Chameleon: An Adaptive AI-Driven Honeypot Architecture Using Threat-Calibrated Particle Swarm Optimization and Semantic Deception Rapidly-Exploring Random Trees

Chameleon: An Adaptive AI-Driven Honeypot Architecture Using Threat-Calibrated Particle Swarm Optimization and Semantic Deception Rapidly-Exploring Random Trees Rohit Swami, Tushar Singh, Akash Warde, and Sri Muthu

arXiv:2608.15407v1 [cs.CR] 15 Aug 2026

Dept. of AI and Data Science, SIES Graduate School of Technology, Mumbai, India rohitrsaids124@gst.sies.edu.in, tusharnsaids124@gst.sies.edu.in, akashkwaids124@gst.sies.edu.in, srimuthulakshmananraids124@gst.sies.edu.in Guide: Dr. Neethu Anna Sabu (neethus@sies.edu.in)

Abstract—An invariant behavioral profile is the defining vulnerability of traditional honeypot installations: a skilled adversary can confirm the presence of a deception environment within only a few diagnostic commands, limiting its intelligence value. High-cost commercial deception products (USD 100,000– 150,000 per year) share a related weakness in that their response engines are not coupled to real-time model-driven feedback. Chameleon is an openly distributed adaptive honeypot platform introduced here to address both shortcomings. Three core components are integrated: a bidirectional long short-term memory (BiLSTM) classifier achieving 99.61% accuracy across seven threat categories at approximately two milliseconds CPU latency; a locally deployed Qwen3.5-0.8B language model (Qwen Team, 2026; Unsloth, 2026) delivering 90% contextual generation accuracy at 4.5 milliseconds average latency; and two domainspecific meta-heuristic engines. Threat-Calibrated Particle Swarm Optimization (TC-PSO) dynamically reshapes swarm inertia and objective amplification in proportion to the classifier’s anomaly output, enabling real-time adjustment of connection-holding delays. Semantic Deception Rapidly-Exploring Random Trees (S-RRT) drives deception schema evolution via exponentially scaled pheromone updates derived from a language-model severity assessment, while a depth-decay multiplier enforces a finite memory footprint. Across five benchmark runs (seeds 42–46), TC-PSO outperformed standard PSO by 48.1% in mean fitness (2.60 to 3.85) with a 32.7% convergence gain, and S-RRT exceeded standard RRT by 258.9% in best-run fitness (450.2 to 1,615.8), achieving a 329.2% gain at critical severity and a 24.9% memory reduction (p < 0.01). Operating costs are approximately USD 17 per month, a roughly 490-fold reduction versus commercial alternatives. Index Terms—Adaptive honeypot; Threat-calibrated particle swarm optimization; Semantic deception RRT; BiLSTM classification; Meta-heuristic optimization; Deception technology

1. I NTRODUCTION Alert fatigue has become a structural liability within modern security operations centers. Industry reporting indicates that individual organizations process in excess of eleven thousand security events daily, yet triage capacity is so constrained that roughly seventy percent of those events receive no analyst attention (Ponemon Institute, 2023; Apruzzese et al., 2023). Perimeter-oriented defenses—next-generation firewalls,

signature-driven intrusion detection, and endpoint platforms— are architecturally ill-suited to this environment because each depends on previously catalogued threat patterns and can offer only reactive, post-event visibility against living-off-the-land movement or zero-day exploitation. Deception infrastructure has long been proposed as a means of redistributing that asymmetry. Attacker engagement environments that appear operationally genuine allow defenders to observe adversarial tactics, techniques, and procedures in forensic detail while the adversary remains unaware that the session is controlled. The problem is that the dominant openly available sensors—Cowrie, a configurable SSH/Telnet emulation platform (Cowrie Project, 2023), and T-Pot, a consolidated multi-sensor aggregator (T-Pot Community, 2023)— share a common structural deficiency: their behavioral output is determined at deployment time and cannot be revised in response to the evolving session. As few as one or two probe commands, such as a uname -a query, may be sufficient for an informed attacker to match the response pattern against a known honeypot fingerprint and withdraw. Commercially available alternatives such as Illusive Networks and Thinkst Canary address this deficiency only partially, and none connects response generation to a live machine-learning pipeline. Market valuations underscore the scale of the unmet need: the deception technology sector was estimated at approximately 2.7 billion US dollars in 2024 and is projected to reach between 5.2 and 6.0 billion US dollars by 2029 (MarketsandMarkets, 2024), with open-source capability lagging well behind the commercial frontier. Chameleon is introduced here as a production-complete, MITlicensed adaptive honeypot designed to close this capability gap. Five primary contributions are advanced. First: TC-PSO, which, to the authors’ knowledge, represents the inaugural application of a BiLSTM-derived anomaly signal to govern particle inertia and objective amplification within a honeypot tarpit-delay optimization context. Second: S-RRT, the first variant of the rapidly-exploring random tree family in which pheromone reinforcement is scaled exponentially through a

