An Adaptive and Behavior-Aware Aggregation Strategy for IOT-Based Hierarchical Federated Learning Systems

Abstract

Federated Learning improves privacy by keeping data decentralized, making it particularly suitable for distributed Internet of Things (IoT) environments where large numbers of edge devices collaboratively train machine learning models. However, the lack of visibility into clients’ local training processes creates vulnerabilities to Byzantine attacks that can compromise the global model. Existing robust aggregation methods still face limitations due to their stateless behavior and reliance on limited behavioral indicators, which makes them less effective against adaptive adversaries. Therefore, a dynamic, multi-dimensional, and history-aware trust mechanism is required to achieve more reliable aggregation. This thesis proposes TWA-HFL, a dynamic trust-aware aggregation framework for Hierarchical Federated

Learning that continuously evaluates client reliability throughout the training process. The proposed ap- proach combines multiple behavioral indicators, including update similarity, peer consistency, magnitude

stability, and historical behavior, into an adaptive trust score. A temporal regulation mechanism is intro- duced to monitor trust evolution over time by rewarding consistent clients and reducing the contribution of potentially malicious participants. Furthermore, trust-enhanced clustering and weighted aggregation are applied to select reliable updates and improve resilience against Byzantine attacks. Extensive experiments were conducted on MNIST, FashionMNIST, and CIFAR-10 datasets under various

attack scenarios and malicious client ratios. The experimental results demonstrate that TWA-HFL consis- tently outperforms traditional and robust aggregation baselines, achieving higher accuracy and stronger

robustness against adaptive Byzantine attacks.