Multi-Factor Trust-Driven Secure Communication Model for Cloud-Based Digital Twins

Unknown

Conceptio › Archive › arXiv CS

arXiv CSopen access

Multi-Factor Trust-Driven Secure Communication Model for Cloud-Based Digital Twins

Unknown · 2026 · arxiv_cs

arXiv CS · Papers · License: Open Access · 2026

Open Source ↗Direct PDF ↓

cloud distributedcomputing parallelcomputing

distributed computing, parallel computing, cloud

Multi-Factor Trust-Driven Secure Communication Model for Cloud-Based Digital Twins

arXiv:2605.23566v1 [cs.DC] 22 May 2026

Deepika Saxena, Senior Member, IEEE and Ashutosh Kumar Singh, Senior Member, IEEE

Abstract— Cloud-based Digital Twin (DT) platforms enable real-time monitoring, simulation, and collaborative decision-making across distributed clients. However, ensuring secure and trustworthy communication remains a critical challenge due to heterogeneous client behavior, resource contention, and evolving adversarial threats. This paper proposes the Multi-Factor Trust-Driven Secure Communication (MT-SeCom) framework to enforce resilient and intelligent collaboration in DT-enabled cloud environments. MT-SeCom operates through four coordinated phases: (i) Multi-Factor Trust Monitoring, capturing temporal, contextual, and federated trust signals; (ii) Adaptive Trust Evaluation, adjusting trust weights based on network dynamics and threat intensity; (iii) Transformer-Based Trusted Client Classification, combining anomaly detection with supervised learning to accurately identify malicious or unreliable nodes; and (iv) Resilient Communication Management, optimizing routing, isolating compromised clients, and ensuring service continuity. A real-world testbed and comprehensive experiments demonstrate that MT-SeCom significantly enhances secure communication, mitigates cascading adversarial effects, and maintains high resilience under fluctuating attack conditions. MT-SeCom achieves an average 18.7% improvement in threat detection accuracy and a 24.3% reduction in anomaly occurrences compared to existing methods, confirming its robustness, scalability, and practical suitability for heterogeneous cloud-based DT ecosystems. Index Terms— Digital Twin, Cloud Security, Trust Management, Resilient Communication, Transformer-based Anomaly Detection

I. I NTRODUCTION IGITAL Twin (DT) technology has emerged as a cornerstone of next-generation cyber–physical systems, enabling applications such as predictive healthcare, intelligent transportation, smart energy grids, and industrial IoT ecosystems [1]. By creating dynamic virtual replicas of physical assets, DTs enhance operational resilience, performance, and resource efficiency. However, reliance on cloud infrastructures exposes these systems to significant cybersecurity risks, as

Deepika Saxena is with the Division of Information Systems, University of Aizu, Japan and also with Department of Computer Science, Vizja University, 01-043 Warsaw, Poland. (Email: [email protected], [email protected]). Ashutosh Kumar Singh is with the Department of Computer Science and Engineering, Indian Institute of Information Technology Bhopal, Bhopal 462003, India, and also with Department of Computer Science, the Vizja University, 01-043 Warsaw, Poland. (E-mail: [email protected]).

sensitive data exchanged between physical and digital counterparts can be intercepted, manipulated, or misused. Recent reports indicate that in the first half of 2025, healthcare organizations experienced 107 cyberattacks compromising over 1.6 million patient records, with Microsoft 365 as a common vulnerability [2]. The manufacturing sector accounted for 22% of sector-attributed attacks, with ransomware incidents rising sharply [3]. Such threats highlight the critical need for robust, trust-aware, and resilient communication mechanisms. Existing solutions typically address security, anomaly detection, or trust management separately, leaving a gap in unified frameworks capable of ensuring end-to-end trustworthy communication among heterogeneous DT clients. This motivates the development of a multi-factor trust-driven approach that dynamically evaluates trust while maintaining secure, reliable, and adaptive communication in cloud-based DT networks. A. Related Work Recent research has explored the integration of Digital Twins (DTs) with advanced communication, security, and learning techniques to enhance performance and resilience across diverse domains. Lu et al. [4] proposed digital twin edge networks (DITENs) that leverage federated learning and asynchronous updates to reduce communication overhead. Tang et al. [5] developed a blockchain-based healthcare DT framework employing attribute-based encryption and missing value prediction to ensure privacy and data integrity. Soula et al. [6] introduced a DT-based trust model for IoT anomaly detection, improving scalability and energy efficiency. Other efforts have focused on blockchain-enabled resource monitoring [7], [8], trust prediction in Internet of Vehicles (IoV) using Generative Adversarial Networks (GANs) [9], and resilient IoT/IoV-based DT frameworks for critical cyber-physical systems [10], [11]. Bera et al. [12] examined DT-driven healthcare communication security in 5G/B5G networks, addressing slice isolation and device authentication. Wang et al. [13] provided a comprehensive survey of the Internet of Digital Twins (IoDT), detailing key architectural, communication, and security challenges. Mrabet et al. [14] have proposed multi-factor trust and resilience models that integrate anomaly detection, blockchain-based trust, privacy-preserving analytics, and secure virtualization to strengthen edge security in smartcity environments. Collectively, these studies demonstrate that DTs can enable intelligent, secure, and efficient cyber-physical systems. However, most solutions address trust, privacy, or resilience in isolation and lack a unified framework for adaptive,

© 2026 IEEE. This article has been accepted in IEEE Transactions on Industrial Informatics Journal © 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. This work is freely available for survey and citation.

