqc quantum key distribution network in vienna. New journal of physics, 11(7):075001, 2009.

[49] Marco Tomamichel and Anthony Leverrier. A largely self-contained and complete security proof for quantum key distribution. Quantum, 1:14, 2017. [50] William K Wootters and Wojciech H Zurek. A single quantum cannot be cloned. Nature, 299(5886):802–803, 1982.

[37] Christopher Portmann and Renato Renner. Cryptographic security of quantum key distribution. CoRR, abs/1409.3525, 2014.

[51] Feihu Xu, Xiongfeng Ma, Qiang Zhang, Hoi-Kwong Lo, and Jian-Wei Pan. Secure quantum key distribution with realistic devices. Reviews of Modern Physics, 92(2):025002, 2020.

[38] Quintessence Labs. qoptica 100 cv-qkd system. https: //www.quintessencelabs.com. Product specification sheet.

[52] Hongzhao Yang, Shuaishuai Liu, Shenshen Yang, Zhenguo Lu, Yanxia Li, and Yongmin Li. High-efficiency rate-adaptive reconciliation in continuous-variable quantum key distribution. Phys. Rev. A, 109:012604, 2024.

[39] Timothy C Ralph. Continuous variable quantum cryptography. Physical Review A, 61(1):010303, 1999. [40] Jothi Ramalingam, Satyabrat Rath, Lakshmi Kuppusamy, and Cheng-Chi Lee. Accelerating QKD postprocessing by secure offloading of information reconciliation. Comput. Electr. Eng., 120:109721, 2024.

[53] Juan Yin, Yuan Cao, Yuhuai Li, Shengkai Liao, Liang Zhang, Jigang Ren, Wenqi Cai, Weiyue Liu, Bo Li, Hui Dai, Guangbing Li, Qiming Lu, Yunhong Gong, Yu Xu, Shuanglin Li, Fengzhi Li, Yayun Yin, Ziqing Jiang, Ming Li, Jianjun Jia, Ge Ren, Dong He, Yilin Zhou, Xiaoxiang Zhang, Na Wang, Xiang Chang, Zhencai Zhu, Naile Liu, Yuao Chen, Chaoyang Lu, Rong Shu, Chengzhi Peng, Jianyu Wang, and Jianwei Pan. Satellitebased entanglement distribution over 1200 kilometers. Science, 356(6343):1140–1144, 2017.

[41] Renato Renner, Nicolas Gisin, and Barbara Kraus. Information-theoretic security proof for quantum-keydistribution protocols. Phys. Rev. A, 72:012332, 2005. [42] E. Rescorla. The Transport Layer Security (TLS) Protocol Version 1.3. RFC 9846, July 2026. [43] Masahide Sasaki, Mikio Fujiwara, H Ishizuka, W Klaus, K Wakui, M Takeoka, S Miki, T Yamashita, Z Wang, A Tanaka, et al. Field test of quantum key distribution in the tokyo qkd network. Optics express, 19(11):10387– 10409, 2011.

A

Ethics Considerations

This work does not involve human subjects, personal data, user studies, or data derived from individuals. Our experiments are conducted in a controlled local setting using QKD-derived experimental data and locally implemented post-processing components. We do not perform network scanning, interact with third-party production systems, exploit deployed services, or evaluate attacks against real users or organizations. The paper proposes and evaluates a key-establishment architecture and does not report any vulnerability in an existing system. Accordingly, there is no responsible-disclosure process required for this work. The experimental results are intended to assess the feasibility and performance of the proposed design under the stated assumptions, rather than to certify a production deployment.

[44] Peter W. Shor. Algorithms for quantum computation: Discrete logarithms and factoring. In FOCS, pages 124– 134, 1994. [45] Dimitrios Sikeridis, Panos Kampanakis, and Michael Devetsikiotis. Post-quantum authentication in TLS 1.3: A performance study. In NDSS, 2020. [46] M. Stanley, Y. Gui, D. Unnikrishnan, S.R.G. Hall, and I. Fatadin. Recent progress in quantum key distribution network deployments and standards. In Journal of Physics: Conference Series, volume 2416, page 012001. IOP Publishing, 2022. 16

B B.1

Related Work

QKD raw-key materials as a service and lets users locally mask and process them into final keys. RAMP resolves the trust challenge by incorporating quantum-safe cryptography and formalizing security against the service infrastructure.

QKD Networks

QKD networks [36,43,47] rely on trusted repeaters, metropolitan backbones, and secret key management to extend QKD beyond single-link distances. Satellite-assisted QKD [27, 53] further extends geographic reach with space infrastructures. The standardization of QKD networks has become an active research topic, driven by the challenges of coordinating heterogeneous deployments and interfaces [46]. Despite the progress, existing QKD networks remain device-centric, where each participating node has to host specialized QKD hardware, maintain quantum channels, and execute post-processing. The architectures also require the intermediate nodes, i.e., relays or repeaters, to be trusted when routing the secret keys. These systems improve QKD coverage, but typically feature limited elasticity under multi-user demand. In contrast, RAMP is a service-centric framework, which only requires the service providers to possess QKD hardware. It decouples the key provision from its consumption, enabling more flexible consumption of QKD-derived material.

