PaperlayerView exact revision on arXiv

Source-linked AI summary

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, Sri Muthu

arXiv:2608.15407v1cs.CRcs.AIcs.NE

TL;DR

Traditional honeypots expose fixed behavioral profiles, while their response engines lack real-time model-driven adaptation. Chameleon integrates threat classification, language-model feedback, and domain-specific optimization, outperforming standard PSO and RRT benchmarks across replicated simulations.

  • Problem

    Openly available honeypots use deployment-time behavioral outputs that cannot adapt to evolving sessions, allowing informed attackers to identify deception with minimal probing.

  • Method

    Chameleon couples a BiLSTM threat classifier and compact language model with threat-calibrated PSO and semantic-deception RRT engines for adaptive honeypot behavior.

  • Results

    258.9% improvement in best-run fitness over standard RRT, while TC-PSO and S-RRT both significantly outperformed their respective standard baselines across five replicated runs.

  • Takeaways & Limitations

    CPU-only adaptive deception can outperform unmodified PSO and RRT while costing approximately USD 17 per month versus proprietary commercial alternatives.

  • Takeaways & Limitations

    All experiments were conducted in simulation, so validation against real-world attacks requires live deployment with appropriate ethical approval.

Abstract

from arXiv · show

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 domain-specific 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.

1. INTRODUCTION · 2. BACKGROUND AND RELATED WORK · A. Static Honeypot Platforms

The paper frames alert overload and reactive perimeter defenses as limits on conventional security operations, while static honeypots remain vulnerable to behavioral fingerprinting. Chameleon is introduced as an adaptive, production-complete honeypot that couples threat classification, language-model feedback, and meta-heuristic deception control.

  • 1. INTRODUCTION: 11,000+ daily security events leave roughly 70% without analyst attention, making alert fatigue a structural security-operations liability.The cited industry reporting describes constrained triage capacity across individual organizations.
  • 1. INTRODUCTION: Signature-driven intrusion detection and endpoint platforms depend on catalogued threat patterns, limiting visibility against living-off-the-land movement and zero-day exploitation.These perimeter-oriented defenses are characterized as reactive and post-event.
  • 1. INTRODUCTION: Deception infrastructure redistributes defender-attacker asymmetry by presenting operationally genuine environments that expose adversarial tactics while concealing session control.The passage identifies attacker unawareness as central to forensic observation.
  • 1. INTRODUCTION: Chameleon introduces TC-PSO and S-RRT as domain-specific contributions for anomaly-governed swarm behavior and language-model-scaled deception-tree pheromone reinforcement.The passage presents TC-PSO as governing particle inertia and objective amplification, and S-RRT as using exponentially scaled reinforcement with depth-decay memory control.
  • 1. INTRODUCTION: 99.5% combined accuracy at approximately 4.5 milliseconds mean CPU latency characterizes Chameleon’s feedback-coupled BiLSTM classifier and compact Qwen3.5-0.8B cascade.The passage also places this cascade within a production-complete instrumentation stack.
  • A. Static Honeypot Platforms: 2004 formalized virtual honeypots as software-emulated services that attract and contain adversarial traffic with less operational risk than physical decoys.Provos is credited with first formalizing the virtual honeypot paradigm.
  • A. Static Honeypot Platforms: Behavioral fingerprintability is identified as the primary impediment to sustained deception because deterministic response patterns enable rapid sensor identification.Spitzner’s interaction-level taxonomy is cited for this characterization.

B. Language-Model-Enhanced Terminal Emulation · C. Swarm and Trajectory-Planning Optimizers in Security · D. Hybrid Classification Pipelines and Commercial Platforms

The reviewed work spans language-model terminal emulation, swarm and trajectory-planning optimization, and hybrid detection or commercial deception platforms. Chameleon is distinguished by compact contextual generation, severity-aware optimization, and feedback integration across deception subsystems.