severity index computed by a language model, combined with a provable depth-decay memory constraint. Third: a feedbackcoupled two-stage cascade pairing the BiLSTM threat classifier with a compact Qwen3.5-0.8B model (Qwen Team, 2026; Unsloth, 2026), achieving approximately 99.5% combined accuracy at a mean latency near 4.5 milliseconds on CPUonly hardware. Fourth: a production-complete instrumentation stack encompassing Merkle-chained forensic logs anchored to Ethereum Sepolia, honeytoken canary files triggering STIX 2.1 bundle generation upon access, and a Paramiko-based SSH sensor with full pseudo-terminal emulation. Fifth: reproducibility through MIT-licensed source code, benchmark datasets, and a deterministic test suite that allows full independent replication.

2.6%. That formulation, however, treats all traffic identically regardless of assessed severity, and no prior work couples swarm inertia to a live anomaly-scoring model in a honeypot engagement setting. In trajectory planning, the rapidlyexploring random tree was introduced by LaValle (2006), and the pheromone-reinforcement principle underlying the exponential update of S-RRT is based on the ant-colony model of Dorigo, Maniezzo, and Colorni (1996). Flat-pheromone RRT applied to deception-network topology mutation can achieve effective graph coverage but is, in the general case, prone to unconstrained node proliferation under sustained high-severity scenarios (LaValle, 2006)—a pathology that S-RRT’s depthdecay mechanism is expressly designed to eliminate.

2. BACKGROUND AND R ELATED W ORK

D. Hybrid Classification Pipelines and Commercial Platforms

A. Static Honeypot Platforms The virtual honeypot paradigm was first formalised by Provos (2004), who showed that software-emulated services were able to attract and contain adversarial traffic with less operational risk than physical decoy systems. Spitzner (2003) later proposed the canonical interaction-level taxonomy in which behavioural fingerprintability—the vulnerability of a sensor to rapid identification based on deterministic response patterns—was identified as the primary impediment to sustained deception. Cowrie (Cowrie Project, 2023) records the Debian shell environment with a configurable but immutable file system state, a configuration that has been well documented and catalogued by attacker communities. T-Pot (T-Pot Community, 2023) brings many sensor engines behind a common logging and visualisation stack, but does not include any mechanism for dynamic response adaptation. None of these architectures, or their derivative implementations, includes any form of real-time, classifier-driven behavioural adjustment. B. Language-Model-Enhanced Terminal Emulation

Dai et al. (2024) report that a CNN-BiLSTM hybrid classifier augmented with an attention mechanism achieved up to 99.84% accuracy on the CIC-DDoS2019 dataset, but functions as a self-contained detector with no connection to deception or optimisation subsystems. Transformer-based anomaly detectors have shown competitive accuracy on flow-based network traffic benchmarks (Manocchio et al., 2024), again without feedback integration. On the commercial side, proprietary ensemble deception platforms typically operate as black-box, cloudhosted systems and charge between one hundred thousand and one hundred fifty thousand US dollars per year, reflecting current market positioning for enterprise deception technology (MarketsandMarkets, 2024). A context-sensitive deception reverse-proxy approach was described by Fraunholz et al. (2018), but without any connection-holding delay optimization capability. Chameleon is distinguished from each of these lineages by its end-to-end feedback-coupled architecture— spanning threat classification, meta-heuristic optimization, and language-model-driven response generation—deployed on opensource, CPU-only infrastructure.