B.2

B.3

Hybrid PQC–QKD Key Establishment and Key Combination

Hybrid constructions combine independently protected key sources so that the resulting session key can remain secure when one source of key fails [7, 13, 17]. When deployed for QKD-as-a-service, a straightforward adaption is for service providers to complete QKD post-processing, deliver a final QKD key to the users, and let the users combine it with an independently established pre- or post-quantum secret. Such a design could provide analogous source-failure confidentiality as RAMP while requiring less QKD processing at the users. To our knowledge, prior hybrid and QKD-service constructions do not address this pre-reconciliation service boundary or the associated problem of protecting provider-known noisy QKD inputs while preserving user-side reconciliation. In RAMP, providers expose pre-reconciliation QKDderived material, and the users execute the classical QKD security pipeline locally. The problem is addressed through our reconciliation-preserving masking and formal analysis of the resulting service boundary.

Offloading and QKD-as-a-Service

Offloading processing of raw-key materials has recently been studied [28, 40]. They consider partial delegation to a third party, rather than systematically offloading it to users. A branch of QKD as a service (QaaS) [9, 10] studies how to arrange resources over existing optical networks via network orchestration or physical design to match the multi-user requirement. Cao et al. [9] consider multi-tenant secret-key assignment within QKD networks, where software-defined networking (SDN) and quantum key pools are used to allocate finite secret-key rates among tenant requests. The follow-up SDQaaS [10] further develops SDN-based QaaS functions for service creation, modification, routing, and assignment of secret-key rates over an existing QKD infrastructure. These works improve the operational efficiency of QKD networks through resource management and service provisioning. They do not address the trust problem that arises when service providers may abuse the keys that were assigned to users. Kozlovics et al. [23] propose a service framework that enables users to obtain QKD keys from two key distribution centers (similar to the service providers in RAMP) over classical networks. Their design deploys a butterfly-style delivery protocol in which the two centers send different halves of a QKD-generated key, ensuring that no single classical channel carries the complete key. Consequently, a user must communicate with both centers to recover the QKD keys. This mechanism strengthens key delivery over classical channels, but it does not hide the key from the key distribution centers. Different from the prior works, RAMP makes access to

C C.1

More on Cryptographic Building Blocks Digital Signature Scheme (SIG)

Correctness. A digital signature scheme SIG is correct if, for any security parameter 1λ , any (pk, sk) ∈ KeyGen(1λ ), and any message m ∈ M with signature σ ∈ Sign(sk, m), it holds that Vf(pk, m, σ) = 1 with overwhelming probability. Security. The standard security notion for signature schemes is existential unforgeability under chosen-message attacks (EUF-CMA). It requires that even if an adversary is allowed to obtain signatures on arbitrary messages of its choice, it cannot produce a valid signature on any message that it did not query. Definition 6 (EUF-CMA Security). SIG is existentially unforgeable under chosen-message attacks if, for any probabilistic or quantum polynomial-time adversary A , the following probability is negligible in λ:

Pr

h

(pk, sk) ← KeyGen(1λ ), i . (m∗ , σ∗ ) ← A Sign(sk,·) (pk) : ∗ ∗ ∗ Vf(pk, m , σ ) = 1 ∧ m ∈ /Q

where Q is the set of messages queried by A to its signing oracle. 17

Quantum Resistance. In response to the quantum threat, NIST has approved ML-DSA [34], derived from CRYSTALSDilithium, as a post-quantum digital-signature standard. MLDSA relies on the computational difficulty of solving the module learning with errors problem [25], which is believed to be difficult for adversaries with quantum capabilities.

C.2

according to the standard secure-channel notions for symmetric authenticated communication [24, 32]. -SEC λ We write AdvSC SC,A (1 ) as shorthand for the aggregate advantage of A against these confidentiality and integrity -SEC λ notions. SC is secure if AdvSC SC,A (1 ) is negligible in λ for every probabilistic or quantum polynomial-time adversary A . -SEC λ In our security analysis of RAMP, AdvSC SC,A (1 ) upperbounds the combined distinguishing and integrity-failure loss incurred by one target secure-channel execution. The quantity is negligible when SC provides secure channel security.

Authenticated Key Exchange (AKE)

Security. In general, AKE is secure if it is infeasible for the adversary to distinguish a key generated by the protocol from an independent random value sampled from the key space K .

Quantum Resistance. The trust root of SC could be a highentropy symmetric credential established during the prior provisioning phase. Its resistance to quantum adversaries instead depends on the security parameters and symmetric primitives used by the instantiated channel. A concrete instantiation is TLS 1.3 with an externally provisioned PSK in PSK-only mode [19, 42]. TLS 1.3 explicitly supports externally established PSKs and derives fresh connection-specific traffic keys from the shared PSK. Symmetric primitives with enough key sizes remain appropriate candidates against quantum-capable adversaries [33], which are also discussed in Section C.4.

