Adaptive Differential Privacy for Federated Intrusion Detection in IoT Networks Using Meta-Learning Control

Abstract

Federated Learning enables distributed intrusion detection in IoT networks but suffers from gradient leakage vulnerabilities that can expose sensitive client data through reconstruction attacks. This paper presents an adaptive privacy-preserving framework that dynamically adjusts Differential Privacy parameters using a meta-learning-based controller, optimising the privacy utility trade-off throughout FL training. We have previously proposed a deep reinforcement learning-based intrusion detection framework that achieved high detection accuracy in IoT environments, but its gradient exchanges remained vulnerable to potential data leakage, motivating the need for a stronger privacy-preserving mechanism. Unlike conventional static-noise DP approaches, our method continuously modulates gradient clipping and Gaussian perturbation levels in response to learning dynamics. We evaluate the framework on three benchmark datasets: ToN-IoT, CICIOT2023, and CICIDS2017, demonstrating that adaptive DP significantly reduces gradient reconstruction vulnerability compared to unprotected FL while maintaining detection accuracies of 99.68%, 98.33%, and 97.25%, and Macro-F1 scores of 98.04%, 84.67%, and 86.08% on ToN-IoT, CICIoT2023, and CICIDS2017, respectively, under a cumulative privacy budget of 4.76. Our approach outperforms fixed-noise baselines by preserving higher detection performance at equivalent privacy levels. These findings establish adaptive privacy modulation as a practical and secure solution for federated intrusion detection in real-world IoT deployments.