Disparity-Aware Federated Learning for Intrusion Detection Systems in Imbalanced Non-IID Settings

Abstract

Many variants of Federated Learning have been proposed to settle different challenges that come with numerous practical applications, one of which is dealing with non-IID data sources. As decentralized data sources in real life are bound to be non-IID, this is one of the hardest challenges, and yet the earliest federated algorithms struggle to resolve this issue, resulting in worse non-IID performance. Also, applications that require capturing really intricate insights from data while upholding the latest data privacy standards, such as Intrusion Detection Systems (IDS) have enabled the use of FL in those domains. In this article, we propose a novel Disparity-Aware federated learning approach that tackles non-IID and data imbalance from both global and local learning steps of FL. Our method capitalizes on state-of-the-art loss functions to tackle data imbalance at the client level and a class distribution-dependent clustering algorithm at the server to tackle class distribution skew. The nature of the process renders it applicable even in asynchronous federated learning schemes. Experiments with multiple benchmark intrusion detection datasets reveal improved performance over traditional deep learning approaches as well as earlier federated learning techniques.