BEAT: An Integrated Blockchain-Edge AI Trust Framework for IoT Security - Design, Prototype, and Performance Evaluation

Abstract

Trust management in heterogeneous Internet of Things (IoT) deployments remains unresolved because most frameworks treat blockchain reputation persistence and edge-AI anomaly detection as separate subsystems, so the graded evidence a detection model produces cannot be indexed or propagated by a ledger that was built without reference to it. This study presents BEAT (Blockchain–Edge AI Trust), which closes that gap through co-design of three coupled layers: a Graph Attention Network–Long Short-Term Memory (GAT-LSTM) edge module that converts device interaction graphs into normalised trust evidence vectors; a Hyperledger Fabric 2.5 ledger whose TrustCC chaincode enforces monotone trust updates over a formally defined trust lattice; and a Proximal Policy Optimisation (PPO) admission controller whose advantage over contextual-bandit and PID baselines is established both theoretically and experimentally. A companion gossip protocol, GARP, propagates reputation deltas across edge orchestrators with a proven logarithmic convergence guarantee. Prototyped on a 12-node Raspberry Pi 5 cluster attached to a three-peer Fabric channel and evaluated on UNSW-NB15, N-BaIoT, TON IoT, and CICIoT2023, BEAT reaches a cross-dataset mean F1 of 0.961, an on-chain throughput of 1,847 transactions per second, and a median trust-decision latency of 38 ms. BEAT outperforms a GNN-Transformer baseline (F1 0.944) and a CNN-BiLSTM-Transformer baseline (F1 0.939) on UNSW-NB15, and ablation results show that removing any single BEAT layer costs at least four F1 points cross-dataset, evidence that the system-level integration, not the detection model alone, drives the gain. Formal Byzantine resilience bounds accompany the prototype evaluation, alongside experimental Sybil-resistance and evidence-poisoning results.