Fifty Shades of Darknet Siddique Abubakr Muntaka∗ , Jacques Bou Abdo∗ ∗ School of Information Technology
University of Cincinnati Cincinnati, OH 45221, USA {[email protected], [email protected]}
Layer 0: Regular Internet (G0 )
Notice—
V0 : all endpoints
arXiv:2605.19437v1 [cs.NI] 19 May 2026
“The Invisible Internet”
This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. Submitted to: IEEE MILCOM 2026, Washington, DC, USA (under review).
Layer 1: I2P Darknet (G1′ , Shades 1–7) I2P (G1 )
“Invisible Within Invisible”
Layer 2: Exclusive Network (G2 , Shade 8)
Abstract—The Invisible Internet Project (I2P) is a peer-topeer anonymous overlay network whose architecture includes a structurally distinct sublayer not characterized in existing security literature. We term this sublayer the Exclusive Network: nodes here host operational services and draw on I2P’s routing resources, but publish no RouterInfo record to the network’s distributed database (NetDB). In a controlled three-node testbed, we demonstrate that an Exclusive Network node survives sequential floodfill queries from a pool of routers with zero NetDB hits, while its hosted service remains continuously accessible to authorized peers. This property is exploitable by documented I2P-based malware, for example, I2PRAT (RATatouille), for persistent command-andcontrol operations against national assets or corporate networks. The structure is analogous to nation-state Operational Relay Box (ORB) infrastructure. The existence of this sublayer, together with the inability of top-down empirical mapping to characterize it, motivates a move toward formal analytical methods to understand the emergence and behavior of covert networks within I2P. Index Terms—Covert Communication, Exclusive Network, Invisible Internet, Command and Control (C2), Cyber Warfare, Operational Relay Box (ORB), Advanced Persistent Threat (APT).
I. I NTRODUCTION The Invisible Internet Project (I2P), among other anonymity networks, carries significant implications for national security and homeland security [1], [2]. Whereas its protocol justifiably enforces net neutrality and free speech [3], its anonymous design provides the technical capabilities required for conducting covert command-and-control (C2) activities, persistent orchestration channels, and communications that defy attribution efforts to a degree that makes it difficult to implement cyber attribution, cyber deterrence, and cyber warfare strategy [4], [5]. For practitioners and defenders operating in these domains, understanding complex anonymous systems like I2P is not optional. The network’s architecture determines what attribution techniques can and cannot observe, and that boundary directly shapes the feasibility of any response. The predominant methodology for studying I2P has been top-down empirical mapping for two decades [6]–[8]: query the distributed network database (NetDB), collect RouterInfo (RI) records from floodfill nodes, enumerate observable peers, and
V1′ : NetDB-visible routers V2 = V1 \ V1′ : no RI published
Fig. 1. Three-layer network hierarchy. Layer 1 (G1′ ) is the observable I2P darknet (Shades 1–7). Layer 2 (G2 ) is the Exclusive Network: routers that publish no RouterInfo, invisible even within I2P.