  • B. Language-Model-Enhanced Terminal Emulation: 90% contextual generation accuracy and 4.5 milliseconds average latency were achieved by Chameleon’s fine-tuned Qwen3.5-0.8B terminal-emulation model.The model uses low-rank adaptation and is substantially smaller than the fine-tuned Llama-3-8B model reported by Otal and Canbaz.
  • B. Language-Model-Enhanced Terminal Emulation: 0.695 cosine similarity was reported for generated versus expected Cowrie terminal outputs using a fine-tuned Llama-3-8B interaction engine.This result comes from Otal and Canbaz (2024).
  • C. Swarm and Trajectory-Planning Optimizers in Security: 99.85% binary-classification accuracy was reported for a PSO-optimized CatBoost classifier on the RT IoT2022 dataset.Benmalek and Seddiki (2025) reported this result, with PSO outperforming baseline methods such as QAE-f16 by 2.6%.
  • C. Swarm and Trajectory-Planning Optimizers in Security: PSO’s canonical inertia setting is w = 0.729, while prior formulations treat all traffic identically regardless of assessed severity.The passage states that no prior work couples swarm inertia to a live anomaly-scoring model in a honeypot engagement setting.
  • D. Hybrid Classification Pipelines and Commercial Platforms: 99.84% accuracy was achieved by a CNN-BiLSTM classifier with attention on the CIC-DDoS2019 dataset.Dai et al. (2024) describe this system as a self-contained detector without connections to deception or optimization subsystems.
  • D. Hybrid Classification Pipelines and Commercial Platforms: Transformer-based anomaly detectors show competitive accuracy on flow-based network-traffic benchmarks, but lack feedback integration with deception or optimization subsystems.The passage attributes this finding to Manocchio et al. (2024).

3. MATERIAL AND METHODS · A. System Overview · B. Threat-Calibrated Particle Swarm Optimization

Chameleon implements a closed adaptive feedback architecture in which BiLSTM anomaly scores guide parallel optimization of connection-holding delays and deception filesystem trees. TC-PSO calibrates exploration and reward amplification to threat severity while preserving minimum search capacity.

  • A. System Overview: 8,400 Python lines and 2,200 JavaScript/TypeScript lines implement Chameleon, including 1,882 lines for meta-heuristic optimization.The platform uses a four-stage payload-normalization pipeline before classification.
  • A. System Overview: θ = 0.85 separates sessions escalated to the deception engine from those receiving lightweight static responses.The threshold keeps latency below two milliseconds for escalated classification handling.
  • A. System Overview: Qwen3.5-0.8B is locally hosted, low-rank adapted at rank 8 to 16, and trained across six hundred iterations for contextual deception responses.Testing uses a simulated Ubuntu 22.04 host seeded with honeytoken canary files and preserves up to twenty command-response turns per source-IP.
  • A. System Overview: BiLSTM anomaly score A(t) drives parallel TC-PSO connection-delay tuning and S-RRT deception-filesystem-tree evolution in Chameleon’s closed optimization loop.TC-PSO maintains one particle swarm per attack category, while S-RRT maintains a population of deception filesystem trees.
  • B. Threat-Calibrated Particle Swarm Optimization: w(t) declines from 0.729 at A(t) = 0 to 0.365 at A(t) = 1.0, with a lower bound preserving exploratory capacity under maximum severity.TC-PSO replaces fixed inertia with a classifier-coupled term using wbase = 0.729, σmin = 0.3, and α = 0.5.
  • B. Threat-Calibrated Particle Swarm Optimization: 30% is the maximum objective-value amplification at A(t) = 1.0, concentrating optimization pressure on prolonged engagement with genuine adversaries.The multiplicative term is (1 + β · A(t)) with β = 0.3; the reward uses w1 = 0.65, w2 = 2.5, w3 = 0.25, and Ibonus = 0.25 × (commands −5).

C. Semantic Deception Rapidly-Exploring Random Trees

S-RRT adapts deception-tree pheromone reinforcement to payload severity while constraining memory through depth-dependent expansion decay. It preserves standard-RRT behavior at minimum severity and strongly amplifies reinforcement for high-severity engagements.