Otal and Canbaz (2024) evaluated an SSH honeypot in3. M ATERIAL AND M ETHODS teraction engine based on a fine-tuned Llama-3-8B model, reporting a cosine similarity of 0.695 between generated and A. System Overview Chameleon is realized across approximately 8,400 lines expected Cowrie terminal outputs. Chameleon differs from this approach by employing a much smaller Qwen3.5-0.8B of Python and 2,200 lines of JavaScript and TypeScript, model (Qwen Team, 2026; Unsloth, 2026) fine-tuned with low- with the meta-heuristic optimization module alone accounting rank adaptation, which achieves 90% contextual generation for 1,882 lines. Incoming payloads first traverse a fouraccuracy with an average latency of 4.5 milliseconds. Multi- stage normalization pipeline that applies Unicode NFKC session coherence is maintained via a rolling per-IP interaction homoglyph collapsing, structural character mapping, controlhistory of up to twenty turns, and credential-exfiltration pattern character suppression, and token-level normalization; this prescanning improves detection sensitivity beyond what the processing layer defends downstream models against encodingmanipulation evasion strategies. Normalized payloads are then classifier alone provides. submitted to a two-stage classification subsystem: a characterC. Swarm and Trajectory-Planning Optimizers in Security level convolutional-recurrent pre-filter followed by the primary Particle swarm optimisation, formalised by Kennedy and BiLSTM scorer, collectively distinguishing seven behavioral Eberhart (1995) and extended to the canonical w = 0.729 categories—SQL injection, cross-site scripting, server-side inertia setting by Shi and Eberhart (1998), has been used for includes, path traversal, remote code execution, brute force, intrusion-detection optimization, with a binary-classification and benign activity. Payloads with maximum class probability accuracy of 99.85% reported by Benmalek and Seddiki (2025) of θ = 0.85 or higher are escalated to the deception engine, for a PSO-optimized CatBoost classifier on the RT IoT2022 while payloads below threshold are assigned lightweight static dataset, outperforming baseline methods such as QAE-f16 by responses, keeping the latency below two milliseconds for

the approximately ninety-five percent of sessions classified as with w1 = 0.65, w2 = 2.5, w3 = 0.25, Ibonus = 0.25 × benign. (commands − 5), and β = 0.3. The multiplicative term (1 Sessions are escalated to a locally hosted Qwen3.5-0.8B + β · A(t)) amplifies objective value by up to 30% at A(t) model, deployed from the official Unsloth Qwen3.5-0.8B = 1.0, concentrating optimization pressure on configurations GGUF checkpoint (Qwen Team, 2026; Unsloth, 2026), fine- that prolong engagement with genuine adversaries rather than tuned via low-rank adaptation at rank 8 to 16 across six hundred expending search effort on benign traffic. training iterations. The model is tested against a simulated C. Semantic Deception Rapidly-Exploring Random Trees Ubuntu 22.04 host seeded with honeytoken canary files, e.g., Baseline RRT used for deception filesystem evolution uniaws production keys.csv and .env.backup. A rolling history formly escalates pheromone (∆τ = 0.5) regardless of payload of up to twenty command-response turns per source-IP is severity, and is analytically unbounded in node growth O(bd ) maintained to preserve session continuity. Three reliability mechanisms ensure production availability: (1) a circuit breaker, for sustained high-severity operation (LaValle, 2006). S-RRT which enforces a thirty-second cooldown period after three has two corrective modifications. First, Payload Severity Index consecutive failures within thirty seconds; (2) a time-to-live Ψ ∈ [1.0, 3.0] assesses the payload’s severity by syntactically response cache, which stores one thousand entries with a and semantically analysing the intercepted command sequence twenty-four-hour expiry, reducing effective model calls by through the Qwen model (Qwen Team, 2026; Unsloth, 2026). thirty to sixty percent; and (3) an OGCPS credential-pattern Pheromone reinforcement is then updated as: scanner for exfiltration detection. ∆τ ′ = ∆τ · exp(Ψ − 1) (4) After classification, two optimisation processes run in parallel. TC-PSO maintains a separate particle swarm for each At Ψ = 1.0 the exponential term evaluates to unity, thus attack category to tune the connection-holding delay. S-RRT exactly recovering the standard-RRT update and ensuring maintains a population of deception filesystem trees. Both backward compatibility. At Ψ = 3.0, pheromone reinforcement use the BiLSTM anomaly score A(t) as the key adaptation is amplified 7.39-fold, guiding tree evolution towards deception signal, forming the closed optimisation feedback loop that sets schemas optimised for multi-stage, high-severity engagement Chameleon apart from the previous systems studied in Section profiles (Dorigo, Maniezzo, & Colorni, 1996). Memory growth 2. is constrained by a depth-dependent expansion decay: B. Threat-Calibrated Particle Swarm Optimization In its conventional formulation, PSO propagates particle velocity according to vi (t + 1) = w · vi (t) + c1 r1 (pbest − xi ) + c2 r2 (gbest − xi ) (1) with w = 0.729 as a fixed constant and c1 = c2 = 1.49445 as cognitive and social scaling coefficients (r1 , r2 ∼ Uniform[0, 1]). Applying uniform inertia across all attack categories is suboptimal for connection-delay tuning: benign traffic demands broad exploratory behavior, whereas sessions associated with remote code execution or SQL injection require rapid convergence to maximum engagement delays. TC-PSO resolves this tension by replacing the static inertia term with a classifier-coupled dynamic scalar: w(t) = wbase · max(σmin , 1 − α · A(t))

