TrustAdaptRL: an online reinforcement learning framework for attack-resilient trust management in Fog-IoT networks

Fog-empowered IoT networks (Fog-IoT) can support latency-sensitive, large-scale applications; however, they remain susceptible to adaptive trust attacks such as collusion, on–off, and grayhole attacks, while still requiring reliable QoS. Available trust-management schemes often use static weighting or rely on short-lived uncertain information from probabilistic or anomaly-detector-based analysis, and thus suffer from limited adaptivity to changing adversarial behaviours, ineffective security–performance management, and low scalability or poor understanding. This study aims to overcome these limitations by introducing an online reinforcement-learning framework, TrustAdaptRL, to enable attack-resilient trust management in Fog-IoT (F-IoT) networks. Behavioural, contextual, temporal, security, and cross-neighbour observations are used as inputs to formulate the long-term trust-updating decision process. An attack-aware reward function explicitly rewards grayhole, on–off, and collusive behaviours while maintaining packet-delivery and delay performance.The inference is finished at the IoT devices, while the replay-buffer management, model optimisation, and policy update are executed at the fog nodes. Evaluation is based on a public-traffic-driven simulation using realistic traffic and security attributes created in cooperation with TON_IoT, BoT-IoT, and IoT-23, as well as syntactic topology, trust labels, and attacks based on specific routes. TrustAdaptRL achieves 93.7% trust accuracy, surpassing the highest baseline by 4.3% points. The benefit of grayhole, on–off, and collusion detection increases by 6.4%, 13.8%, and 15.3%, respectively, and the packet delivery ratio is 95.1%. Cross-cluster transfer, statistical testing, and explainability analysis results further demonstrate stable, interpretable, and scalable transfer of trust within dynamic Fog-IoT environments. The complete TrustAdaptRL implementation and reproducibility materials are publicly available at: https://github.com/Pallavi-1981/TrustAdaptRL .