Definition 7 (Session-Key Security). Consider an adversary A that controls the communication network and may interact concurrently with protocol sessions, subject to the exposure rules. For a completed and unexposed target session (U, U′ , sid) between two parties U and U′ , let k denote its λ session key. We define the following advantage AdvSK AKE,A (1 )

Pr

h

$ i 1 b← − {0, 1}, k0 ← AKE(U,U ′ , sid), k1 ←$ K − . 2 b′ ← A (kb ) : b′ = b

We say that AKE provides session-key security if (i) matching sessions completed by uncorrupted U and U′ output the λ same session key, and (ii) AdvSK AKE,A (1 ) is negligible in λ.

C.4

Correctness. Following Definition 3, SSC is correct if, for any 1λ , any k ∈ KeyGen(1λ ), any output length n, any keystream z ∈ Keystream(k, n), and any n-bit message m, it holds that Dec(Enc(m, z), z) = m with overwhelming probability.

Quantum Resistance. For post-quantum AKE, the established session key of an unexposed accepted session must remain computationally indistinguishable from an independently sampled key even when the adversary has quantum computational capability. We instantiate post-quantum AKE in our construction with mutually authenticated TLS 1.3 that employs NIST-standardized post-quantum primitives for key establishment [35] and peer authentication [34].

C.3

Synchronous Stream Cipher (SSC)

Security. The standard security notion for SSC is indistinguishability from random bits against chosen-plaintext attack (IND-CPA). Intuitively, it requires that for any efficient adversary, a ciphertext is computationally indistinguishable from a truly random sequence of the same length, even if the adversary can choose the plaintext to be challenged with adaptively. Definition 9 (IND-CPA Security of SSC). An SSC = (KeyGen, Keystream, Enc, Dec) is indistinguishable under chosen-plaintext attacks if, for any probabilistic or quantum polynomial-time adversary A , the following advantage -CPA λ AdvIND SSC,A (1 ) is negligible in λ:

Secure Channel

Security. We adopt the standard secure-channel requirements of confidentiality and network authentication/integrity [24, 32], where the provisioned symmetric credential replaces the online establishment of the channel trust root.

k ← KeyGen(1λ ),

Definition 8 (Secure Channel Security). Let SC be a stateful two-party channel protocol rooted in a provisioned credential. We say that SC is secure if, whenever the target credential remains unexposed, the channel provides:

$

Pr

1. confidentiality of protected messages against an active network adversary; and

h

i 1 (m, n) ← A (), b ← − {0, 1}, z ← Keystream(k, n), c0 ← Enc(m, z), − . 2 c1 ←$ {0, 1}|c0 | , b′ ← A (cb ) : b′ = b

Quantum Resistance. While public-key primitives like RSA or ECC are broken by Shor’s quantum algorithm [44], symmetric-key primitives (and stream ciphers built upon them) offer better resistance. The foremost quantum threat is

2. channel authentication and integrity, such that an honest receiver does not accept an unauthorized modification, injection, or replay under the target channel context, 18

H0 : H0 is exactly the real experiment of Definition 4. Both providers may be corrupted and may expose their complete target-session state. The honest users establish kAKE ← AKE(U,U ′ , sid), derive z ← SSC.Keystream(kAKE , n), and execute masking and post-processing as in Section 4. By definition, we have

Grover’s algorithm [18], which gives a square-root speed-up for brute-force key search with quantum computers. That is, an n-bit key can in theory be searched in approximate O(2n/2 ) quantum steps using a quantum computer. This also implies that doubling key size can restore the original security level. The report by NIST [33] remarks that symmetric algorithms remain viable in a quantum future provided key lengths are increased appropriately. More recent analyses and quantum resource estimates [16] refine this understanding by considering circuit depth, qubit overhead, and fault-tolerance constraints. The consensus remains that the stream ciphers retain resilience against realistic quantum attacks, assuming proper implementations and avoidance of structural weaknesses. We instantiate SSC with AES in counter mode (AES-CTR). It is deemed quantum-safe with an appropriate key length [1].

D

λ AdvPC Π,A (1 ) = p0 −

H1 : H1 is identical to H0 , except that the common key used by the target session for masking is replaced by an independently sampled key k⋆ ←$ K , where K is the AKE/SSC key space. The users derive z ← SSC.Keystream(k⋆ , n) and execute the remaining parts of RAMP unchanged. Condition M guarantees that the target AKE session is unexposed. Therefore, if A distinguishes H0 from H1 , we obtain an adversary B1 against the AKE session-key security. Specifically, B1 embeds its AKE challenge key as the masking key of the target RAMP session and simulates the remaining RAMP execution for A . If the AKE challenger supplies the real session key, the simulation is distributed as H0 ; if it supplies an independent key from K , the simulation is distributed as H1 . Under the advantage convention of Definition 7,

Security Arguments of RAMP

D.1 Proof of Provider-Compromise Final-Key Confidentiality (Theorem 1) We first show the pseudorandomness of the keystream generated in SSC by Lemma 1. Lemma 1 (Pseudorandom Keystream). Let SSC be an XORbased synchronous stream cipher. If SSC is IND-CPA secure as in Definition 9, then for any probabilistic or quantum polynomial-time adversary A and any output length n decided by A , the following advantage is negligible in λ:

λ |p0 − p1 | ≤ 2 AdvSK AKE,B1 (1 ).