(2)

where wbase = 0.729, σmin = 0.3, α = 0.5, and A(t) ∈ [0, 1] denotes the real-time BiLSTM anomaly output. The max(·) operator enforces a lower bound on inertia regardless of anomaly magnitude, preserving a minimum degree of exploratory capacity even under maximum-severity conditions (A(t) = 1.0), where w(t) reaches its floor of 0.365. At benign baselines (A(t) = 0), inertia equals the conventional value of 0.729. A complementary modification amplifies the reward signal in proportion to session severity: F ′ (t) = (w1 Cexec )−(w2 Pdrop )+w3 Ibonus ·(1+β ·A(t)) (3)

′ Pexpand = Pexpand · max(ϵ, 1 − d/dmax )

(5)

Here Pexpand is the base expansion probability, Pexpand ∈ [0.1, 0.8], dmax = 6 is a hard depth ceiling, and ϵ = 0.1 is a probability floor. The linear decay function decreases the expected branching from approximately 3.0 children at the root node down to 0.3 at max depth, resulting in an analytic expected number of nodes of approximately 22.74 compared to 15,625 for an unconstrained tree of the same depth, a theoretical reduction of approximately 687-fold. Empirical validation confirms a node growth factor of 1.00x over 25 generations. D. Machine Learning Models The threat scoring module employs the bidirectional long short-term memory (BiLSTM) neural net (Hochreiter & Schmidhuber, 1997) trained on a balanced dataset containing 50,000 examples collected from the CSIC HTTP corpus, NSL-KDD, and a crafted honeypot interaction dataset. The architecture consists of a 128-D character embedding layer, a 256-unit BiLSTM, dropout of rate 0.3, a 128-unit dense layer with ReLU activation, a second dropout of rate 0.2, and a seven-class softmax output layer. Model weights are around 50 MB in PyTorch format. Language generation uses the Qwen3.50.8B model, deployed via its official GGUF checkpoint (Qwen Team, 2026; Unsloth, 2026), adapted to honeypot terminal interactions using the low-rank approximation method (Hu et al., 2022) of rank r = 8–16 and scaling parameter α = 16, conducted over six hundred optimization iterations, and trained on 2,400 examples of attacker-terminal interaction pairs.

Fig. 1. Chameleon five-layer adaptive honeypot architecture showing the feedback-coupled path from BiLSTM classifier through TC-PSO and S-RRT to the Qwen3.5-0.8B language model.

E. Supporting Infrastructure The production deployment is supported by a fourcomponent supporting stack. All session records have tamperevident forensic continuity through a SHA-256 Merkle hash chain periodically committed to the Ethereum Sepolia test network (Ethereum Foundation, 2026). Honeytoken canary files embedded within the simulated filesystem trigger automated STIX 2.1 threat bundle generation (OASIS Open, 2021) upon access, allowing direct ingestion into downstream security information and event management platforms. A Paramikobased SSH listener on port 2222 provides full pseudo-terminal emulation for capturing interactive sessions. The four-stage input normalisation pipeline detailed in Section 3.1 also serves as a defensive boundary against encoding-based classifier evasion.

to-head algorithmic comparison benchmarks—was used as the primary validation instrument.1 4. R ESULTS A. TC-PSO Benchmark Results Five replicated runs at A(t) = 0.85 produced a mean final fitness of 3.85 for TC-PSO versus 2.60 for the standardPSO reference, a relative gain of 48.1%. Convergence speed improved by 32.7% over the same comparison, and TC-PSO produced the superior outcome in every individual run. All differences were significant at p < 0.01 across the five-run ensemble. Table 1 presents the summary comparison. B. S-RRT Benchmark Results

