ClusFed: A Clustering-Based Defense for Secure Federated Learning

Abstract

As IoT devices continue to proliferate, the demand for secure and efficient machine learning solutions becomes increasingly critical. Federated Learning (FL) offers a promising approach by enabling decentralized model training across distributed clients while preserving data privacy. However, FL systems are susceptible to adversarial threats, such as data and model poisoning attacks, where compromised clients send corrupted updates to undermine the global model’s performance. In this paper, we develop an innovative, lightweight Byzantine resistance strategy, called ClusFed, which uses adaptive client selection mechanisms. By leveraging clustering techniques to dynamically differentiate between honest and malicious clients, based on their local model updates. ClusFed effectively mitigates the impact of malicious nodes while sometimes exploiting their data, depending on the type of attack. This approach ensures robust performance even with up to 40% of clients being malicious. Extensive experiments on non independent, identically distributed (non-IID) partitions of MNIST, CIFAR-10, Shakespeare and a time series Water Leak datasets demonstrate that ClusFed consistently outperforms state-of-the-art Byzantine-resilient methods like FedMedian, Multi-Krum, and TrimmedMean. The results highlight ClusFed’s ability to maintain high accuracy, achieving up to 93% for MNIST and 76% for CIFAR-10, 86% for Water Leak and 40% for Shakespeare dataset, as well as stability in various attack scenarios while guaranteeing efficient learning times.