FedPurge: Byzantine-Robust Federated Aggregation via Cosine Filtering and Trusted-Subset Median for 6G Edge Intrusion Detection

Abstract

Federated learning (FL) offers a compelling framework for collaborative intrusion detection across distributed 6G edge nodes without centralizing sensitive network traffic data. However, existing aggregation strategies remain vulnerable to poisoning attacks, particularly when malicious clients constitute a large fraction of the federation. In this paper, we propose FedPurge, a Byzantine-robust aggregation strategy that combines cosine-similarity-based client filtering with coordinate-wise median aggregation on a dynamically maintained trusted subset. While FedMedian fails at majority poisoning ratios due to its lack of a pre-filtering stage, FedPurge eliminates malicious contributions before aggregation, maintaining robustness even when 50% of clients are adversarial. Experiments on the UNSW-NB15 intrusion detection dataset under non-IID data distribution (Dirichlet $\alpha=0.5$) show that FedPurge maintains 0.766 accuracy under 50% malicious client participation, compared to 0.449 for both FedAvg and FedMedian, a 31.7 percentage point improvement. Ablation studies across poisoning ratios from 10% to 50% and non-IID levels from $\alpha=0.1$ to $\alpha=1.0$ confirm consistent superiority of FedPurge in high-poisoning regimes.