end-to-end trust-aware communication across heterogeneous DT clients. B. Research Gaps and Contributions Despite advances in DT security and trust frameworks [4], [5], [8], [9], [6], [10], [11], several gaps remain: • Most trust management approaches focus on a single evaluation dimension, such as behavior monitoring, peer reputation, or cryptographic validation, which leads to an incomplete trust representation in DT communication. • Only a few frameworks adjust trust scores in real time based on environmental changes or adversarial activities, often resulting in delayed or inaccurate trust decisions under hostile conditions. • Trust evaluation in existing methods is often disconnected from communication resilience mechanisms, making DT systems more vulnerable to coordinated attacks. To address these gaps, the Multi-Factor Trust-Driven Secure Communication (MT-SeCom) model is proposed. Its main contributions include: • The proposed framework unifies multiple trust dimensions by integrating temporal, contextual, and federated trust signals, enabling comprehensive evaluation of client reliability under dynamic and adversarial conditions. • An adaptive trust evaluation mechanism employs a Transformer-based classifier to dynamically adjust the weighting of trust factors, combining anomaly detection with supervised learning to support robust and contextaware decision-making. • The resilient communication management scheme constructs optimized, attack-aware network topologies that maximize link reliability, minimize latency, and isolate malicious clients, thereby ensuring secure and dependable communication across cloud-based DT networks. MT-SeCom provides a cohesive and adaptive trust-driven communication framework, unifying multi-factor trust evaluation, anomaly detection, and resilient link management to ensure end-to-end security, reliability, and performance in cloudbased DT ecosystems. C. Paper Organization Section II presents the problem formulation including the system model, cyberthreat model, and problem statement with design goals. Section III details the proposed MT-SeCom framework, covering multifactor trust monitoring, adaptive trust evaluation, transformer-based client classification, and resilient communication management. Section IV describes the operational workflow, algorithmic summary, and complexity analysis of the proposed model. Section V presents the experimental setup, performance evaluation, comparative analysis, and security assessment of MT-SeCom. Finally, Section VI concludes the study and outlines directions for future research. II. P ROBLEM F ORMULATION This research addresses the challenge of securing collaborative cloud-based Digital Twin (DT) applications against malicious clients and unreliable behaviors while ensuring resilient

communication and trustworthy execution of component tasks. The problem is formalized through: System Model (Section IIA) defining participating entities and their roles; Cyberthreat Model (Section II-B) specifying potential adversarial threats; and Problem Statement and Design Goals (Section II-C) presenting the optimization objectives and design constraints. A. System Model As illustrated in Fig. 1, consider a system model that enables the execution of Digital Twin Applications at the cloud platform. This system comprises six interdependent entities: Clients (C), Collaboration Applications (A), Virtual Nodes (VN), Physical Nodes (PN), and the Cloud Platform (CP), defined as: Security Threats

Challenges Adversarial Threats Behavioural Dynamics Cascading Anomalies Trust Resilience

Constraints

Model Poisoning Threat

Digital Twin Application

Research Design Objectives Resource abuse Threat

benign VN

benign client

cyberthreat flow

benign sub-application

malicious VN

hidden malicious client

normal operation flow

malicious sub-application

Fig. 1. System model and problem illustration

Clients (C): The clients represent operators executing DT applications, C = {C1 , . . . , Cn }. There are two types: – Benign clients (C B ): Legitimate participants supporting collaboration. – Malicious clients (C M ): Insider adversaries that mimic benign behavior to inject malicious workloads or compromise results. • Collaboration DT Applications (A): Each DT application comprises z sub-applications, A = {A1 , . . . , Az }, executed by specific clients. Sub-applications are classified as benign or malicious based on their behavior. • Virtual Nodes (VN): The logical execution environments hosting sub-applications operate as virtual nodes, enabling scalability, isolation, and controlled resource allocation. They can be benign (executing legitimate workloads) or malicious (hosting adversarial sub-applications). • Physical Nodes (PN). These are physical servers that consolidate multiple virtual nodes while offering computation, storage, and networking resources required to execute Digital Twin (DT) workloads. • Cloud Platform (CP): This platform orchestrates virtual and physical resources, ensuring elastic, resilient, and isolated execution. It supports both normal operations and propagation of cyberthreat flows.

•

B. Cyberthreat Model Consider n clients {C1 , . . . , Cn } executing sub-applications {A1 , . . . , Az } on Q virtual nodes {V N1 , . . . , V NQ } hosted on P physical nodes {P N1 , . . . , P NP }, as shown in Fig. 1. A subset of hidden adversaries, denoted C M , may control one or more malicious virtual nodes {V NjM } with the objective of compromising benign components. The execution of subapplication Ak on V Nj running on P Ni is represented by Ωijk : P Ni × V Nj × Ak and models a workload instance with observable timing, resource, and communication characteristics. Unless stated otherwise, threat indicators lie in [0, 1], with larger values indicating higher impact. This cyberthreat model comprises the following threats, and Table I summarizes the notation used throughout the model. TABLE I N OTATION USED IN THE CYBERTHREAT MODEL Symbol P Ni V Nj Ak Ωijk Ξx x Vijk ∆tijk Rijk Txfederated (t) Γ(t) Φ(·)

Meaning physical node identifier or host virtual node running on P Ni sub-application or task execution instance of (i, j, k) threat impact score for threat x vulnerability factor for threat x timing deviation from nominal execution resource contention level federated trust of client x at time t aggregation weighting state aggregation function producing manipulation score

Range / Units integer index integer index integer index dimensionless [0, 1] [0, 1] seconds normalized, [0, 1] [0, 1] [0, 1] [0, 1]

Definition 1 (Model Poisoning Threat): Model poisoning occurs when a malicious virtual node injects corrupted parameters or manipulated gradients into collaborative learning or control updates. A malicious node V Nĵ targets sub-task Ak̂ and contaminates the execution of Ak running on V Nj at P Ni . The resulting impact is represented in Eq. (1): P MP ΞM = Ωijk · Ωiĵ k̂ · P Ni · Vijk , (1) iĵ k̂→ijk MP where Vijk ∈ [0, 1] quantifies the susceptibility of the execution pipeline to poisoning. Larger values indicate weak authentication, insecure update validation, or insufficient isolation. Definition 2 (Synchronization Disruption Threat): Synchronization disruption captures situations where malicious virtual nodes intentionally delay, reorder, or accelerate task updates, which creates temporal inconsistencies among dependent processes. The threat is modeled in Eq. (2): SD ΞSD = Ωijk · Ωiĵ k̂ · ∆tijk · Vijk , iĵ k̂→ijk

(2)

where ∆tijk measures the deviation from nominal execution SD ∈ [0, 1] expresses the degree to time (in seconds), and Vijk which such deviations can destabilize coupled DT processes. Definition 3 (Resource Abuse Threat): Resource abuse refers to adversarial behavior that results in abnormal consumption of CPU, memory, or network bandwidth, degrading normal system performance. The effect is captured in Eq. (3): RA ΞRA = Ωijk · Ωiĵ k̂ · Rijk · Vijk , iĵ k̂→ijk

(3)

where Rijk ∈ [0, 1] measures normalized contention level and RA Vijk ∈ [0, 1] reflects weaknesses in resource isolation that allow interference across co-located tasks.