H2 : H2 is identical to H1 , except that the SSC keystream is replaced by z⋆ ←$ {0, 1}n . All subsequent masking and post-processing operations use z⋆ . Since k⋆ is independently distributed as an SSC key, Lemma 1 implies that replacing SSC.Keystream(k⋆ , n) by a uniform n-bit string is computationally indistinguishable. Hence there exists an IND-CPA adversary B2 such that

$

Pr

h

1 . 2

i 1 k ← KeyGen(1λ ), b ← − {0, 1}, z0 = Keystream(k, n), z1 ←$ {0, 1}n , − . 2 b′ ← A (zb ) : b′ = b

-CPA λ |p1 − p2 | ≤ 2AdvIND SSC,B2 (1 ).

Lemma 1 can be proved as follows. Assume there exists an adversary A that could distinguish Keystream(k, n) from a uniform string with non-negligible advantage. We build an adversary B against the IND-CPA security of SSC. B submits the plaintext m = 0n and output length n as the challenge to the challenger in SSC. It receives c that is either

It remains to analyze the information-theoretic part of H2 . We consider the two QKD modes separately. • DV-QKD. U′ holds the binary reference string x′ ∈ {0, 1}n before starting reconciliation. After masking in H2 , U′ has rk′mask = x′ ⊕ z⋆ .

Enc(0n , Keystream(k, n)) = 0n ⊕ Keystream(k, n)

The provider-compromise experiment permits A to know x′ completely. This does not reduce the entropy introduced by the mask. Since z⋆ is uniform and independent of all information available when x′ is fixed, rk′mask is itself uniformly distributed over {0, 1}n , even conditioned on the complete provider-side view.

= Keystream(k, n) or a uniform string, depending on the challenge bit. B forwards c to A and outputs whatever A outputs. Thus, any non-negligible distinguishing advantage of A gives the same advantage against IND-CPA security of SSC. It thus contradicts the fact that SSC is IND-CPA secure.

• CV-QKD. In the reverse-reconciled CV-QKD, as described in Section 4.3, U ′ holds the binary reference string x′ ∈ {0, 1}n , while U holds its corresponding LLR vector x = (L1 , . . . , Ln ). RAMP transforms them as

Proof of Theorem 1 We prove it by a sequence of hybrid games. Let A be an admissible adversary in PCΠ,A (1λ ). For each hybrid Hi , let pi = Pr[A outputs b′ = b in Hi ].

′ rkmask = x ′ ⊕ z⋆ ,

19

ei = (−1)z⋆i Li . L

As in the DV case, the binary reference x′ ⊕ z⋆ is uniformly distributed even if the corrupted providers know the original x′ and L completely. e is auxiliary soft informaThe transformed LLR vector L tion used locally by the decoder; it is not released to the service providers or included in the public transcript. Its role is to preserve the likelihood relation associated with the masked reference bits: flipping xi′ is accompanied by exchanging the corresponding binary likelihood labels, represented by negating Li . Thus, the CV transformation does not supply additional entropy; rather, it ensures that reconciliation remains consistent with the uniformly masked binary reference string.

λ AdvPC Π,A (1 ) is negligible, proving provider-compromise finalkey confidentiality. Importantly, this argument does not rely on QKD-secrecy or acquisition-security assumptions. The adversary may know the complete unmasked QKD-derived inputs. In the providercompromise branch, the entropy used for final-key extraction is introduced by the independent user-side mask, while privacy amplification removes the bounded information subsequently exposed through public post-processing. □

D.2 Proof of Masking-Compromise Final-Key Confidentiality (Theorem 2) We prove Theorem 2 in two steps. We first account for the confidentiality and integrity of the two target acquisition channels using the secure-channel security. Specifically, we use -SEC λ AdvSC SC,A (1 ) as defined in Definition 8 to represent the security loss of each target acquisition channel. We then show that, once the masking state is revealed, RAMP masking constitutes only a publicly invertible relabeling of the underlying QKD post-processing execution. Recall that the masking-compromise experiment exposes the complete cryptographic masking state, including kAKE and z, but does not compromise either user endpoint. Conditions Q and A remain in force. Thus, the QKD-derived materials are not disclosed by the service providers and are delivered to the users without adversarial disclosure or modification. The proof does not need the secrecy or pseudorandomness of the masking state. Instead, it uses the fact that a disclosed mask induces a public relabeling of the reconciliation inputs.

Therefore, for both QKD modes, the value from which the reconciled key is derived contains an n-bit uniformly masked binary reference string. More generally, let E denote the adversary’s complete state immediately after masking and before the subsequent keydependent public communication. The state E may contain x′ , the original raw-key materials, metadata, acquisition credentials, the AKE transcript, and all earlier public messages.
In H2 , Hmin rk′mask | E = n. This equality does not arise from the QKD source. It follows solely from the independent uniform mask z⋆ . Let C denote all key-dependent public information subsequently released by reconciliation and correctness verification. By the output-length required in RAMP (Section 4.4), this public leakage is accounted for before privacy amplification. Equivalently, if leak denotes the corresponding conditionalmin-entropy loss, then
Hmin rk′mask | E,C ≥ n − leak.