analyse LeaseSet distributions to characterise network topology and behaviour. This approach has produced genuine insight into the observable network. Its structural limitation has gone unaddressed. The NetDB is populated entirely by voluntary publication [9], [10]. A node that withholds its RouterInfo contributes no record for any probe to retrieve. Increasing probe frequency, expanding floodfill coverage, or refining query scope all operate on the same directory that was never populated. This is not a calibration problem. It is an architectural property of the I2P protocol, and it defines a space that every empirical mapping technique in the literature is structurally incapable of reaching. In I2P, there is one such layer that plays this precise function. We label this layer the Exclusive Network and represent it graphically as Layer 2 in Fig. 1. Nodes at this layer build tunnels, operate functional eepsites, and route garlicencrypted traffic through the wider I2P infrastructure, hence utilizing bandwidth and tunnel capacity resources provided by others. These nodes are free riders: present in the true network yet wholly absent from any observable directory. The barrier to entry is low, allowing threat actors ranging from criminal groups to nation-state operators to deploy persistent covert infrastructure with no artifact exposed to any forensic investigator. This gap carries practical consequences. For example, documented use cases such as the I2P-based remote access trojan (RAT), including I2PRAT (RATatouille) [11], leverage this behavior to facilitate persistent C2. Operational Relay Box (ORB) networks for nation-state campaigns [12] also achieve this through jurisdictional dispersion of compromised relay infrastructure. Although two decades of I2P measurement efforts have not characterised this layer, Sections III and VI
seek to remedy this issue. The incomplete nature of top-down empirical mapping sets an upper limit on how much can be learned from measurement alone. This work provides empirical evidence of that limitation, as well as an explanation for its underlying cause, thus providing motivation for the development of analytical techniques to complement the existing empirical methods. The contribution of this study is as follows: 1) We demonstrate the Exclusive Network as a structurally distinct sublayer within I2P whose nodes can operate as covert infrastructure and remain undetectable by existing mapping techniques, connecting this property to documented I2P-based malware and nation-state ORB infrastructure. II. N ETWORK A RCHITECTURE AND F ORMAL M ODEL A. Three-Layer Hierarchy
C Client
R1
R2
R3
Rdv
Outbound LS: G-hash only (ℓ-hop) G
H1 Inbound (ℓ-hop)
H2
E
Never published
Shade 8 host
Fig. 2. I2P inbound tunnel. The LeaseSet (LS) publishes only the gateway hash G; the Shade 8 hosting endpoint E is absent from the LeaseSet and from the NetDB. Rdv = rendezvous point; ℓ = tunnel length.
neighbours [14]. Garlic encryption bundles multiple payloads into a single transmission, obscuring message boundaries from any intermediate relay [9]. The NetDB is a Kademlia-derived distributed hash table (DHT) [15] whose storage responsibility rotates daily through an XOR routing key expressed as: rk(d) = SHA-256(Hd ⊕ SHA-256("yyyyMMdd"))
The I2P (garlic network) is not simply a layer added to the existing Internet to provide anonymity [9]. Rather, the design of the I2P network creates a hierarchical nesting structure across three levels. We model this hierarchy using nested graphs, and the invisibility property arises directly as a result of the model architecture.
where Hd = SHA-256(dbytes ) and dbytes is the destination’s public-key. The floodfill whose 256-bit hash sits XOR-nearest to rk(d) becomes the designated storage node:
I2P routes traffic through independent unidirectional tunnel chains of depth ℓ, where each relay knows only its immediate
where fcap maps observable capability properties to shades 1 through 7 per Table I.
(1)
f ∗ (d) = arg min [Hf ⊕ rk(d)]Z
(2) f ∈F Definition 1: I2P Network Hierarchy. Let G0 = (V0 , E0 ) be the Internet, where V0 is the set of reachable endpoints. where [·]Z denotes unsigned integer interpretation. Equation (2) The I2P overlay will form a proper subgraph represented as underpins Method C in Section V. G1 = (V1 , E1 ), where V1 ⊂ V0 represents all active I2P router An eepsite’s inbound LeaseSet publishes only the gateway endpoints (Layer 1). Any RouterInfo-based mapping of this hash and a tunnel identifier (Fig. 2), disclosing nothing about overlay recovers only the observable subgraph G1′ = (V1′ , E1′ ), the identity of the hosting router at the tunnel endpoint. An where V1′ ⊆ V1 comprises those routers that has published operator in the Exclusive Network withholds the RouterInfo signed RouterInfo (RI) records to the global NetDB managed entirely, removing that endpoint from V1′ and placing it in V2 . by floodfill routers. The Exclusive Network is the residual set III. S HADE TAXONOMY G2 = (V2 , E2 ), where V2 = V1 \ V1′ (Layer 2), comprising routers structurally absent from the NetDB. The Shade Taxonomy formalises the visibility gradient across The non-emptiness of V2 is an immediate result of protocol