Definition 4 (Anomalous Behavior Propagation): Anomalous behavior propagation describes cascading effects in which local compromises produce secondary failures in dependent applications or nodes. The propagation is represented in Eq. (4): X ΞABP = Ξxiĵ k̂→ijk ·Ωi∗ j ∗ k∗ ·P Ni∗ ·Vi∗ j ∗ k∗ , iĵ k̂→i∗ j ∗ k∗ x∈{M P,SD,RA}

(4) where Vi∗ j ∗ k∗ ∈ [0, 1] determines how easily local anomalies propagate to downstream tasks. Definition 5 (APT and Collusion-based Trust Manipulation): Advanced persistent adversaries gradually accumulate high trust and later exploit it during critical operations. A malicious client Cî , collaborating with a colluding set C, manipulates federated trust aggregation as: T ΞAP (t) = Φ Tîfederated (t), {Tcfederated (t)}c∈C , Γ(t) , (5) î,C where Φ(·) denotes a robust operator (trimmed-mean or median). The trimmed-mean removes the highest and lowest δ% values before averaging, while the median suppresses outliers. Resilience holds if the colluding fraction is below δ (or below 50% for median aggregation). Here, Txfederated (t) ∈ [0, 1] is the trust of client x, C is the colluding set, Γ(t) ∈ [0, 1] conT (t) ∈ [0, 1] denotes the manipulation trols weights, and ΞAP î,C magnitude. C. Problem Statement with Design Goals The objective is to quantify multi-factor trust for each client and sub-application, composed of temporal (Titemporal ), contextual (Ticontext ), and federated (Tifederated ) trust, while minimizing untrusted entities and ensuring secure, resilient communication. This is formulated as follows in Eq. (6): MAX

Ctrusted ,Lcomm

subject to

αTitemporal + βTicontext + γTifederated

i∈C

Titemporal , Ticontext , Tifederated ∈ [0, 1], Lcomm (i, j) ≥ Lmin ,

∀i, j ∈ Ctrusted

Lcomm (i, j) ≤ Llatency , max

Ctrusted ⊆ C,

∀i ∈ C

∀i, j ∈ Ctrusted

Lcomm ⊆ C × C.

(6) Here, Ctrusted represents the set of trusted clients, Lcomm represents communication links among them, and α, β, γ weight the trust dimensions. Communication links satisfy reliability and latency thresholds Lmin and Llatency , while max privacy and workload constraints ensure secure, heterogeneous DT execution. Design Constraints: The framework is required to operate under practical DT communication constraints. Trust evaluation must be lightweight to support real-time decisions, while communication paths must satisfy strict latency and reliability bounds. The design should accommodate heterogeneous workloads and diverse DT applications without degrading performance, and must preserve the privacy of benign clients throughout trust computation and communication.

Design Goals:

To enable accurate multi-dimensional trust evaluation for clients and tasks in DT environments. • To proactively detect and isolate malicious nodes to prevent system disruption or task compromise. • To ensure resilient, efficient communication among trusted entities while remaining scalable and generalizable across cloud-based DT ecosystems. •

considered are r ∈{Output Integrity, Timeliness, Resource Usage, Reliability, Privacy Compliance}. The representative formulations are: (Output Integrity)

(Timeliness)

(Resource Usage)

III. P ROPOSED M ODEL This section presents the Multi-Factor Trust-Driven Secure Communication (MT-SeCom) model for secure, resilient, and optimized execution of cloud-based DT applications. MTSeCom jointly evaluates trust, classifies clients, and adapts routing so that malicious or unstable clients are isolated while preserving performance. As illustrated in Fig. 2, MT-SeCom enables trustworthy collaboration by operating in four phases: (i) Multi-factor Trust Monitoring, (ii) Adaptive Trust Evaluation, (iii) Trusted Client Classification, and (iv) Resilient Communication Management. A. Trust Monitoring Layer Let n clients {C1 , ..., Cn } execute sub-applications: {A1 , ..., An } on Q virtual nodes {V N1 , ..., V NQ } hosted on P physical nodes {P N1 , ..., P NP }. For each client Ci , multifactor trust signals are partitioned into three complementary components including temporal trust, contextual trust, and federated trust defined as follows: Definition 6 (Temporal Trust): This metric captures how a client’s performance evolves over time, penalizing negative drifts. Let Si (t) ∈ Rm denote the m-dimensional vector of normalized performance signals for client i at time t, including performance accuracy (P AC), mission completion rate (M CR), average resource consumption (ARC), and previous violation counts (P V C). The temporal trust (Titemporal ) of client i is defined using Eq. (7): (S) Titemporal (t) = exp − λ ∆S⊤ , (7) i (t)w with ∆Si (t) = Si (t) − Si (t − 1). Here, λ > 0 is a decay factor, and w(S) is a non-negative weight vector satisfying 1⊤ w(S) = 1, reflecting the relative importance of each signal. A typical instantiation is Si (t) = ⊤ P ACi (t), M CRi (t), ARCi (t), P V Ci (t) , where, all elements are normalized to [0, 1]. For negative indicators (e.g., P V C), normalization is inverted so that higher values consistently indicate better performance. Definition 7 (Contextual Trust): Contextual trust quantifies how reliable a client i is under the prevailing operating conditions. It aggregates multiple behavioral indicators capturing stability, correctness, and communication quality. The contextual trust score at time t is defined as Eq. (8): ! X (r) context Ti (t) = σ αr Fi (t) , (8) r (r) Fi (t)

where ∈ [0, 1] is the normalized contribution of P factor r, αr ≥ 0 is its weight, and r αr = 1. The factors

(Reliability)

(Privacy Compliance)

(t) = 1 − ErrRatei (t), (t) = e−β∆ti , Usagei − Qi , 0, 1 , (t) = 1 − clip Qi (t) =

PDRi , 1 + Jitteri

(t) = 1 − LeakageScorei .