Lemma 2 (Public-Mask Invariance). Conditioned on the public session context, the masking state, including the masking keystream z, is generated independently of the secret contents of the QKD-derived reconciliation inputs. For DV-QKD masking (Section 4.3), (x, x′ ) 7−→ (x ⊕ z, x′ ⊕ z) is an invertible relabeling that preserves the complete discrepancy pattern. For CV-QKD masking (Section 4.3), x′ 7−→ x′ ⊕ z, together with Li 7−→ (−1)zi Li , is an invertible relabeling of the corresponding binary soft-decision reconciliation problem. If the instantiated reconciliation procedure is equivariant under these relabelings, then the masked and unmasked executions have the same accept/abort behavior and the same reconciliation-leakage bound, and their reconciled binary reference strings differ only by the public XOR offset z.

Conditioned on successful reconciliation, both honest users hold the same masked reference value on which privacy amplification is applied. By the assumptions in Theorem 1, the universal-hash privacy-amplification output length ℓ is selected after accounting for this post-masking leakage. With the leftover-hash guarantee the resulting final key is εPA close to an independent uniformly random ℓ-bit string. Consequently, replacing the real final key in H2 with an independent uniform string changes the adversary’s view by at most εPA . Combining the bounds by the triangle inequality yields λ AdvPC Π,A (1 ) = p0 −

1 2

≤ |p0 − p1 | + |p1 − p2 | + p2 −

Proof of Lemma 2. We analyze the QKD modes separately.

1 2

• DV-QKD. For any fixed z ∈ {0, 1}n , applying the DV masking operation twice gives

λ IND-CPA λ ≤ 2AdvSK AKE,B1 (1 ) + 2AdvSSC,B2 (1 ) + εPA

(x ⊕ z) ⊕ z = x,

The first two terms are negligible by the assumed postquantum security of AKE and SSC, while εPA is negligible for the selected privacy-amplification parameters. Hence

(x′ ⊕ z) ⊕ z = x′ .

Hence the transformation is a bijection whose inverse is itself. Moreover, (x ⊕ z) ⊕ (x′ ⊕ z) = x ⊕ x′ . Namely, 20

H0 : H0 is exactly the masking-compromise experiment of Definition 5. The users acquire the target raw-key materials and metadata through the real secure channels instantiated from their independently provisioned credentials. By definition, we have

every disagreement position between the two reconciliation inputs is preserved exactly. In particular, the error pattern and error rate seen by the reconciliation procedure are unchanged. • CV-QKD. In the reverse-reconciled construction, U ′ holds the binary reference string x′ , while U holds the corresponding LLR vector x = (L1 , . . . , Ln ). Applying the same transformation twice restores both values as (xi′ ⊕ zi ) ⊕ zi = xi′ ,

λ AdvMC Π,A (1 ) = p0 −

1 . 2

H1 : H1 is identical to H0 , except that the acquisition channels between U and P is replaced according to the experiment of secure channel security. Specifically, for the network-visible protection of the target acquisition payload, the protected raw-key material and metadata are replaced by equal-length dummy values. The honest endpoint nevertheless receives the original provider-supplied payload internally. Executions in which the adversary causes U to accept an unauthorized modification, injection, or replay on this channel are treated as a secure-channel integrity failure. If A distinguishes H0 from H1 , or causes the accepted acquisition payload to differ from the provider-supplied payload, we obtain an adversary B1 against the secure-channel security (Definition 8). Therefore,

(−1)zi (−1)zi Li = Li .

The reliability of every soft-decision position is unchanged. When zi = 1, the binary label is exchanged and the corresponding LLR sign is reversed; when zi = 0, neither is changed. The same binary likelihood problem thus remains under a different labeling after masking. For the linear syndrome reconciliation in CV-QKD, if H is the parity-check matrix, then H(x′ ⊕ z) = Hx′ ⊕ Hz. Hence the syndrome of the masked reference differs from the original syndrome only by the public offset Hz. Under the assumed reconciliation equivariance, decoding the sign-relabeled LLRs with this relabeled syndrome outputs the correspondingly relabeled reconciled binary string and preserves the decoding success or failure.

