A Fog-Enabled Architecture for Securing Hierarchical Federated Learning from Data Poisoning

Federated Learning (FL) enables distributed model training across multiple clients without requiring the exchange of raw data, thereby enhancing data privacy. However, FL remains vulnerable to data poisoning attacks from malicious clients. In this paper, we propose an enhanced four-layer hierarchical federated learning (HFL) architecture that integrates fog nodes between the edge and cloud layers to provide a robust, multi-tier defense against adversarial manipulation. Our approach introduces a two-layer outlier detection mechanism: (1) hard filtering at the edge using a median-based strategy to exclude malicious client updates, and (2) soft reweighting at the fog layer to attenuate the influence of suspicious edge aggregates. We formally define the system and threat models, present our defense algorithm, and provide theoretical and empirical analysis of its security advantages. Experiments on the MNIST dataset demonstrate that the proposed framework significantly improves resilience to label-flipping and data poisoning attacks. Even with 60% of clients behaving maliciously, our approach maintains an accuracy of 88%, compared to only 71% for standard HFL. Across a range of attack scenarios, the dual-layer defense consistently outperforms the baseline HFL, effectively detecting and mitigating both overt and subtle adversarial behaviors and ensuring a more stable and reliable model convergence.