Here, σ(·) is the logistic normalization that keeps Ticontext (t) ∈ (0, 1). ErrRatei (t) denotes the normalized error (e.g., accuracy loss or deviation) in the outputs of client i. ∆ti is the reporting delay and β > 0 controls its penalty. Usagei is the observed resource consumption relative to the allocated quota Qi and is bounded by clip(·) to [0, 1]. Network reliability is captured through the packet delivery ratio PDRi and latency variability Jitteri . LeakageScorei measures the likelihood of privacy leakage. Larger values of Ticontext (t) therefore indicate more stable, efficient, and trustworthy behavior in the current context. Definition 8 (Federated Trust): Federated trust mitigates collusion and advanced persistent threats by combining peer feedback with robustness and temporal decay. Let Ni (t) denote the peer set of client i at time t as depicted in Eq. (9): Tifederated (t) = RobAgg { w̃j→i (t) sj→i (t) | j ∈ Ni (t) } , (9) where sj→i (t) ∈ [0, 1] is the feedback from peer j, and RobAgg(·) denotes a trimmed-mean or median operator that suppresses extreme opinions and reduces collusion. The credibility weight is given by Eq. (10): w̃j→i (t) = ηe−ρ∆tj + (1 − η)

1 , 1 + Var(Tj (1:t))

(10)

where e−ρ∆tj enforces time decay so that recent behavior contributes more strongly, and Var(Tj (1:t)) penalizes peers exhibiting unstable trust trajectories. The parameter η ∈ [0, 1] controls the balance between recency and stability. To ensure that peer opinions do not overshadow direct evidence, Tifederated (t) is bounded as Eq. (11): Tifederated (t) ← min Tifederated (t), κ1 Titemporal (t)+κ2 Ticontext (t) (11) with κ1 + κ2 = 1. This design limits gradual trust accumulation, attenuates colluding peers, and prevents federated trust from overriding temporal or contextual evidence. MT-SeCom updates trust incrementally using sliding windows, retaining only recent telemetry. The per-client update cost remains O(m), where m denotes the number of monitored factors. This incremental design enables continuous trust estimation with low memory and computational overhead, even at large scale. To defend against collusion and gradual trust inflation, federated aggregation combines (i) robust statistics

Temporal Trust Adaptive Fusion Engine reweighted dynamically

Outputs (Shifted Right)

Contextual Trust

Input Embedding

Output Embedding

Positional Encoding

Federated Trust Multi-Head Attention

Masked Multi-Head Attention

Add & Norm

Final Trust Score Temporal Trust

Multi-Head Attention

Feed Forward

Adaptive Fusion Engine reweighted dynamically

Add & Norm

Contextual Trust

DT Collaborative Communication Virtual Network (DTCCVN)

Computation of optimized links by selecting reliable links only

Feed Forward Add & Norm

Linear

Federated Trust

Softmax (Output Probabilities)

Final Trust Score

Eliminate Non-Trusted Clients

Temporal Trust Adaptive Fusion Engine reweighted dynamically Masked Multihead attention

Contextual Trust Encoder self-attention: tokens look at each other. Queries, keys, values are computed from encoded states

Multi-head attention Add & Norm Feed Forward

Federated Trust Feed-forward network: processes each token’s context to refine its representation.

Final Trust Score Multi-factor TRUST Monitoring Layer

Adaptive Trust Evaluation Engine Layer

Add & Norm Multi-head attention Add & Norm

Add & Norm

Feed Forward

Residual connections and layer normalization

Softmax (output probabilities)

Decoder-self attention (masked): each token attends to only earlier tokens to prevent seeing the future. Decoder-encoder attention: target tokens looks at the source queries-from decoder state; keys and values from encoder states

Add & Norm Linear

Feed-forward network: processes each token’s context to refine its representation.

Multi-factor Trusted Client Evaluation Layer

Resulted Optimized Resilient Communication Links Network

Resilient Communication Management Layer

Fig. 2. Proposed Model

(trimmed-mean or median), (ii) credibility-aware weighting with temporal decay and variance penalization, and (iii) an upper bound that constrains federated trust by temporal and contextual evidence. The aggregation remains resilient provided the fraction of colluding peers does not exceed the robustness tolerance (e.g., trimming parameter in trimmed-mean or 50% in median aggregation). As a result, malicious feedback cannot dominate aggregation unless adversarial participation exceeds the robustness tolerance threshold. B. Adaptive Multi-Factor Trust Evaluation The multi-factor trust score for client P Ci is computed 3 dynamically as Eq. (12), where γp ≥ 0 and p=1 γp = 1. Ti (t) = γ1 Titemporal (t) + γ2 Ticontext (t) + γ3 Tifederated (t), (12) The trust engine continuously updates Ti (t) in real time. Temporal and contextual trust values are derived from performance and environmental signals, while federated trust is aggregated from peer feedback. A feedback controller adaptively adjusts (γ1 , γ2 , γ3 ): contextual trust is emphasized under network instability, temporal trust dominates for longterm reliability, and federated trust is prioritized in uncertain or adversarial conditions based on observed violation rates, link instability, and anomaly frequency. Because each trust component is updated locally and independently, the trust engine aggregates compact summaries rather than full historical traces. This reduces recomputation overhead and allows Ti (t)