-SEC λ |p0 − p1 | ≤ AdvSC SC,B1 (1 ). H2 : Similar to H1 , H2 further applies the same transformation to the acquisition path between P′ and U ′ . Thus, there exists an adversary B2 such that

Therefore, for either QKD mode, knowing z permits the masked reconciliation execution to be mapped efficiently and reversibly to the corresponding unmasked execution. The transformation changes neither the reconciliation acceptance behavior nor its leakage bound. □ Instantiation to the evaluated CV-QKD pipeline. The CVQKD implementation in Section 6 uses a binary reference string, LLR-based belief-propagation decoding, and linear LDPC syndromes. Hence its masking transformation follows the CV instance of Lemma 2. The LLR sign change exchanges the two binary likelihood labels without changing their magnitudes, and the syndrome changes only by the public offset. Correctness-verification messages and the resulting acceptance bit are treated as part of the public reconciliation transcript and are included in the leakage accounting. Privacy amplification is applied to the reconciled binary string. RAMP does not re-derive the finite-size security analysis of the underlying CV-QKD protocol. Provided the instantiated QKD protocol supplies a valid finite-size secrecy bound, Lemma 2 shows that the relabeling in RAMP preserves the reconciliation structure and leakage quantities to which that bound applies.

-SEC λ |p1 − p2 | ≤ AdvSC SC,B2 (1 ). We now analyze H2 . Condition Q guarantees that the target QKD execution satisfies the secrecy assumptions of the instantiated QKD protocol and that the provider-side state containing the target raw-key materials is not exposed to A . Consequently, immediately before masking, the secretbearing QKD state held by the users has the same secrecy guarantee as in the corresponding underlying QKD execution. Let AuxQKD denote the side information covered by the composable QKD secrecy definition, including the adversarial quantum side information and the public transcript of the underlying QKD classical processing. By the two secure-channel hybrids above, the acquisition traffic in H2 contributes no information about the protected QKD-derived contents beyond permitted public channel information, and the users’ accepted acquisition payloads equal the corresponding provider-supplied payloads except for the secure-channel failure probabilities already accounted for. The masking-compromise experiment then reveals both kAKE and z to A , thus no secrecy of AKE or SSC is used. By construction, the reconciliation inputs are fixed before the users execute AKE and derive z. Hence revealing the masking state does not reveal an additional function of the user-side secret QKD-derived inputs. For every resulting public value

Proof of Theorem 2. We prove it by a sequence of hybrid games. Let A be an admissible adversary in MCΠ,A (1λ ). For each hybrid Hi , let pi = Pr[A outputs b′ = b in Hi ]. 21

D.3

of z, Lemma 2 shows that the RAMP masking operation is an invertible relabeling of the reconciliation problem. It remains to verify that privacy amplification preserves the same security guarantee after this relabeling. Let τz (v) = v ⊕ z denote the relabeling of the reconciled binary reference string. Section 4.4 applies privacy amplification using a universal hash family. Let this family be H = {hs }s . Executing privacy amplification on the masked reconciled string is equivalent to applying hs ◦ τz to the corresponding unmasked string. For every fixed z, the composed family Hz = {hs ◦ τz }s remains universal. Specifically, for any distinct v, v′ ∈ {0, 1}n , τz (v) ̸= τz (v′ ), and therefore

The preceding theorems directly realize the confidentiality boundary defined in Section 3.2. Corollary 1 (Dual-Domain Confidentiality). RAMP realizes final-key confidentiality, provided that M ∨ (Q ∧ A) holds. If M holds, Theorem 1 applies even when the complete provider-side QKD and acquisition state is exposed. If M does not hold but both Q and A hold, Theorem 2 applies despite complete exposure of the masking state. The two cases rest on different security foundations. Under provider compromise, RAMP provides computational confidentiality through AKE, SSC, and privacy amplification. Under masking compromise, RAMP retains the secrecy guarantee supplied by the underlying QKD execution, provided the QKD-derived materials remain protected at the service boundary and during acquisition. RAMP makes no final-key confidentiality claim when ¬M ∧ (¬Q ∨ ¬A). This is the explicit confidentiality boundary of RAMP.

Pr[hs (τz (v)) = hs (τz (v′ ))] ≤ 2−ℓ . s

Thus, the public XOR relabeling does not weaken the privacy amplification. Together with Lemma 2, the masked execution has the same reconciliation leakage and the same privacyamplification security conditions as the corresponding unmasked QKD execution. For the LLR transformation in CV-QKD, the LLRs are auxiliary soft information used only to perform reconciliation, while the privacy amplification is applied to the resulting reconciled binary reference string. The sign transformation ensures that this binary string is exactly the relabeled version required by τz . Let Eqkd denote the adversarial side-information system covered by the composable QKD secrecy definition, including the quantum side information by A and the public transcript of the underlying QKD classical processing. Let ERAMP additionally include the public RAMP session context, the exposed masking state, and the post-masking public transcript. By public-mask invariance and the preservation of universality under the public relabeling τz , the RAMP post-processing execution has the same secrecy guarantee as the corresponding underlying QKD execution. We denote the subnormalized accepted states as acc ρacc KE = pacc ρ̄KE ,

E

QKD Testbed

The qOptica 100 platform implements a Gaussian-modulated coherent-state CV-QKD protocol with heterodyne detection and reverse reconciliation, consistent with practical longdistance CV-QKD systems [21]. The transmitter and receiver are connected through a deployed metropolitan fibre loop of 13.7 km, so the acquired raw-key material reflects transmission over real installed field fibre rather than a laboratory spool, back-to-back, or simulated channel. The one-way propagation delay over the 13.7 km loop (≈ 68 µs) and the syndrome-serialization latency are negligible relative to the user-side processing times reported. CV-QKD needs to specify classical side information, i.e., metadata, required for post-processing. This includes:

acc ρacc E = pacc ρ̄E

• Shot-noise calibration that is used to normalize quadrature measurements. • Detector parameters, e.g., detection efficiency ηB and trusted electronic noise vel .