I2P routers, making the boundary between the observable design, not a measurement artefact. I2P imposes no require- network and the Exclusive Network precise. Prior literature ment on routers to publish their RouterInfo (RI); directory treats I2P as a binary distinction between floodfill and nonparticipation is fully voluntary [9]. Thus, a router may join V1 , floodfill routers for mapping purposes [6]. That framing is build outbound tunnels as a client, and contribute zero vertices insufficient. The taxonomy captures eight discrete visibility classes derived entirely from observable RouterInfo fields, to G1′ . demonstrating that structural invisibility is the endpoint of Proposition 1: Structural Incompleteness. For any RI-based a spectrum rather than a binary property. Shades 1 through 7 mapping approach M : G1 → G1′ , we have G1′ ⊊ G1 when at reside in Layer 1 (V ′ ); Shade 8 defines Layer 2 (V ). 2 1 least one router operates in exclusive mode. Definition 2: Shade Classifier. Let κ(r) denote the capabilities This follows from voluntary RI publication. From a network string of router r, α(r) ∈ {0, 1} indicate whether a direct science perspective, analogous structures appear as the “dark transport address is published, ι(r) ∈ {0, 1} indicate introducer matter” of complex networks [13]: entities that shape system presence, and δ(r) ∈ {0, 1} indicate whether a RouterInfo behaviour while absent from every observable adjacency record for r is present in the global NetDB. The shade class structure. The observable completeness ratio ρ = |V1′ |/|V1 | of r is: and its complement ξ = 1 − ρ quantify this gap, and are ( 8 if δ(r) = 0 evaluated empirically in Section VI. σ(r) = (3) fcap κ(r), α(r), ι(r) otherwise B. Garlic Routing and the NetDB
TABLE I S HADE TAXONOMY. κf : FLOODFILL FLAG ; κH : HIDDEN FLAG ; κU : FIREWALLED FLAG . T HE L AYER COLUMN REFERS TO F IG . 1. Shade Name 1 2 3 4 5 6 7 8
Beacon Relay Passive Cloaked Veiled Declared Phantom Exclusive
Criteria
C2 Role
κf , α=1 High-cap, α=1 Low-cap, α=1 κU , α=1 α=0, ι=1 κH , α=0 α=0, ι=0, δ=1 δ=0
NetDB anchor Traffic relay BW donor Hidden relay Covert relay Semi-hidden Ghost node Stealth C2
Layer 1 1 1 1 1 1 1 2
local I2P proxy using the SAM bridge (port 7656) and engages in communication with the C2 eepsite having its b32 address hardcoded into the malicious executable code. In cases where the server-side node is set up as the Exclusive Network (Shade 8), the outgoing network traffic from the victimized machine will look like regular I2P traffic [11] [17]. Attribution then requires either binary forensics on the implant or comprehensive traffic correlation across the entire I2P overlay. C. Operational Relay Box Parallel
Operational Relay Box (ORB) networks, attributed to multiple Chinese APT groups, construct multi-hop proxy chains Shades 1 through 7 form a concealment progression from through compromised small office home office (SOHO) routers Shade 1 (Beacon), which carries the floodfill flag and anchors and cloud infrastructure to obscure operation origins [12]. the NetDB, through firewalled, introducer-only, and hidden- Where ORB achieves unattributability through jurisdictional flag variants, to Shade 7 (Phantom), which has no address and dispersion of relay infrastructure, the Exclusive Network no introducer yet still exists in the NetDB. All seven satisfy achieves it through protocol-level directory non-publication. Both instantiate Gdark ⊂ G: an operational subgraph that δ(r) = 1: present in the directory, however hard to contact. Shade 8 is categorically different. It satisfies δ(r) = 0: contributes to network behaviour while remaining absent from no NetDB record exists, and none can be retrieved by any every observable directory. This shared structure positions the RouterInfo-based method regardless of how many floodfills model developed in Section II as directly applicable to the are probed. A Shade 7 router is hard to reach; a Shade 8 structural analysis of ORB-class architectures. router is structurally absent from the NetDB. The classifier in V. E MPIRICAL M ETHODOLOGY Eq. (3) encodes this boundary by checking δ(r) before any capability inspection, since capability fields are unavailable A. Three-Node Testbed when no RouterInfo record exists. The empirical validation of The testbed comprised three nodes running this property is the subject of Section VI. Ubuntu 24.04 LTS with I2P 2.12.0 (API 0.9.69). All configuration scripts are available on GitHub [16]. IV. T HREAT M ODEL A. Operational C2 Architecture A threat actor deploying a Shade 8 node as C2 infrastructure configures the router’s router.config file with parameters that suppress all directory participation. Our exclusive network script implements two progressively deeper profiles [16]: router.isHidden=true router.hiddenMode=true i2np.udp.addressSources= # empty i2np.ntcp2.autoip=false router.floodfillParticipant=false router.maxParticipatingTunnels=0 router.sharePercentage=0 router.enablePeerTest=false router.dynamicKeys=true # ephemeral identity i2np.udp.requireIntroductions=true
Listing 1. Exclusive profile (10 parameters). Ghost profile adds 8+ more, including firewalled declaration and laptop-mode identity rotation.