  • Severity-calibrated reinforcement: S-RRT uses Qwen-assessed Payload Severity Index Ψ ∈[1.0, 3.0] to scale pheromone reinforcement for deception-schema evolution.The index evaluates intercepted command sequences through syntactic and semantic analysis.
  • Severity-calibrated reinforcement: At Ψ = 1.0, S-RRT exactly recovers the standard-RRT update, while Ψ = 3.0 amplifies pheromone reinforcement 7.39-fold.This provides backward compatibility and targets multi-stage, high-severity engagement profiles.
  • Memory-constrained expansion: 22.74 expected nodes versus 15,625 for an unconstrained tree of the same depth yields an approximately 687-fold theoretical reduction.Depth-dependent expansion decay uses dmax = 6 and reduces expected branching from approximately 3.0 children at the root to 0.3 at maximum depth.

D. Machine Learning Models

The threat scoring module uses a BiLSTM trained on a balanced 50,000-example dataset and configured with layered embedding, recurrent, dropout, dense, and seven-class output components. The architecture couples the BiLSTM classifier to TC-PSO, S-RRT, and the Qwen3.5-0.8B language model.

  • 50,000 balanced examples from CSIC HTTP, NSL-KDD, and a crafted honeypot interaction dataset train the threat-scoring BiLSTM.
  • The model comprises a 128-D character embedding, 256-unit BiLSTM, dropout rates of 0.3 and 0.2, a 128-unit ReLU dense layer, and seven-class softmax output.
  • The five-layer architecture routes BiLSTM classifier feedback through TC-PSO and S-RRT to the Qwen3.5-0.8B language model.

E. Supporting Infrastructure · F. Experimental Protocol

The supporting infrastructure preserves forensic continuity and converts honeytoken access into standardized threat intelligence. Experiments used reproducible, CPU-only Apple M-Series hardware and a deterministic test suite spanning correctness, convergence, memory, and algorithmic comparisons.

  • E. Supporting Infrastructure: A SHA-256 Merkle hash chain provides tamper-evident continuity for all session records.The chain is periodically committed to the Ethereum Sepolia test network.
  • E. Supporting Infrastructure: Honeytoken canary files trigger automated STIX 2.1 threat-bundle generation when accessed.The bundles support direct ingestion into downstream security information and event management platforms.
  • F. Experimental Protocol: Experiments ran on Apple M-Series ARM64 hardware with 16 GB of unified memory under macOS.Python 3.14.0 and pytest 9.0.2 were used as the runtime environment.
  • F. Experimental Protocol: No GPU acceleration was employed at any stage of the experiments.
  • F. Experimental Protocol: Five replicated runs used seeds 42 through 46 for both TC-PSO and S-RRT.
  • F. Experimental Protocol: A four-module deterministic test suite validated equation correctness, boundary behavior, convergence, memory bounds, and algorithmic comparisons.

4. RESULTS · A. TC-PSO Benchmark Results · B. S-RRT Benchmark Results

Across five-run benchmarks, TC-PSO improved adaptive tarpitting optimization over standard PSO, while S-RRT substantially improved deception-schema fitness over standard RRT, especially at critical severity. The results also report bounded memory use, statistically significant differences, and released replication materials.