to scale efficiently with both the number of clients and the event rate. C. Trusted Client Classification Client trust is inferred using a Transformer-based classifier Fθ that maps the multi-factor signal vector Si to a binary decision (Eq. 13): ( 1, if client i is trusted, yi = Fθ (Si ) ∈ {0, 1}, yi = 0, if client i is malicious. (13) The Transformer-based classifier operates on compact multi-factor trust features instead of raw telemetry streams. This keeps inference linear in sequence length and easily parallelizable across clients, enabling MT-SeCom to sustain high event rates without introducing noticeable latency in the decision pipeline. The Transformer encoder models temporal and contextual dependencies in Si , yielding an embedding hi = Transformerθ (Si ). Classification integrates supervised prediction and anomaly detection via Eq. (14): ( 1, fanom (hi ) < τ and fsup (hi ) = 1, yi = (14) 0, otherwise, where fanom (·) denotes the anomaly score, fsup (·) the supervised classifier output, and τ the detection threshold. This joint decision rule reduces false acceptance of anomalous clients while preserving sensitivity to legitimate behavior.

To exploit temporal trust dynamics, the Transformer encoder applies multi-head self-attention over trust sequences. Unlike recurrent models, self-attention avoids stepwise state propagation and captures long-range dependencies without vanishing gradients. The mechanism emphasizes salient trust events (e.g., performance drifts, policy violations, reputation drops) while suppressing benign fluctuations. By jointly encoding temporal, contextual, and federated indicators, the model learns their interactions and detects patterns formed by multiple weak signals, yielding a more discriminative representation of client behavior and more reliable trust decisions in MT-SeCom. D. Resilient Communication Management

IV. O PERATIONAL D ESIGN AND C OMPLEXITY A NALYSIS The operational workflow of MT-SeCom is summarized in Algorithm 1. Each phase is designed to operate incrementally and in parallel, thereby enabling scalable execution under high client populations and event rates. Algorithm 1: Operational Summary of MT-SeCom /* Multi-Factor Trust Monitoring for each client Ci do 2 Update temporal trust Titemporal using Eq. (7) 3 Update contextual trust Ticontext using Eq. (8) 4 Update federated trust Tifederated via Eq. (9)

/* Adaptive Trust Evaluation */ The Resilient Communication Manager (RCM) maintains 5 for each client Ci do secure and efficient connectivity among trusted DT clients by 6 Compute overall trust: adapting the communication graph network G(V, E), where Ti = γ1 Titemporal + γ2 Ticontext + γ3 Tifederated V denotes clients with yi = 1 and E represents active links. 7 Adapt (γ1 , γ2 , γ3 ) based on violation rate, At each decision step, the RCM derives a resilient subgraph instability, and anomaly trends ′ ′ G = (V \ Vmalicious , E ) by balancing reliability and latency / Trusted Client Classification * */ while isolating links associated with suspicious nodes. The 8 for each client C do i resulting link set satisfies using Eqs. (15) and (16): 9 Compute embedding hi = Transformerθ (Si ) X X Classify using Eq. (14) link_reliability(i, j)−β latency(i, j),10 E ′ = arg max E⊆G

(i,j)∈E

/* Resilient Communication Management */ subject to 11 Construct trusted set Vtrusted = {Ci | yi = 1} malicious nodes and links to obtain sub-network (i, j) ∈ / E ∀(i, j) ∈ Emalicious , L̄(E) ≤ Lmax , Res(E) ≥ τres .12 Prune ′ G (16) ′ Because solving this optimization exactly under dynamic 13 Optimize E under resilience/latency constraints using RL policy conditions is expensive, MT-SeCom approximates it using ′ ′ 14 Deploy updated topology G = (Vtrusted , E ) reinforcement learning. The system state s summarizes the (15)

current network context, including trust scores, link reliabilities, queue backlogs, and average path delay. An action at modifies the active communication network by activating, deactivating, or rerouting selected trusted links. The reward function encourages resilient yet efficient routing and is given by Eq. (17): rt = α1 Res(Et ) − α2 L̄(Et ) − α3 penaltymal (Et ),

(17)

where, the resilience term reflects path survivability under failures, the latency term penalizes delay, and the final term discourages routes traversing low-trust nodes. Learning is implemented using a model-free actor–critic scheme that updates the policy from transitions (st , at , rt , st+1 ). Training continues until the moving average of cumulative reward stabilizes within a tolerance window, indicating convergence. During operation, the policy adapts gradually as traffic conditions and trust dynamics evolve, enabling the RCM to reduce exposure to unreliable links while maintaining communication constraints. To further reduce computational cost, the RCM operates only on the pruned network consisting of trusted clients. As a result, routing optimization scales with the size of the active trusted sub-network (approximately O(|E ′ | log |V |)) rather than the entire network. This network-pruning strategy substantially decreases routing complexity while preserving resilience guarantees.

Complexity Analysis MT-SeCom is designed for scalable operation through incremental updates, compact feature processing, and trusted network optimization. Trust metrics are computed locally using sliding windows, requiring only recent telemetry and yielding O(n · m) time and O(n) space, where n is the number of clients and m the monitored signals. Adaptive weighting introduces only O(n) overhead. The Transformerbased classifier operates on short trust sequences, resulting in O(n · L) inference cost with O(n · d) memory. Communication optimization is performed only over the pruned trusted subgraph G ′ = (V ′ , E ′ ), with routing and refinement scaling as O(|E ′ | log |V ′ |) and reinforcement-learning updates as O(|E ′ | · TRL ). Overall, MT-SeCom achieves low computational and memory overhead while supporting large client populations and high event rates. V. P ERFORMANCE E VALUATION A. Experimental Set-up The simulation experiments are conducted on a highperformance server equipped with dual Intel® Xeon® Silver

