FedDynamic: trust-aware secure aggregation for federated learning in dynamic IoT

Abstract

Federated learning (FL) enables collaborative model training across distributed Internet of Things (IoT) devices while preserving data privacy. However, in dynamic and open IoT environments, untrustworthy clients, adversarial attacks, and insufficient traceability hinder the robustness and security of FL systems. To address these challenges, we propose FedDynamic, a trust-aware FL framework designed for dynamic IoT scenarios. FedDynamic features dual-layer authentication and adaptive client selection. It combines device-level fingerprint verification and model watermark embedding to ensure the authenticity and integrity of local updates. A multi-metric trust scoring mechanism evaluates clients based on training efficiency, watermark consistency, and model accuracy. To handle participation dynamics, we introduce a negative-feedback-driven dynamic weighting strategy that adjusts aggregation weights by promoting reliable clients and suppressing risky ones. Notably, our experiments simulate realistic client behavior with dynamic join-and-exit patterns, reflecting the intermittent availability of IoT devices in practical deployments. Experimental results demonstrate that FedDynamic maintains high global accuracy, fast convergence, and strong robustness under adversarial and dynamic participation conditions. In addition, it achieves end-to-end update traceability and accountability, making it suitable for secure and scalable FL deployment in practical IoT systems.