Thus the RAMP key is εqkd -close to a uniformly random key independent of the adversary’s complete side information. Combining the prior hybrid with the QKD secrecy bound, we have

• Channel estimates, e.g., transmittance T and excess noise ξ, obtained during parameter estimation.

1 2

≤ |p0 − p1 | + |p1 − p2 | + p2 −

More on QKD Experiment

E.1

Hence, for the subnormalized accepted branch,
acc D ρacc KERAMP , τK ⊗ ρERAMP ≤ εqkd .

λ AdvMC Π,A (1 ) = p0 −

Dual-Domain Compromise Resilience

• Reconciliation configuration, e.g., LDPC code structure, code rate R, and block size.

1 2

• Protocol metadata, e.g., basis information (if applicable), block partitioning, and parameter estimation splits.

-SEC λ SC-SEC λ ≤ AdvSC SC,B1 (1 ) + AdvSC,B2 (1 ) + εqkd . Therefore, RAMP provides masking-compromise final-key confidentiality when Q ∧ A holds. □

The providers supply calibration and protocol metadata, including shot-noise normalization, detector parameters, code 22

configuration, and block-partition information. Securitycritical channel quantities such as transmittance T and excess noise ξ are estimated by the users locally during the parameter estimation rather than trusted as provider-supplied values. The authentication of provider metadata guarantees its origin and transport integrity, which does not imply correctness under provider corruption. The metadata is required for the construction of loglikelihood ratios (LLRs) and the correct operation of reconciliation and privacy amplification. In our implementation, quadrature samples are pre-normalized with respect to shot noise; the remaining parameters are still required to correctly model the effective channel and compute soft information.

• Both-quadrature parameter estimation. Channel parameters are estimated jointly over the amplitude and phase quadratures across the full parameter-estimation sample budget, tightening the excess-noise estimate and thereby the Holevo bound χE|B . • FFT-based privacy amplification. The Toeplitz hash is evaluated by blocked FFT, giving an O(n log n) extractor that scales to aggregated inputs of order 109 bits. Near-capacity reconciliation in practice still requires careful handling of LLR scaling, numerical stability, and decoder convergence, which we address in the implementation.

F E.2

Post-Processing Details

Efficient error correction in CV-QKD requires transforming continuous Gaussian variables into a form amenable to binary error-correcting codes. We employ multidimensional reconciliation using 8-dimensional (octonion-based) mappings, following Leverrier et al. [26], in which blocks of heterodyne measurements are mapped via orthogonal transformations to an equivalent binary-input additive white Gaussian noise (BiAWGN) channel. This produces soft information in the form of LLRs, which serve as input to a syndrome-based LDPC decoder. Error correction uses large, sparse quasi-cyclic LDPC codes of block length nH = 2 × 106 bits and rate R = 1/10, obtained from the qOptica QKD source files, with beliefpropagation decoding on soft inputs. The effective channel model incorporates shot noise (normalised to unity), trusted detector electronic noise, and excess noise estimated during parameter estimation, which together determine the SNR and reconciliation performance. The full pipeline explicitly realizes parameter estimation, multidimensional reconciliation, LDPC decoding, verification, and privacy amplification. We denote the user who processes data from the QKD transmitter (resp. receiver) by U (resp. U′ ). In RAMP, masking is applied prior to reconciliation by flipping both the raw bits of U′ and the corresponding LLRs of U using a shared keystream, ensuring that decoding performance is unaffected while hiding the masked keys from the providers. Correctness is enforced via Toeplitz-hash verification, and privacy amplification is performed after aggregating multiple successfully decoded blocks to mitigate finite-size effects. For reproducibility, we make explicit three implementation choices that determine the efficiency of the user-side pipeline:

Findings in Experiment: Asymmetry, Optimization, and Autonomy

The breakdown of the processing time shown in Table 1 provides two insights on performance. First, due to the asymmetric user-side workload, a practical deployment of RAMP can be strategically adjusted based on user-side capability to improve performance. Second, since the dominant operation is LDPC reconciliation, optimized implementations and acceleration from commodity GPUs [31] are expected to enable speedups without changing the workflow in RAMP. Our evaluation also highlights a feature of RAMP, i.e., rawkey acquisition and final-key distillation are decoupled. Users can autonomously decide when and how much raw keys to process, provided they follow the post-processing procedure.

