Privacy Preserving Federated Learning for Distributed Intrusion Detection: Differential Privacy Guarantees, NonIID Convergence, and Byzantine Robustness

Abstract

Federated learning (FL) enables collaborative intrusion detection system (IDS) training across organizations that cannot share raw network traffic due to privacy, regulatory, or competitive constraints. However, naive FL deployments are vulnerable to gradient inversion attacks that reconstruct private training data from gradient updates, suffer severe accuracy degradation under nonidentically distributed (nonIID) data partitions across heterogeneous network environments, and are susceptible to Byzantine manipulation by adversarial clients. This paper presents FedIDSDP, a differentially private federated learning framework for distributed IDS that simultaneously addresses all three challenges. The framework applies Renyi Differential Privacy (RDP) accounting with perlayer gradient clipping, a communication efficient Top gradient scarification scheme achieving 90% compression with less than 1.2% accuracy loss, and an adaptive Fed Prox regularization term dynamically calibrated to client data heterogeneity via Maximum Mean Discrepancy (MMD). Theoretical analysis establishes (epsilon=1.0, delta=10^5)DP guarantees across 200 communication rounds for five heterogeneous client organizations. Byzantine fault tolerance under 20% malicious client fraction is provided by Krum aggregation. Empirical evaluation on CICIDS2018 with Dirichlet nonIID partitioning (alpha=0.5) demonstrates FedIDSDP achieves 97.8% F1score, within 1.2 percentage points of centralized training, while providing provable privacy guarantees and 73% communication overhead reduction. Ablation studies confirm each component's contribution; distribution shift experiments validate adaptive weight updating.