With these settings, the router publishes no RI to any floodfill, refuses to relay traffic for other nodes, disables peertesting probes that would reveal reachability, and rotates its cryptographic identity on every restart without affecting eepsite keys, which are stored separately in eepPriv.dat. The hosted eepsite remains reachable to partners who hold the b32 address. B. Connection to Documented Malware Remote access trojan variants that exploit the I2P architecture, for example I2PRAT (RATatouille) [11] implement this paradigm explicitly. The backdoor initiates a connection to the
VM1 (Exclusive Host). Router H1 = PB5dY5gvdEpj..., configured via exclusiveStealth-network.sh (ghost profile, Listing 1). Hosts eepsite sid001.i2p with custom TLD routing managed by customtld-manager.sh. VM2 (Authorised Partner). Router H2 = 6FRyiaaN..., configured via setup-i2p-proxy.sh with SOCKS5 access to VM1’s eepsite. VM2 holds VM1’s b32 address obtained through an out-of-band channel. VM3 (Adversary Scanner). Runs node-lookup.py with knowledge of H1 ’s router hash but no knowledge of its b32 eepsite address, and separately runs b32-lookup.py with knowledge of the b32 address but no knowledge of the hosting router hash. These two complementary probes simulate an adversary approaching from either direction, together constituting the empirical test of Shade 8 structural invisibility. B. NetDB Measurement Approach RouterInfo records are collected from the local NetDB directory (~/.i2p/netDb/) by parsing binary routerInfo-*.dat files via the POSIX strings utility, extracting capabilities, transport addresses, version strings, and known-peer counts. The I2P console API at 127.0.0.1:7657/netdb provides a second source. Our snapshot consisted of 3,242 RI entries, with 1,556 (48.0%) having the floodfill capability flag set, consistent with established I2P measurement practice [6], [18].
Algorithm 1 Shade Classification Protocol Input: Router hash h, local NetDB D, floodfill set F, batch size b Output: Shade class σ(h) ∈ {1, . . . , 8} 1: if h ∈ D then 2: Extract (κ, α, ι) from local RI 3: return fcap (κ, α, ι) ▷ Shades 1–7 4: end if 5: Query console: RI ← /netdb?r=h 6: if RI ̸= ∅ then 7: return fcap (κ, α, ι) from RI 8: end if 9: for i ← 0 to |F| step b do 10: for f ∈ F [i : i+b] do 11: Probe: /netdb?r=f ▷ Expand console view 12: end for 13: RI ← /netdb?r=h 14: if RI ̸= ∅ then 15: return fcap (κ, α, ι) from RI 16: end if 17: end for 18: return 8 ▷ Shade 8: δ(h) = 0 confirmed
Algorithm 2 XOR Routing Key Association (Method C) Input: Target hash Ht , eepsite set S, floodfill map F (hash → bytes) Output: R ⊆ S: eepsites for which Ht is responsible 1: mk ← SHA256(UTC_date_as_yyyyMMdd) 2: R ← ∅ 3: for s ∈ S with known b32 do 4: Decode Hs from b32 prefix 5: rks ← SHA256(Hs ⊕ mk ) 6: dt ← [Ht ⊕ rks ]Z 7: closest ← true 8: for f ∈ F, f ̸= Ht do 9: if [Hf ⊕ rks ]Z < dt then 10: closest ← false; break 11: end if 12: end for 13: if closest then 14: R ← R ∪ {s} 15: end if 16: end for 17: return R