F.1

Asymmetry in User-Side Processing Time

The processing cost of RAMP is asymmetric between the two users in our CV-QKD instantiation. This asymmetry comes from the structure of syndrome-based reconciliation. For masking in reverse reconciliation (Section 4.3), U′ holds the masked binary string rk′mask , while U holds the corresponding masked LLRs. U′ computes the syndrome as s = Hrk′mask mod 2, where H is the LDPC parity-check matrix. This operation is essentially a lightweight sparse matrix-vector multiplication. On the other hand, U performs the heavyweight part of reconciliation. It runs iterative LDPC decoding over soft inputs and verifies the decoded result. This finding has direct implications for deployment. RAMP does not require both users to have symmetric computational capability. When role assignment is configurable, a resourceconstrained device can participate in the syndrome-generation side, while a more capable endpoint handles the decodingintensive reconciliation. Such a role assignment is particularly relevant in service-oriented settings, where users may operate laptops, mobile devices, or embedded IoT devices with heterogeneous computational resources.

• Vectorised belief propagation. The decoder operates on the Tanner-graph edge list using vectorised check- and variable-node updates, rather than scalar per-node loops. This removes the dominant cost of the reference decoder while reproducing identical decoding behaviour (and hence identical FER). 23

Therefore, the practicality of RAMP should be understood not only from the total processing time, but also from the per-user workload distribution. The computation is concentrated on one side, while masking and syndrome generation remain lightweight. It provides a useful deployment insight, i.e., RAMP can support heterogeneous users by assigning the reconciliation-heavy role to the more capable endpoint.

F.2

Such deferred processing requires appropriate local protection of the acquired raw-key materials. Users should store raw-key blocks together with their metadata, key identifiers, and state information, and protect them using standard authenticated encryption or quantum-safe storage mechanisms where appropriate. The raw-key state must also remain consistent with the provider-side lifecycle, so that each block is used only once and eventually marked as consumed or expired. From the perspective of service providers, RAMP removes reconciliation and privacy amplification from the providers (cf. partial delegation in prior works [28, 40]). Service providers can concentrate on operating QKD devices, maintaining rawkey inventories, and serving acquisition requests, rather than performing reconciliation and privacy amplification for every user session. In this way, service providers produce and manage QKD-derived materials as a backend resource, while users consume and process them on demand. Therefore, RAMP does not merely offload computation from providers to users. It upgrades the operational model of QKD, where raw-key generation, raw-key acquisition, and final-key distillation no longer need to occur as one tightly coupled procedure. This decoupling is central to making QKD-derived entropy available as a flexible service for heterogeneous users.

Optimization with Commodity Hardware

Table 1 identifies LDPC reconciliation as the bottleneck of our prototype, which is introduced by standard CV-QKD postprocessing. Since our implementation does not utilize hardware acceleration, the result should be viewed as a conservative CPU baseline rather than an inherent limit of RAMP. Prior work on CV-QKD reconciliation [31] has shown that LDPC decoding is highly amenable to parallel acceleration on commodity GPUs (with a single NVIDIA GeForce GTX 1080 GPU in their implementation), especially when the code structure is designed to improve memory access and messagepassing efficiency. More recent work [52] further suggests GPU-accelerated rate-adaptive reconciliation for CV-QKD by optimizing the storage of the parity-check matrix. These observations indicate that the dominant cost observed in our evaluation is an optimizable implementation bottleneck of reconciliation, rather than a protocol-level overhead. More specifically, a cost-effective method toward overall performance improvement is to assign the decoding-heavy role identified in Section F.1 to an endpoint with commodity computing resources like a GPU. The other endpoint could remain on the lightweight side. Note that the optimization remains independent of specialized QKD hardware. With commodity GPUs for QKD postprocessing available, end users could configure and consume the service with general-purpose computing platforms.

G

Use of Generative AI

The main body of this paper, including the conceptual design, technical construction, analyses, evaluation, and novel intellectual contributions, was produced by the authors. Generative AI tools were used in supporting roles. ChatGPT was used to assist with language polishing. ChatGPT and Claude helped build the prototype and fix bugs during development. All suggested fixes from generative AI were manually reviewed, tested, and validated by the authors before being included. Grammarly was used to check grammar and improve the presentation of the text. The authors take full responsibility for the accuracy, correctness, and integrity of all material in the paper.

F.3 On-Demand Key Consumption and User Autonomy In conventional QKD deployments, the parties that benefit from the final keys are also involved in the quantum-state generation and transmission. In RAMP, service providers generate and maintain raw-key materials, while users decide when and how much of the acquired materials to process into final keys. This separation gives users greater autonomy. A user can acquire raw-key materials when the service is available and defer post-processing until local computation, energy, or application demand makes it appropriate. The amount of material processed in each session can also be adjusted according to the available resources and the desired key length, as long as it satisfies the conditions of QKD post-processing. For example, the users could process smaller raw-key batches in a resource-constrained device, while aggregating larger batches when strong computation resources are available. 24

Related documents

• Quantum computing service with quality of service (QoS) enforcement via out-of- … — Amazon Technologies, Inc. (US12430170B2)#607111
• Interconnecting Regional QKD Networks: Hybrid Key Delivery Across Quantum Domains#974390
• The SRP Authentication and Key Exchange System#134916
• A hybrid encryption algorithm based on AES and Kyber for secure image protection in the post-quantum era#1226471
• QSpy: A Quantum RAT for Circuit Spying and IP Theft#1120047