DBST-FL: Dynamic Behavioural and Semantic Trust for Robust Federated Learning in Industrial IoT.

Federated learning (FL) has emerged as an effective paradigm for collaborative model training in Industrial Internet of Things (IIoT) environments by enabling distributed devices to learn shared models without exchanging raw data. However, existing FL defence mechanisms predominantly rely on either behavioural analysis of client updates or semantic validation of model performance, limiting their ability to detect sophisticated poisoning and stealthy backdoor attacks that evade single-dimensional trust assessment. This paper proposes DBST-FL, a dynamic behavioural and semantic trust framework for robust federated learning in the Industrial IoT. The proposed framework evaluates each client through two complementary trust dimensions: a behavioural trust layer that measures gradient alignment, historical consistency, and collective deviation and a semantic trust layer that assesses benign utility and template-free semantic stress validation using server-side data. The two trust scores are integrated through a non-compensatory multiplicative trust fusion mechanism, ensuring that weaknesses in one trust dimension cannot be masked by strengths in the other. The resulting trust score guides a trust-aware aggregation strategy that reduces the influence of malicious participants while preserving the contributions of reliable clients. Extensive experiments are conducted on the Edge-IIoTset and UNSW-NB15 datasets using ANN, 1D-CNN, and LSTM models under multiple poisoning and backdoor attack scenarios. The proposed framework achieves overall classification performance competitive with the strongest robust aggregation baselines while consistently delivering stronger resilience against adversarial attacks and lower backdoor attack success rates than representative trust-based and Byzantine-robust aggregation methods, all while maintaining linear per-round computational complexity suitable for large-scale IIoT deployments. The results demonstrate that integrating behavioural and semantic trust within a unified aggregation framework provides an effective and scalable defence against advanced adversarial threats in federated learning.