C. Attribution Methods The attribution framework consists of five approaches: three incorporated into Algorithm 1 (local NetDB inspection, console cache query, and floodfill probe expansion) and two described below. Method A is a control verification from the hosting VM; Methods B+D and C are the adversarial approaches applied by VM3. Method A (b32 Derivation). On the hosting VM, the canonical b32 address follows directly from eepPriv.dat: b32(r) = Base32 SHA-256(dr [0 : ds ]) (4) ds = 387 + L, where L = uint16_be(dr [385 : 387]) is the key-certificate length. Confirmed on I2P 2.12.0: cert type 5, ds = 391 bytes, b32 = poitcahygw7f2zz7.... Method B+D (Gateway Scan). All known active LeaseSets are inspected for Lease entries whose gateway hash prefix matches the target. With ℓ-hop tunnels (ℓ ≥ 1), a gateway match indicates routing participation, not hosting. Method C (XOR Routing Key). Using Eqs. (1) and (2), for each known b32 address, the responsible floodfill is identified. If the target matches, it stores that LeaseSet per protocol. Algorithm 1 formalises the Shade classification; Algorithm 2 formalises Method C. VI. E XPERIMENTAL R ESULTS A. Shade 8 Empirical Proof Algorithm 1 ran on VM3 targeting H1 PB5dY5gvdEpj..., with b = 5 and |F| = 1,556: 1) Local NetDB (3,242 RI files): no hit. 2) Console cache: no hit.
=
Fig. 3. Shade 8 classification output from node-lookup.py. After 500 floodfill probes from a pool of 1,556, router H1 (PB5dY5...) produces zero NetDB hits, confirming Layer 2 exclusive status.
3) Floodfill probe (500 floodfills, batches of 5, re-check after each): no hit at any checkpoint. The Shade 8 criterion is satisfied: ¬ RIlocal (H1 ) ∧ ¬ RIconsole (H1 ) ∧
^
¬ RIf (H1 ) (5)
f ∈F500
Fig. 3 confirms zero NetDB hits across all three sources. Simultaneously, VM2 maintained continuous access to sid001.i2p, confirming that structural invisibility and operational functionality coexist in Layer 2. Fig. 4 plots NetDB hit count against cumulative probes for three router classes, confirming H1 remains in V2 = V1 \ V1′ across all 500 probes.
NetDB hits
3
H1 (Shade 8) Shade 1 (Beacon) Shade 7 (Phantom)
2 1 0 0
0
10
0 20
0
30
0
40
0
50
0
60
0
70
Cumulative floodfill probes
Fig. 4. NetDB hit count vs. cumulative floodfill probes for three router types. H1 (Shade 8, Layer 2) remains absent across all 500 probes; observable Layer 1 nodes are found within tens to hundreds of probes.
Fig. 6. Method C (XOR proximity) output for router F1 (SnuNBZ65...). XOR analysis across 1,536 floodfills and 172 active LeaseSets identifies 2yxn3ei... as the nearest eepsite storage node, confirming full NetDB integration under Algorithm 2.
This mirrors the structural challenge of ORB attribution [12]: in both cases, the observable network is a proper subgraph of the true operational topology. I2P achieves invisibility through protocol-level non-publication; ORB achieves it through jurisdictional dispersion of compromised relay infrastructure. Fig. 5. Shade 1 (Beacon) classification output for router F1 (SnuNBZ65...). Known Routers: 7,778; Known LeaseSets: 213; caps XfR. All five detection methods return positive results, confirming full observability in Layer 1.