For five replicated runs with seed values 42 through 46, SRRT achieved a best-run fitness of 1,615.8 versus 450.2 for the standard-RRT benchmark, a 258.9% improvement. At critical payload severity (Ψ = 3.0), S-RRT provided a fitness value of F. Experimental Protocol 1,135.75 as opposed to 264.64 for standard RRT, a relative gain All reported experiments were executed on Apple M-Series of 329.2%. A 24.9% reduction in memory consumption was ARM64 hardware provisioned with 16 GB of unified memory obtained with respect to the unconstrained baseline, confirming running macOS, with Python 3.14.0 and pytest 9.0.2 as the the bounding effect of Equation (5). All differences reported runtime environment. No GPU acceleration was employed were statistically significant (p < 0.01). Table 2 shows the at any stage. Five replicated runs were conducted under comparison. The exponential scaling predicted analytically from Equation seeds 42 through 46 for both TC-PSO and S-RRT. TC(4) was verified by varying Ψ between 1.0 and 3.0. At low PSO benchmarking held A(t) = 0.85 throughout; S-RRT was severity, the performance differences between S-RRT and characterized across the full payload severity range Ψ = 1.0 to standard RRT were modest. At critical severity (Ψ = 3.0), 3.0, with critical-severity results reported at Ψ = 3.0. Statistical the difference increased considerably, with S-RRT showing reliability was assessed at the p < 0.01 significance threshold across all five runs. A four-module deterministic test suite— 1 Test suite and benchmark scripts are released with the MIT-licensed source covering equation correctness and boundary behavior (TC-PSO repository to enable independent replication: https://github.com/RohitSwami3 and S-RRT), convergence and memory-bound proofs, and head- 3/Chameleon-cybersecurity-ml

Fig. 2. PSO convergence: adaptive tarpitting delay optimization across 100 iterations showing swarm mean, global best, and standard deviation band converging to the 4.5 s optimal delay. TABLE I TC-PSO VERSUS STANDARD PSO BENCHMARK SUMMARY (A(t) = 0.85, 5 INDEPENDENT RUNS , SEEDS 42–46). Metric Average final fitness Convergence speed Statistical basis

Std. PSO 2.60 baseline 5 runs (seeds 42–46)

TC-PSO 3.85 faster 5 runs (seeds 42–46)

∆ +48.1% +32.7% p < 0.01

TABLE II S-RRT VERSUS STANDARD RRT BENCHMARK SUMMARY (5 INDEPENDENT RUNS , SEEDS 42–46). Metric Best-run fitness Critical-severity fitness Memory usage Statistical basis

Std. RRT 450.2 264.64 baseline 5 runs (seeds 42–46)

a 329.2% fitness advantage (1,135.75 vs. 264.64), consistent with the theoretical pheromone amplification of 7.39 at this severity. C. Classification and Language Model Performance The BiLSTM classifier was evaluated on the combined CSIC HTTP, NSL-KDD, and honeypot-interaction test partition. It achieves an accuracy of 99.61%, with macro precision close to 99.5%, macro recall close to 99.4%, macro-F1 close to 99.45%, and a false-positive rate of 0.39% at a CPU inference time close to two milliseconds. This is comparable to the 99.84% accuracy benchmark reported by Dai et al. (2024) for an isolated CNN-

S-RRT 1,615.8 1,135.75 reduced 5 runs (seeds 42–46)

∆ +258.9% +329.2% −24.9% p < 0.01

BiLSTM-attention hybrid architecture; unlike that standalone classifier, however, Chameleon’s BiLSTM stage is directly coupled to the real-time deception and optimisation subsystems via the anomaly score A(t). Qwen3.5-0.8B (Qwen Team, 2026; Unsloth, 2026) achieved 90% contextual generation accuracy under expert review, a result not directly comparable to the cosine-similarity-based evaluation (0.695) reported by Otal and Canbaz (2024) for a fine-tuned Llama-3-8B configuration. The BiLSTM stage in the cascaded pipeline processes about ninetyfive percent of sessions at sub-two-millisecond latency, and the remaining five percent trigger language-model inference at fifty

Fig. 3. BiLSTM training and validation loss (left) and accuracy (right) across 8 epochs on the 50,000-sample corpus, showing convergence to >99.5% validation accuracy.

Fig. 4. TC-PSO versus standard PSO: fitness convergence (left) and optimal delay convergence (right) across five independent runs (seeds 42–46) at A(t) = 0.85.

to one hundred milliseconds, giving a weighted average pipeline latency of about 4.5 milliseconds and combined accuracy of almost 99.5%. D. Validation Suite The deterministic test suite ran with no failures across all validation modules: twenty-four tests for equation correctness and boundary coverage of Equations (2) and (3); twenty-seven tests for Equations (4) and (5) together with Ψ scaling; thirtyone tests for TC-PSO convergence proofs and S-RRT memorybound proofs; and nine tests for head-to-head benchmarking, for a total of ninety-one tests. An operationally acknowledged edge case, in which low inertia at A(t) = 1.0 can cause early single-run convergence under severe conditions, is not counted in the pass/fail tally and is handled in production via multi-run aggregation.