  • A. TC-PSO Benchmark Results: 48.1% relative gain: TC-PSO achieved mean final fitness of 3.85 versus 2.60 for standard PSO across five runs.TC-PSO produced the superior outcome in every individual run.
  • A. TC-PSO Benchmark Results: 32.7% convergence-speed improvement: TC-PSO outperformed standard PSO, with all five-run differences significant at p < 0.01.The comparison used A(t) = 0.85.
  • B. S-RRT Benchmark Results: 258.9% improvement: S-RRT achieved best-run fitness of 1,615.8 versus 450.2 for standard RRT across seeds 42 through 46.All reported differences were statistically significant.
  • B. S-RRT Benchmark Results: 24.9% memory reduction: S-RRT reduced memory consumption relative to the unconstrained baseline, confirming Equation (5)'s bounding effect.Performance differences were modest at low severity but increased considerably at Ψ = 3.0.
  • B. S-RRT Benchmark Results: Replication materials: test-suite and benchmark scripts were released with the MIT-licensed source repository.The repository is identified as github.com/RohitSwami3 3/Chameleon-cybersecurity-ml in the supplied passage.
  • A. TC-PSO Benchmark Results: 4.5 s optimal delay: the PSO convergence figure shows adaptive tarpitting-delay optimization across 100 iterations converging toward this delay.The figure reports swarm mean, global best, and a standard-deviation band.
  • B. S-RRT Benchmark Results: 329.2% fitness advantage: at critical payload severity Ψ = 3.0, S-RRT reached 1,135.75 versus 264.64 for standard RRT.This result is consistent with theoretical pheromone amplification of 7.39 at that severity.

C. Classification and Language Model Performance

Chameleon combines a highly accurate, low-latency BiLSTM threat classifier with a locally deployed Qwen3.5-0.8B model for contextual generation. The classifier’s anomaly score directly feeds the real-time deception and optimization subsystems, producing an approximately 4.5-millisecond weighted pipeline latency and nearly 99.5% combined accuracy.

  • Classification Performance: 99.61% accuracy was achieved by the BiLSTM classifier across the combined CSIC HTTP, NSL-KDD, and honeypot-interaction test partition.Macro precision was close to 99.5%, macro recall close to 99.4%, and macro-F1 close to 99.45%.
  • Classification Performance: 0.39% false-positive rate and approximately two-millisecond CPU inference time characterized the BiLSTM classifier.The classifier was evaluated on the combined CSIC HTTP, NSL-KDD, and honeypot-interaction test partition.
  • Language Model Performance: 90% contextual generation accuracy was achieved by Qwen3.5-0.8B under expert review.This result was not directly comparable to the cosine-similarity evaluation of 0.695 reported for a fine-tuned Llama-3-8B configuration.
  • Integrated Performance: A(t) directly couples the BiLSTM stage to Chameleon’s real-time deception and optimization subsystems.The anomaly score enables the classifier output to participate in the adaptive system rather than operating as a standalone classifier.

D. Validation Suite · 5. DISCUSSION · A. Mechanistic Basis of TC-PSO Performance

The validation suite passed all 91 deterministic tests, while the discussion attributes TC-PSO’s gains to threat-calibrated inertia and objective adaptation that alter exploration as severity rises. Supporting results include bounded S-RRT memory, a 329.2% critical-fitness advantage, and BiLSTM evaluation with few classification errors.

  • D. Validation Suite: 91 deterministic tests ran without failures across equation, scaling, convergence, memory-bound, and benchmarking validation modules.The suite included 24, 27, 31, and 9 tests across those modules, respectively.
  • A. Mechanistic Basis of TC-PSO Performance: 48.1% fitness advantage and 32.7% convergence acceleration for TC-PSO arise from two coupled threat-responsive mechanisms.Under low threat, inertia remains 0.729 for broad exploration; as severity approaches A(t) = 1.0, it decreases to 0.365.
  • A. Mechanistic Basis of TC-PSO Performance: TC-PSO’s mechanistic design preserves exploratory search under low threat and shifts trajectories toward longer connection-holding delays as perceived severity increases.The reported inertia schedule links anomaly severity to the search trajectory rather than maintaining fixed inertia.
  • A. Mechanistic Basis of TC-PSO Performance: 329.2% critical fitness advantage is shown for S-RRT over standard RRT at Ψ = 3.0, alongside bounded memory growth over 20 generations.The figure compares node count across generations and final performance at Ψ = 3.0.
  • A. Mechanistic Basis of TC-PSO Performance: 20 false positives and 12 false negatives occurred in the BiLSTM classifier’s 50,000-sample test-set evaluation.The confusion matrix also shows 4,543 true benign and 3,728 true malicious classifications.
  • 5. DISCUSSION: The fixed-inertia, single-objective PSO formulation used in general-purpose intrusion detection lacks both threat-responsive mechanisms, explaining the comparative performance gap.The passage identifies Benmalek and Seddiki (2025) as an example of this general-purpose formulation.