4114 CPUs (40 cores, 2.20 GHz). The setup synthesizes realistic VM time-series data by concatenating multiple perVM CSV traces (stored in GCD_VMs) [15], each containing CPU and memory utilization values normalized and windowed into sequences of length SEQ_LEN for Transformerbased anomaly prediction. To train the Transformer classifier, ground-truth labels are constructed automatically during attack injection. Each VM sequence is labeled as trusted when it evolves under normal conditions, and malicious when the simulate_attack routine injects abnormal behavior. The classifier is therefore trained in a supervised manner on binary labels {0, 1}. The dataset is split into nonoverlapping training, validation, and test partitions using a 70 : 15 : 15 protocol, while ensuring that sequences originating from the same VM do not appear across different splits. The Transformer architecture (D_MODEL, NHEAD, NUM_LAYERS) is optimized using the Adam optimizer with learning rate LR. Binary cross-entropy is used as the classification loss, and model selection is guided by validation loss with early stopping patience of five epochs to prevent overfitting. For anomaly prediction, reconstruction error is trained using MSE loss, followed by a median + Median Absolute Deviation (MAD) thresholding rule for anomaly scoring. Thresholds are fixed across experiments for reproducibility. Table II summarizes the experimental configuration. VM-to-PM assignment is performed via allocate_vms_pm, parameterized by VM_COUNT and PM_COUNT, while system metadata (for example Model_Drift, PDR, CPU_cap) is generated using a fixed seed for reproducibility. Attacks are injected using simulate_attack with a Bernoulli distribution governed by ATTACK_RATE, resulting in stochastic injection at random time steps during operation. When triggered, the injected attack persists for a configurable duration window (ATTACK_DURATION), thereby modeling sustained malicious behavior rather than single isolated spikes. This design enables evaluation under both intermittent and continuous adversarial conditions. Trust metrics (Contextual, Temporal, Federated, and Adaptive) are computed to emulate dynamic trustaware decision-making, where VMs with Adaptive_Trust > 0.5 guide a greedy optimizer for selecting communication paths under latency and resource constraints. TABLE II E XPERIMENTAL PARAMETERS AND THEIR VALUES Parameter VM_COUNT PM_COUNT SEQ_LEN ATTACK_RATE RNG Seed Transformer architecture Training hyperparameters Thresholding rule Link selection constraints

Default / Range 100-1600 VNs 30-500 10 0.05 42 D_MODEL=64, NHEAD=4, NUM_LAYERS=2 BATCH_SIZE=128, EPOCHS=5, LR=1e-3 median + 3*Median Absolute Deviation (MAD) L_max = 150, τres = 0.1

B. MT-SeCom Results The performance metrics in Table III demonstrate the scalability and robustness of the proposed MT-SeCom framework

under increasing cloud–DT workloads. Trust evaluation accuracy remains above 93% and peaks at 95.20% for 800 VNs, while the mean error stays extremely low (< 7 × 10−4 ). This accuracy–error consistency indicates the effectiveness of the adaptive multi-factor trust computation (Eq. (12)) in modeling evolving behavioral patterns across temporal, contextual, and federated dimensions. Furthermore, the low Hamming loss (HL < 0.08) confirms the reliability of the Transformer-enhanced trust classifier in detecting fine-grained deviations. From a system efficiency standpoint, MT-SeCom maintains stable communication and resource utilization as scale increases. Power usage decreases with larger deployments, indicating reduced per-node processing overhead enabled by the lightweight trust-calculation pipeline. Resource utilization remains balanced (63–66%), ensuring effective task–resource coordination without overloading physical infrastructure. Therefore, the framework demonstrates strong stability and efficiency even as workload scale increases by more than an order of magnitude. TABLE III P ERFORMANCE METRICS FOR MT-S E C OM OVER INCREASING VN S . VN AP N T Acc M E × 10−4 P W (W) RU (%) HL 100 30 85.04 93.47 5.53 6.12E-03 66.05 0.064 400 110 85.03 94.56 6.89 1.96E-04 64.07 0.037 600 180 85.07 93.33 5.98 2.17E-04 64.15 0.027 800 210 86.04 95.20 4.95 3.31E-04 63.04 0.078 1000 300 87.03 94.44 4.05 3.59E-04 63.87 0.052 1600 455 87.07 94.36 3.89 3.92E-04 63.32 0.045 VN: Virtual node; AP N : active physical node; Acc: Average accuracy; M E: Mean error; P W : Power consumption; RU : Resource utilization; HL: Hamming loss

Fig. 3 illustrates the behavior of the multi-factor trust mechanism under increasing adversarial intensity. Across all attack levels (5%–90%) and deployment scales (100–1600 VNs), MTSeCom consistently maintains higher trust scores compared to baseline models. Temporal trust in Fig. 3(a) decreases gradually as attack intensity rises, reflecting the exponential penalization of behavioral drift in (7) and enabling rapid identification of compromised clients. Contextual trust in Fig. 3(b) remains stable despite fluctuating workloads and adversarial behavior, supported by continuous normalization of communication, resource, and privacy metrics in (Eq. 8). Meanwhile, Fig. 3(c) shows that federated trust mechanisms dampen malicious influence through collaborative anomaly detection and validator consensus. The fused behavior in Fig. 3(d) demonstrates the strength of adaptive multi-factor aggregation. Through dynamic adjustment of weights (γ1 , γ2 , γ3 ) in (12), the model prioritizes the most reliable trust channel as adversarial conditions shift, maintaining separation between legitimate and malicious nodes. Overall, the results confirm that resilient trust in large-scale DT environments is best achieved through adaptive, context-aware fusion rather than reliance on a single trust dimension. Fig. 4 summarizes the operational performance of MTSeCom under increasing attack intensity and system scale. In Fig. 4(a), the Transformer-based trust classifier sustains high accuracy (> 90% at low attack rates and > 80% even at 90% attacks), indicating reliable discrimination between benign and malicious clients despite adversarial noise. Consistently,