5. D ISCUSSION A. Mechanistic Basis of TC-PSO Performance The 48.1% fitness advantage and 32.7% convergence acceleration reported for TC-PSO are both the product of two coupled mechanisms. Under low-threat conditions, A(t) is near zero and inertia is kept at its conventional value of 0.729, preserving the wide exploratory sweep needed to avoid committing too soon to locally optimal delay configurations. As perceived threat severity increases towards A(t) = 1.0, inertia is progressively decreased to 0.365, shifting the search trajectory towards longer connection-holding delays that maximise dwell time with authentic adversaries. Simultaneously, the magnified objective function of Equation (3) increases the gradient of the fitness landscape by as much as 30% for maximum-severity sessions, strengthening convergence pressure in the directions

Fig. 5. S-RRT versus standard RRT: node count over 20 generations (left) and final performance comparison at Ψ = 3.0 (right), confirming bounded memory growth and 329.2% critical fitness advantage.

whereas S-RRT amplifies reinforcement in proportion to Ψ. A payload scored at maximum severity (Ψ = 3.0) deposits 7.39 times the pheromone of a benign probe (Ψ = 1.0), and this disparity cascades through the tree’s fitness computation as a multiplicative rather than additive term. High-severity events therefore produce path-amplified learning that saturates branches of the tree oriented towards adversarially relevant deception schemata, directly addressing the severity-agnostic update rule identified here as a general limitation of standard RRT-based deception topology approaches. C. Operational Cost

of greatest operational importance. The single-objective, fixedinertia formulation of PSO used in general-purpose intrusiondetection applications, such as that of Benmalek and Seddiki (2025), lacks both mechanisms, which explains the observed performance gap in comparative testing.

The current estimate for deploying Chameleon on a singlecore cloud instance with 4 GB of RAM is USD 17 per month, compared with entry-level commercial deception platforms costing about USD 8,333 per month, consistent with current industry pricing for enterprise deception technology (MarketsandMarkets, 2024). Running Qwen3.5-0.8B locally, rather than through a cloud API, sidesteps the per-token inference cost that would otherwise be paid continuously during attacker engagement sessions. Adding Merkle roots to the Ethereum Sepolia test network adds negligible gas overhead while providing legally defensible, independently verifiable audit continuity. The resulting cost ratio—Chameleon is about 490 times cheaper—does not come at the expense of the ML-driven adaptivity that distinguishes Chameleon from the open-source and commercial alternatives explored in Section 2.

B. Mechanistic Basis of S-RRT Performance

D. Comparison with Alternative Optimization Approaches

The rise from a 258.9% best-run fitness advantage to 329.2% at critical severity arises from the multiplicative nature of exponential pheromone reinforcement. Standard RRT deposits equal pheromone regardless of whether the triggering event was a benign reconnaissance probe or an active exploitation attempt,

TC-PSO and S-RRT are distinct from other optimisation techniques used in network security. Genetic algorithms have been used to tune IDS hyperparameters, achieving accuracy exceeding 98% for Random Forest and Decision Tree classifiers (Bakır & Ceviz, 2024), but cannot adapt their search based

Fig. 6. BiLSTM classifier confusion matrix on the 50,000-sample test set evaluation, showing 4,543 true benign, 3,728 true malicious, with only 20 false positives and 12 false negatives.

