Adversarially robust dynamic trust evaluation via graph-attentive reinforcement learning for scalable edge IoT security

Abstract

Edge IoT deployments must evaluate the trustworthiness of peer devices continuously, locally, and under resource constraints that static scoring models cannot accommodate—particularly when adversaries adapt their mimicry strategies in response to observed detection outcomes. This study presents the Adaptive Dynamic Trust Evaluation (ADTE) framework, with three core novelties that address the distinct failure modes of prior research. First, an edge-masked heterogeneous attention (EMHA) operator suppresses adversarially injected graph edges through a learnable perturbation-aware gate, for which we prove a certified robustness radius under ℓ ∞ -bounded node-feature perturbation and quantify mask convergence under PGD-augmented training. We additionally demonstrate empirical—though not formally certified—robustness against gradient-based edge-injection attacks and explicitly define the gap between the analytic certificate (node features) and the main graph-poisoning threat model (edge structure). Second, a non-stationary MDP formulation models the adversary’s policy as a time-varying component of the transition kernel; we derive a regret bound of O ( T Δ T ) that holds under both piece-wise stationary and continuously drifting attack distributions by applying a sliding-window restart mechanism, proving that this bound extends to partially observable adversarial policies via a POMDP belief-state reduction. Third, a continual PPO agent augmented with elastic weight consolidation (EWC) prevents catastrophic forgetting over long deployment horizons; we demonstrate stable policy performance across 5000-episode simulations spanning ≈ 3.47 continuous deployment-days at 60 s evaluation windows (equivalent to 300,000 s of gateway operation). Unlike prior trust evaluation frameworks that address, at most, two of the three sub-problems (representation, decision, and adaptation), ADTE simultaneously targets all three, thus delivering certified robustness guarantees that are absent from GNN-based approaches, non-stationary regret bounds missing from RL-based methods, and continual learning stability unavailable in either category. Evaluated on N-BaIoT, MedBIoT, and UNSW-NB15 against eight baselines, ADTE achieves 95.3% trust classification accuracy, 41.8% FPR reduction over FedTrust, and sub-12 m edge latency at 1.8 W incremental power. The 41.8% FPR reduction over FedTrust is measured under equivalent single-gateway operating conditions. These results are validated under gradient-based adaptive graph poisoning, asynchronous decentralised gateway operation, and topology churn up to 1,200 devices—conditions no prior trust evaluation framework has simultaneously addressed.