B. Mechanistic Basis of S-RRT Performance · C. Operational Cost · D. Comparison with Alternative Optimization Approaches

S-RRT’s performance advantage is attributed to severity-proportional exponential pheromone reinforcement, while local deployment keeps Chameleon’s estimated operating cost far below commercial platforms. TC-PSO and S-RRT are distinguished from alternative optimizers by adapting or targeting cybersecurity-specific, real-time threat conditions.

  • B. Mechanistic Basis of S-RRT Performance: 329.2% at critical severity follows from exponential pheromone reinforcement increasing S-RRT’s advantage over standard RRT from 258.9%.The passage attributes this rise to multiplicative reinforcement rather than additive updating.
  • B. Mechanistic Basis of S-RRT Performance: 7.39 times the pheromone is deposited by a maximum-severity payload (Ψ = 3.0) compared with a benign probe (Ψ = 1.0).High-severity events produce path-amplified learning toward adversarially relevant deception schemata.
  • C. Operational Cost: USD 17 per month is the estimated cost of deploying Chameleon on a singlecore cloud instance with 4 GB of RAM, versus about USD 8,333 per month commercially.Local Qwen3.5-0.8B deployment avoids continuously paid per-token cloud inference during attacker engagement sessions.
  • D. Comparison with Alternative Optimization Approaches: Exceeding 98% accuracy, genetic algorithms tune IDS hyperparameters but cannot adapt their search based on live classifier outputs.The comparison cites Random Forest and Decision Tree classifiers.
  • D. Comparison with Alternative Optimization Approaches: Figure 7 illustrates genetic-algorithm evolution behavior on the deception-schema filesystem optimization problem for reference.The figure is presented as a reference comparison for GA behavior.
  • D. Comparison with Alternative Optimization Approaches: 97.1% detection accuracy is achievable with differential evolution and JADE, but these algorithms were not designed specifically for cybersecurity threats.Simulated annealing can in principle locate the global optimum with a sufficiently slow cooling schedule, yet lacks real-time threat-level tuning.
  • D. Comparison with Alternative Optimization Approaches: Ant colony optimisation is identified as the algorithm most similar to S-RRT.The passage introduces this similarity within the comparison of alternative optimization techniques.

E. Limitations · 6. CONCLUSIONS

The paper identifies limitations in Qwen3.5-0.8B fine-tuning and simulation-only evaluation, while concluding that Chameleon enables adaptive deception at substantially lower cost than proprietary alternatives.

  • E. Limitations: 90% contextual accuracy was attained after six hundred LoRA fine-tuning iterations for Qwen3.5-0.8B.The six-hundred-iteration training run was executed, unlike the projected longer trajectory.
  • E. Limitations: 94–96% accuracy is projected at two thousand iterations, but this trajectory was not executed because of limited computational resources.The projection extrapolates the same training trajectory rather than reporting an executed experiment.
  • E. Limitations: All experiments were conducted in simulation, leaving validation against real-world attacks for live deployment.The passage explicitly identifies real-world validation as a requirement beyond the current evaluation scope.
  • 6. CONCLUSIONS: Approximately USD 17 per month is the reported cost of CPU-only Chameleon deployment.The conclusion frames this as a fraction of the cost of proprietary commercial alternatives.
  • 6. CONCLUSIONS: USD 100,000–150,000 annually is the licensing cost cited for proprietary commercial alternatives.The comparison concerns platforms with adaptive response capability.
  • 6. CONCLUSIONS: Chameleon is described as surpassing the adaptive response capability of proprietary commercial alternatives.The conclusion attributes this result to a CPU-only deployment costing approximately USD 17 per month.
Loading 2608.15407v1…