Intrusion Detection via Federated Learning: Tackling Non-IID Data and Poisoning Attacks

Authors: Quan Li, Yuhong Li

Journal: IEEE International Conference on Trust, Security and Privacy in Computing and Communications (TrustCom)

Publication year: 2025

DOI: 10.1109/trustcom66490.2025.00223

Abstract

Intrusion detection based on Federated Learning has been proven to be an effective method to address security risks in the Internet of Things (IoT). However, it faces challenges in dealing with non-independent and identically distributed (non-IID) data and poisoning attacks in IoT. To address these challenges, we propose an IDS approach named Clustering and Defending Against Poisoning Attacks (CDPA). Through clustering, CDPA not only alleviates the challenges of non-IID data, but also reframes the defense against poisoning attacks as one under nearly IID conditions. CDPA uses a novel federated clustering method to ensure that malicious clients are grouped without leaking additional information, and it adaptively selects the number of clusters. Moreover, CDPA proposes a novel federated aggregation algorithm specifically designed to address the previously overlooked challenge of high malicious client ratios within clusters - a vulnerability particularly exacerbated in cluster-based training scenarios. Extensive experiments demonstrate that CDPA can effectively handle non-IID data and defend against poisoning attacks, achieving a performance gain of approximately 20% over the baseline.

View this paper on Academic Now