on live classifier outputs; Figure 7 illustrates GA evolution 6. C ONCLUSIONS behavior on the deception-schema filesystem optimization Three results of note emerge from this investigation. Beproblem for reference. haviorally adaptive deception infrastructure is achievable at Simulated annealing (Černý, 1985) can, in principle, locate a fraction of the cost of proprietary commercial alternatives: the global optimum given a sufficiently slow cooling schedule, Chameleon demonstrates that a CPU-only deployment costing but has no mechanism for tuning its search to real-time threat approximately USD 17 per month can surpass the adaptive levels. The differential evolution algorithm and its variant JADE response capability of platforms licensed at USD 100,000– (Storn & Price, 1997; Zhang & Sanderson, 2009) can reach 150,000 annually. Domain-specific meta-heuristic formulation 97.1% detection accuracy in network security but were not matters: standard PSO and standard RRT, applied without designed for cybersecurity threats specifically. The algorithm modification to the honeypot optimization problem, produce most similar to S-RRT is ant colony optimisation (Dorigo, substantially inferior fitness outcomes compared with TC-PSO Maniezzo, & Colorni, 1996), in that both use pheromones to and S-RRT, which differ from their baselines only through guide search; however, standard ACO has no depth limit, so the addition of threat-aware inertia scaling and exponential the tree grows uncontrolled during heavy attacks—a problem pheromone weighting, respectively. And the performance addressed here by Equation (5). Deep reinforcement learning advantage is statistically robust: all reported improvements— (Ma, Li, Wang, & Wang, 2025) can learn complex defence TC-PSO at +48.1% mean fitness (2.60 to 3.85) and +32.7% strategies but usually requires a very large amount of interaction convergence gain; S-RRT at +258.9% best-run fitness (450.2 data to converge, making these approaches less suitable for to 1,615.8) and +329.2% critical-severity fitness (264.64 to honeypots that must adapt within milliseconds. TC-PSO, by 1,135.75) with 24.9% memory reduction—hold at p < 0.01 contrast, converges in just 50 iterations and achieves a 100% across five independent replication runs (seeds 42–46). The success rate at A(t) = 0.85 across all five benchmark runs, while BiLSTM and Qwen3.5-0.8B cascade that feeds both optimizers S-RRT converges within fifteen generations while maintaining achieves approximately 99.5% combined classification accuracy a hard, analytically verified memory ceiling. None of the other at a mean latency near 4.5 milliseconds on CPU-only hardware, optimizers reviewed achieves both properties simultaneously. It with the full production stack validated by a ninety-one-test should be noted that these comparisons rely on results reported deterministic suite. Chameleon is openly released under the in the respective original papers rather than a direct head-to- MIT license. The authors anticipate that its availability will head test on the Chameleon benchmark; conducting such a lower the barrier to adaptive deception research and contribute controlled empirical comparison against GA, SA, DE/JADE, to a more empirically grounded literature on meta-heuristic ACO, and DQN under identical experimental conditions is design for cybersecurity applications. identified as a high-priority avenue for future investigation. R EFERENCES Apruzzese, G., Laskov, P., Montes de Oca, E., Mallouli, W., Brdalo Rapa, L., Grammatopoulos, A. V., & Di Franco, F. E. Limitations (2023). The role of machine learning in cybersecurity. Digital Four limitations are identified based on the scope of Threats: Research and Practice, 4(1), Article 8. Bakır, H., & Ceviz, Ö. (2024). Empirical enhancement of current results. First, LoRA fine-tuning of Qwen3.5-0.8B was intrusion detection systems: A comprehensive approach with carried out for six hundred iterations, attaining 90% contextual genetic algorithm-based hyperparameter tuning and hybrid accuracy; extrapolating the same training trajectory suggests 94– feature selection. Arabian Journal for Science and Engineering, 96% accuracy at two thousand iterations, though this run was 49(9), 13025–13043. https://doi.org/10.1007/s13369-024-089 not executed owing to limited computational resources. Second, all experiments were conducted in simulation; validating these 49-z Benmalek, M., & Seddiki, A. (2025). Particle swarm results against real-world attacks requires live deployment optimization-enhanced machine learning and deep learning with appropriate ethical approval. Third, TC-PSO and S-RRT techniques for Internet of Things intrusion detection. Data are currently single-objective; extending them to a multiScience and Management, 8, 423–435. https://doi.org/10.1016/ objective framework using NSGA-II (Deb, Pratap, Agarwal, j.dsm.2025.02.005 & Meyarivan, 2002) to jointly optimize connection delay and Černý, V. (1985). Thermodynamical approach to the traveling response quality represents a natural extension. Fourth, the salesman problem: An efficient simulation algorithm. Journal Payload Severity Index currently uses discrete values between of Optimization Theory and Applications, 45(1), 41–51. 1.0 and 3.0; calibrating it on a continuous scale using a labeled Cowrie Project. (2023). Cowrie SSH/Telnet honeypot [Softdataset would improve its precision. Future work includes neural fitness approximation to reduce the computational cost ware]. GitHub. https://github.com/cowrie/cowrie per iteration by an estimated 10–50%, federated TC-PSO across Dai, W., Li, X., Ji, W., & He, S. (2024). Network intrusion distributed honeypots for privacy-preserving threat intelligence, detection method based on CNN, BiLSTM, and attention and benchmarking reinforcement learning against TC-PSO for mechanism. IEEE Access, 12, 53099–53111. https://doi.or connection-delay optimization. g/10.1109/ACCESS.2024.3384528

Fig. 7. GA evolution: deception schema filesystem optimization showing target file frequency and average population fitness across 16 generations, illustrating convergence behavior of genetic algorithm baselines.