0.9

0.8

0.7

0.6

0.7 5

100 VNs

25 50 Attack rate (%) 600 VNs

1000 VNs

100 VNs

1600 VNs

(a) Temporal trust

25 50 Attack rate (%) 600 VNs

1000 VNs

90 1600 VNs

(b) Contextual trust

Average adaptive T

0.9

0.8

0.9

Average T f ederated

Average T context

Average T temporal

0.85

0.8

0.75

0.78 0.76 0.74 0.72 0.7

0.7 5

100 VNs

25 50 Attack rate (%) 600 VNs

1000 VNs

90 1600 VNs

100 VNs

25 50 Attack rate (%) 600 VNs

1000 VNs

90 1600 VNs

(d) Adaptive multi-factor trust

Fig. 3. Multi-factor trust analysis

the low Hamming loss trend in Fig. 4(b) confirms minimal false approvals or rejections. Together, these metrics validate the effectiveness of the hybrid anomaly–classification strategy, where temporal–context learning supports fine-grained behavior modeling and anomaly scoring prevents trust inflation from compromised clients. The benefits of the Resilient Communication Manager (RCM) are reflected in Fig. 4(c)– (d). Trust improvement increases with attack severity as the system selectively isolates low-trust nodes and reroutes communication links through reliable peers. Simultaneously, the communication latency reduction (up to 80% at low attack levels and > 40% even under 90% attacks) confirms that the RL-driven optimization not only preserves defense integrity but also enhances communication efficiency. Overall, the results demonstrate that MT-SeCom maintains secure, adaptive, and scalable trust-based collaboration even under extreme adversarial stress. C. Ablation Analysis Table IV presents the ablation analysis of MT-SeCom across six architectural variants to quantify the contribution of each trust component and mechanism. The full model demonstrates optimal balance with 94.23% accuracy, F1 =0.8231, and a minimal mean error (5.06 × 10−4 ), confirming the complementary value of temporal, contextual, and federated trust integration. TABLE IV MT-S E C OM A BLATION A NALYSIS . MT-SeCom versions MT-SeCom (Full) MT-SeCom+No T temporal MT-SeCom+No T context MT-SeCom+No T f ederated MT-SeCom+No adaptive trust Adaptive Trust+MLP

Acc 94.23 92.65 93.35 92.04 96.33 91.83

P rec. 0.95004 0.95005 0.95039 0.95016 0.94991 0.95012

Recall 0.8310 0.8491 0.8432 0.8562 0.8566 0.8588

F1 0.8231 0.8330 0.8341 0.8391 0.8041 0.8390

HL 0.0644 0.0732 0.0414 0.0262 0.0367 0.0813

ROC − AU C 0.9500 0.9500 0.9501 0.9500 0.9601 0.97003

ME 5.06E-04 4.93E-05 6.76E-04 4.11E-04 3.21E-04 5.34E-03

VN: Virtual node; AP N : active physical node; Acc: Average accuracy; M E: Mean error; P W : Power consumption; RU : Resource utilization; HL: Hamming loss

Removing individual trust factors leads to measurable degradation, excluding T temporal and T federated lowers accuracy to 92.65% and 92.04%, respectively, while omission of T context reduces F1 -score, reflecting its role in situational adaptability. Notably, the “No adaptive trust” variant yields higher accuracy (96.33%) but a weaker F1 -score (0.8041), demonstrating that static weighting overfits and lacks robustness under dynamic conditions. Meanwhile, replacing the Transformer with an MLP significantly degrades performance

(Acc=91.83%), highlighting the necessity of attention-driven sequence modeling. D. Comparison The comparative performance analysis in Fig. 5 demonstrates the superiority of the proposed MT-SeCom framework across multiple evaluation dimensions under varying attack intensities ranging from 5% to 90%. As shown in Fig. 5(a), MTSeCom consistently achieves the highest accuracy, maintaining performance above 85% even under extreme adversarial loads (i.e., > 75% attack rate). In contrast, baseline models such as PCA, Isolation Forest, and OCSVM experience a sharp degradation as attack rates increase, reflecting their limited adaptability to dynamic and coordinated adversarial behaviors. The Transformers-based trust-driven classifier in MT-SeCom enables stronger resilience by jointly modeling temporal drifts, contextual signatures, and federated trust feedback, thereby preventing false trust elevation of malicious clients. Fig. 5(b) shows that MT-SeCom achieves the lowest Hamming loss across all attack intensities. Classical anomaly detectors and shallow learning methods perform adequately under mild adversarial conditions (< 25%) but degrade sharply beyond 50%. MT-SeCom maintains stable performance via adaptive trust–weighted inference, mitigating misclassification propagation. Precision results (Fig. 5(c)) remain near 0.9 even under high adversarial noise, outperforming deep models (LSTM-AE, Seq2Seq) that lack federated trust alignment. F1score and ROC-AUC trends (Figs. 5(d)–5(e)) show balanced detection of malicious and legitimate behavior, with ROCAUC consistently above 0.9, thanks to dynamic trust parameter adjustment (γ1 , γ2 , γ3 ). Fig. 5(f) confirms that MTSeCom minimizes anomaly occurrence under extreme attacks (> 75%), suppressing cascading adversarial effects, whereas reconstruction-based baselines overflag anomalies. VI. C ONCLUSION AND F UTURE W ORK This paper proposed the Multi-Factor Trust-Driven Secure Communication (MT-SeCom) model for secure and resilient collaboration in cloud-based Digital Twin (DT) networks. By combining temporal, contextual, and federated trust with a Transformer-based classifier, MT-SeCom effectively identifies trusted clients and isolates malicious nodes, while the Resilient Communication Manager ensures reliable, low-latency communication. Analytical evaluation demonstrates its adaptability

85 80

6.5 6 5.5

75 70

100 VNs

25 50 Attack rate (%) 600 VNs

1000 VNs

1600 VNs

25 50 Attack rate (%)

100 VNs

(a) Trusted client classification accuracy

600 VNs

1000 VNs

30 Average latency reduction (%)

Average T rust improvement (%)

