Robust Federated Learning via Weighted Median Aggregation*

Abstract

Federated learning (FL) enables multiple clients to collaboratively train a model without sharing their private data, preserving privacy. However, the iterative communication between the server and clients introduces vulnerabilities, allowing non-cooperative nodes to act adversarially and potentially compromise the entire system. These malicious clients can disrupt the learning process, causing the global model to diverge or produce erroneous predictions. In this paper, we introduce Weighted Median Aggregation (WMA), a novel defense mechanism in federated learning. WMA combines distance-based weighting with median-based aggregation to enhance robustness and efficiency, effectively mitigating the influence of malicious clients while preserving contributions from honest ones. Empirical evaluations on real-world datasets demonstrate that WMA outperforms standard methods like FedAvg, Krum, and TrimmedMean in terms of model accuracy, convergence speed, and resilience against Byzantine attacks. Under attack conditions, WMA achieves 85% accuracy, significantly surpassing baseline methods. Our findings underscore the potential of WMA in addressing key challenges in designing Byzantine-resilient federated learning systems.