Deb, K., Pratap, A., Agarwal, S., & Meyarivan, T. (2002). LaValle, S. M. (2006). Planning algorithms. Cambridge A fast and elitist multiobjective genetic algorithm: NSGA-II. University Press. IEEE Transactions on Evolutionary Computation, 6(2), 182– Ma, Y., Li, C., Wang, Y., & Wang, Y. (2025). Application 197. of deep reinforcement learning algorithms for automatic threat Dorigo, M., Maniezzo, V., & Colorni, A. (1996). Ant detection and response in dynamic network environments to system: Optimization by a colony of cooperating agents. IEEE improve cybersecurity. Journal of Computational Methods in Transactions on Systems, Man, and Cybernetics, Part B, 26(1), Sciences and Engineering, 25(3), 2112–2125. https://doi.org/ 29–41. 10.1177/14727978241309550 Ethereum Foundation. (2026). Networks: Sepolia testnet Manocchio, L. D., Layeghy, S., Lo, W. W., Kulatilleke, G. documentation. ethereum.org. https://ethereum.org/developers/ K., Sarhan, M., & Portmann, M. (2024). FlowTransformer: docs/networks/ A transformer framework for flow-based network intrusion Fraunholz, D., Reti, D., Duque Antón, S., & Schotten, H. D. detection systems. Expert Systems with Applications, 241, (2018). Cloxy: A context-aware deception-as-a-service reverse Article 122564. https://doi.org/10.1016/j.eswa.2023.122564 proxy for web services. Proceedings of the 5th ACM Workshop MarketsandMarkets. (2024). Deception technology market— on Moving Target Defense (MTD’18), 40–47. https://doi.org/ Global forecast to 2029 (Market Report TC 5768). Market10.1145/3268966.3268973 sandMarkets Research. Hochreiter, S., & Schmidhuber, J. (1997). Long short-term OASIS Open. (2021). STIX 2.1 specification: Structured memory. Neural Computation, 9(8), 1735–1780. threat information eXpression. OASIS Standard. Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, Otal, H. T., & Canbaz, M. A. (2024). LLM honeypot: S., Wang, L., & Chen, W. (2022). LoRA: Low-rank adaptation Leveraging large language models as advanced interactive of large language models. Proceedings of the International honeypot systems. 2024 IEEE Conference on Communications Conference on Learning Representations. and Network Security (CNS), 1–6. https://doi.org/10.1109/CN Kennedy, J., & Eberhart, R. (1995). Particle swarm opti- S62487.2024.10735607 mization. Proceedings of the IEEE International Conference Ponemon Institute. (2023). The economics of security on Neural Networks, 4, 1942–1948. operations centers: What is the true cost for effective SOC?

Ponemon Institute. Provos, N. (2004). A virtual honeypot framework. Proceedings of the USENIX Security Symposium, 1–14. Qwen Team. (2026). Qwen3.5-0.8B [Large language model]. Hugging Face. https://huggingface.co/Qwen/Qwen3.5-0.8B Shi, Y., & Eberhart, R. (1998). A modified particle swarm optimizer. Proceedings of the IEEE Congress on Evolutionary Computation, 69–73. Spitzner, L. (2003). Honeypots: Tracking hackers. AddisonWesley. Storn, R., & Price, K. (1997). Differential evolution—A simple and efficient heuristic for global optimization over continuous spaces. Journal of Global Optimization, 11, 341– 359. T-Pot Community. (2023). T-Pot: All-in-one multi-honeypot platform [Software]. GitHub. https://github.com/telekom-secur ity/tpotce Unsloth. (2026). Qwen3.5-0.8B-GGUF [Large language model, GGUF format]. Hugging Face. https://huggingfac e.co/unsloth/Qwen3.5-0.8B-GGUF Zhang, J., & Sanderson, A. C. (2009). JADE: Adaptive differential evolution with optional external archive. IEEE Transactions on Evolutionary Computation, 13(5), 945–958.

Related documents

• Chameleon: An Adaptive AI-Driven Honeypot Architecture Using Threat-Calibrated Particle Swarm Optimization and Semantic Deception Rapidly-Exploring Random Trees#615372
• Chameleon: An Adaptive AI-Driven Honeypot Architecture Using Threat-Calibrated Particle Swarm Optimization and Semantic Deception Rapidly-Exploring Random Trees#614405
• The Human Oversight approaches at the forefront of responsible and trustworthy AI, from data-centric adaptive AI-Ops pipelines to Multi-Agent Systems#256690
• AI-Powered Penetration Testing: How Autonomous Offensive Security WorksWhitepaperArtificial Intelligence4 Sep 2026Chris Dale#654785
• TrustAI: A Multi-Source Artificial Intelligence Verification Platform with Authority-Weighted Trust Scoring, Retrieval-First Caching, and Domain-Adaptive API Routing#299558