7.5

Average HL

Average accuracy(%)

·10−2

60 40 20 0

100 VNs

1600 VNs

(b) Hamming loss

25 50 Attack rate (%) 600 VNs

1000 VNs

1600 VNs

100 VNs

25 50 Attack rate (%) 600 VNs

1000 VNs

90 1600 VNs

(d) Latency reduction

Fig. 4. MT-SeCom efficiency analysis

0.9

80 70

Average precision

Average HL

Average accuracy(%)

0.3

90 0.2

0.1

0.7 0.6 0.5

60 50

0.4

25 50 Attack rate (%)

MT-SeCom PCA

MVA Iso. Forest

LSTM-AE OCSVM

Seq2Seq

25 50 Attack rate (%)

MT-SeCom PCA

(a) Accuracy

MVA Iso. Forest

LSTM-AE OCSVM

(b) Hamming loss

Anomaly occurrence (%)

ROC-AUC

0.7 0.6

0.8

0.7 0.5 0.6

0.4 10

MT-SeCom PCA

25 50 Attack rate (%) MVA Iso. Forest

LSTM-AE OCSVM

90 Seq2Seq

(d) F1 Score

MVA Iso. Forest

LSTM-AE OCSVM

90 Seq2Seq

0.6

0.9

0.8

25 50 Attack rate (%)

MT-SeCom PCA

Seq2Seq

0.9

F1 score

0.8

0.4

0.2

MT-SeCom PCA

25 50 Attack rate (%) MVA Iso. Forest

LSTM-AE OCSVM

(e) ROC-AUC

90 Seq2Seq

MT-SeCom PCA

25 50 Attack rate (%) MVA Iso. Forest

LSTM-AE OCSVM

90 Seq2Seq

(f) Anomaly occurrence

Fig. 5. Clients classification analysis: MT-SeCom versus state-of-the-art methods.

to dynamic network conditions and adversarial threats, offering a scalable framework for heterogeneous DT environments. Future work focuses on richer attack models such as bursty, periodic, adaptive, and Markov-driven to capture complex trust dynamics, together with unsupervised trust modeling, real-time RL-based adaptation, cross-domain interoperability, large-scale evaluation, privacy preservation, and quantitative analysis of energy and computational costs for deployment readiness. R EFERENCES [1] D. Saxena and A. K. Singh, “A self-healing and fault-tolerant cloudbased digital twin processing management model,” IEEE Transactions on Industrial Informatics, vol. 21, no. 5, pp. 4233–4242, 2025. [2] E. Woollacott, “Healthcare organizations report rampant email security failures – and microsoft 365 is often the weakest link,” https://www.itpro.com/business/business-strategy/healthcareorganizations-report-rampant-email-security-failures-and-microsoft-365is-often-the-weakest-link, Sep 2025.

[3] Bitsight, “Inside cyber threats on manufacturing in 2025,” 2025. [Online]. Available: https://www.bitsight.com/blog/inside-cyber-threatsin-manufacturing-2025 [4] Y. Lu, X. Huang, K. Zhang, S. Maharjan, and Y. Zhang, “Communication-efficient federated learning for digital twin edge networks in industrial iot,” IEEE Transactions on Industrial Informatics, vol. 17, no. 8, pp. 5709–5718, 2020. [5] Y. Tang, K. Wang, D. Niyato, J. Li, O. A. Dobre, and T. Q. Duong, “Secure data sharing and prediction with digital twin and blockchain in healthcare,” IEEE Communications Magazine, 2025. [6] M. Soula, B. Mbarek, and A. Meddeb, “A real-time trust management model using digital twin in iot networks,” IEEE Access, 2024. [7] Z. Lyu, C. Cheng, H. Lv, and H. Song, “Blockchain based intelligent resource management in distributed digital twins cloud,” IEEE Network, vol. 38, no. 4, pp. 143–150, 2024. [8] S. Son, D. Kwon, J. Lee, S. Yu, N.-S. Jho, and Y. Park, “On the design of a privacy-preserving communication scheme for cloud-based digital twin environments using blockchain,” IEEE Access, vol. 10, pp. 75 365– 75 375, 2022. [9] J. Liu, L. Zhang, C. Li, J. Bai, H. Lv, and Z. Lv, “Blockchain-based secure communication of intelligent transportation digital twins system,” IEEE transactions on intelligent transportation systems, vol. 23, no. 11,

pp. 22 630–22 640, 2022. [10] A. Saad, S. Faddel, T. Youssef, and O. A. Mohammed, “On the implementation of iot-based digital twin for networked microgrids resiliency against cyber attacks,” IEEE transactions on smart grid, vol. 11, no. 6, pp. 5138–5150, 2020. [11] ——, “On the implementation of iot-based digital twin for networked microgrids resiliency against cyber attacks,” IEEE Transactions on Smart Grid, vol. 11, no. 6, pp. 5138–5150, 2020. [12] B. Bera, A. K. Das, and B. Sikdar, “Digital twins-empowered secure network slice access and isolation for consumer healthcare applications,” IEEE Transactions on Services Computing, vol. 17, no. 6, pp. 3429– 3444, 2024. [13] Y. Wang, Z. Su, S. Guo, M. Dai, T. H. Luan, and Y. Liu, “A survey on digital twins: Architecture, enabling technologies, security and privacy, and future prospects,” IEEE Internet of Things Journal, vol. 10, no. 17, pp. 14 965–14 987, 2023. [14] M. Mrabet and M. Sliti, “Towards secure, trustworthy and sustainable edge computing for smart cities: innovative strategies and future prospects,” IEEe Access, 2025. [15] C. Reiss, J. Wilkes, and J. L. Hellerstein, “Google cluster-usage traces: format+ schema,” Google Inc., White Paper, pp. 1–14, 2011.

Record · ID 224486 · SHA-256 7b7f6a5a32391ca8
Retrieved via Conceptio — every document is proof-bundled with source, license, and retrieval metadata.