TrustSense: Structure-Adaptive Trust Inference for Resilient Cooperative IoT Sensing
TL;DR
This work presents TrustSense, a structure-adaptive GNN framework for per-report trust evaluation, which represents active devices as a dynamic graph and maps each report's neighborhood-based reconstruction residual to a trust score.
Abstract
Cooperative IoT sensing must assess reports from noisy, faulty, or compromised devices despite churn and changing inter-device relationships. We present TrustSense, a structure-adaptive GNN framework for per-report trust evaluation. TrustSense represents active devices as a dynamic graph and maps each report's neighborhood-based reconstruction residual to a trust score. A layer-wise soft edge mask downweights less informative dependencies, while forced anchor attention reserves a controlled share of attention for available trusted anchors. At runtime, historical consistency drives anchor-rooted credit diffusion to prioritize returning devices, and accepted reports guide subsequent evaluations. Shared GNN parameters make TrustSense inductive, enabling targeted adaptation to unseen devices without usable history. TrustSense is also location-centric, device-agnostic, and prior-free with respect to the sensing process, assuming neither a parametric sensing-value distribution nor a known propagation model. We evaluate TrustSense on synthetic RF data and real LoED traces with injected attacks. With 10% additive-bias attackers on LoED, it improves F1 by 4.25× over the strongest evaluated baseline and achieves a lower Brier score. These gains persist with up to 50% malicious devices, while inference latency remains within 230-250 ms at several thousand devices.