B. Implications for Cyber Deterrence
Deterrence theory requires that a credible retaliatory or sanctioning response be made possible by attribution [5]. Layer 2 I2P infrastructure severs this chain at its foundation, leaving a C2 operator using an exclusive node with no credible B. Method C: Floodfill Association technical-attribution-based deterrence. Deterrence strategies Algorithm 2 targeted floodfill SnuNBZ65faZL... must therefore shift from infrastructure-based attribution to be(caps XfR, I2P 0.9.68, 107.172.250.117:16657, havioural attribution: the analysis of targeting patterns, exploit Known LeaseSets: 213, Fig. 5) against 1,536 floodfills tooling, and operational rhythms, which is less dependent on and 172 candidate b32 addresses. It identified the visibility of individual network nodes. 2yxn3eio....b32.i2p as the one address in the candidate set for which this floodfill is the XOR-nearest known storage C. Network Science: Dark Vertices node (Fig. 6). Method B+D separately identified a gateway The Shade Taxonomy contributes a vertex-visibility classifiparticipation association for lcjlqzkb....b32.i2p. cation to the network science of anonymous systems. Shade 1 For comparison, Fig. 7 shows relay node gz9qliN5Zx7... routers exhibit high betweenness centrality in G ′ ; they are 1 (caps XR, Shade 2). Zero known routers and zero LeaseSets the primary carriers of routing state. Shade 8 routers have confirm no NetDB storage participation; Methods B and D zero degree in G ′ despite positive degree in the true graph G , 1 1 scan 69 active LeaseSets and return zero gateway matches. analogous to “dark nodes” in social network analysis where Despite full tunnel routing participation, this node carries no influential actors are absent from observable contact graphs [13]. service descriptors, placing it firmly in Layer 1 Shade 2 as a The complement ξ = 1 − ρ bounds the fraction of the true structural contrast to H1 . topology that remains structurally inaccessible regardless of VII. D ISCUSSION
measurement methodology.
A. Attribution and the Observable Graph Limit
D. Dual-Use Dimension
The proof of Eq. (5) demonstrates a hard epistemic boundary for NetDB-based attribution. Every technical methodology applied to G1′ , including IP geolocation, RouterInfo fingerprinting, and floodfill enumeration, is bounded within V1′ . A Shade 8 operator is necessarily in V2 = V1 \V1′ , beyond this boundary by construction. Identifying actors in V2 requires either endpoint forensics or global traffic correlation across all relay hops in G1 , neither tractable at operational tempo for most defenders.
The civil applications of I2P, which safeguard journalists, dissidents, and whistle-blowers in repressive or censored environments, are firmly established and independent of the threat surface analysed here. The dual-use nature of the network cannot be addressed through technical restriction of Layer 2 features without simultaneously undermining those protections. Policy should emphasise behavioural attribution capability rather than protocol interdiction.
Fig. 7. Shade 2 (Relay) classification for router R1 (gz9qliN5..., caps XR). Left: Known Routers: 0, Known LeaseSets: 0, direct IP published, Shade 2: Relay. Right: Methods B and D scan 69 active LeaseSets and return zero gateway associations. Despite full tunnel routing participation, R1 carries no service descriptors and occupies Layer 1 Shade 2.
VIII. C ONCLUSION In I2P, a sublayer exists where nodes consume routing resources without any directory record. Minimal configuration renders a standard I2P router invisible to every observable directory while remaining fully operational as a covert host. Two decades of empirical research have undercharacterized this layer: the directory is never populated, and no increase in probe coverage can retrieve what is never stored. The empirical result is unambiguous: 500 sequential floodfill probes from a pool of 1,556, applied through five attribution methods, returned zero NetDB hits for H1 while its hosted eepsite remained continuously accessible. The bound ξ = 1 − ρ is a hard protocol-design limit. Actors who understand this, such as in I2PRAT [11], use techniques to operate where directory-based attribution is structurally impossible. The same property Gdark ⊂ G defines ORB infrastructure [12]; both lie beyond current empirical methods. The primary implication is methodological: understanding the Exclusive Network demands formal analytical approaches independent of directory observation, and this study establishes precisely why empirical mapping alone is insufficient. Future research will expand the Shade Taxonomy to i2pd and I2P+ variants [14], investigate timing-analysis techniques for Shade 8 de-anonymisation, and deploy I2PRAT (RATatouille) [11] as an active C2 implant against an Exclusive Network node to validate attribution-resistance under adversarial conditions. The G′ ⊂ G incompleteness framework further offers a structural foundation for graph-theoretic modelling of ORB architectures, where nation-state actors achieve comparable unattributability through jurisdictional dispersion rather than protocol-level nonpublication. ACKNOWLEDGMENT The authors acknowledge the support of Multi-domain and Information Operations, Resilience and Anonymity Groupe (MIRAGe-UC) research group at the University of Cincinnati.
R EFERENCES [1] S. A. Muntaka and J. B. Abdo, “Optimizing anonymity and efficiency: A critical review of path selection strategies in tor,” in 2025 IEEE/ACS 22nd International Conference on Computer Systems and Applications (AICCSA). IEEE, 2025, pp. 1–8. [2] T. Rid and B. Buchanan, “Attributing cyber attacks,” Journal of Strategic Studies, vol. 38, no. 1–2, pp. 4–37, 2015. [3] B. Zantout and R. A. Haraty, “I2p data communication system,” in Proceedings of the 10th International Conference on Networks (ICN), Gosier, Guadeloupe, 2011, pp. 401–409. [4] H. Chen, A. Coco, A. Rotondo, and Y. Ying, “The attribution of cyber operations to states in international law,” Geneva Centre for Security Policy (GCSP), Tech. Rep., 2025. [5] H. S. Lin, “Cyber conflict and international humanitarian law,” International Review of the Red Cross, vol. 94, no. 886, pp. 515–531, Jun. 2012. [6] C. Egger, J. Schlumberger, C. Kruegel, and G. Vigna, “Practical attacks against the i2p network,” in Proceedings of the 16th International Symposium on Research in Attacks, Intrusions and Defenses (RAID), Rodney Bay, Saint Lucia, 2013, pp. 432–451. [7] A. Biryukov and I. Pustogarov, “Trawling for Tor hidden services: Detection, measurement, deanonymization,” in Proc. IEEE Symp. Security & Privacy, San Jose, CA, 2013, pp. 80–94. [8] N. P. Hoang, P. Kintis, M. Antonakakis, and M. Polychronakis, “An empirical study of the i2p anonymity network and its censorship resistance,” in Proceedings of the internet measurement conference 2018, 2018, pp. 379–392. [9] I2P Project, “I2P network specifications,” [Online], 2024, available: https://i2p.net/en/docs/specs/. [10] S. A. Muntaka and J. Bou Abdo, “Resilience of the invisible internet project: A computational analysis,” Internet Technology Letters, vol. 8, no. 5, p. e70119, 2025. [11] P. L. Bourhis, “Ratatouille: Cooking up chaos in the i2p kitchen,” Feb. 2025. [Online]. Available: https://blog.sekoia.io/ ratatouille-cooking-up-chaos-in-the-i2p-kitchen/ [12] M. Raggi, “IOC extinction? China-nexus cyber espionage actors use ORB networks to raise cost on defenders,” Google Cloud Mandiant, Tech. Rep., May 2024, available: https://cloud.google.com/blog/topics/ threat-intelligence/china-nexus-espionage-orb-networks. [13] A.-L. Barabási and R. Albert, “Emergence of scaling in random networks,” Science, vol. 286, no. 5439, pp. 509–512, Oct. 1999. [14] S. A. Muntaka and J. B. Abdo, “Systemic flaws in the invisible internet project: Analysis of exploitable design choices,” 2026. [15] P. Maymounkov and D. Mazieres, “Kademlia: A peer-to-peer information system based on the xor metric,” in Proceedings of the 1st International Workshop on Peer-to-Peer Systems (IPTPS), ser. Lecture Notes in Computer Science, vol. 2429. Springer, 2002, pp. 53–65.
[16] S. A. Muntaka, “Fiftyshadesdarknet: Research datasets, scripts, and analysis tools for darknet and anonymous network investigations,” 2026. [Online]. Available: https://github.com/abksiddique/FiftyShadesDarknet [17] Kaspersky GReAT, “MATA: Multi-platform targeted malware framework,” Kaspersky Securelist, Tech. Rep.,
Jul. 2020. [Online]. Available: https://securelist.com/ mata-multi-platform-targeted-malware-framework/97746/ [18] S. A. Muntaka, J. B. Abdo, K. Akanbi, S. Oluwadare, F. Hussein, O. Konyo, and M. Asante, “Mapping the invisible internet: Framework and dataset,” Data in Brief, p. 112175, 2025.