A comparative study of defense strategies for managing malicious participants and its effects in Federated Learning systems

Federated Learning (FL) is a dispersed machine learning technique that allows participants to cooperatively train a model without the need to disclose their private data. However, the distributed nature of FL introduces new threats as it increases the likelihood of compromised participants. Federated Learning is vulnerable to poisoning attacks, in which deceptive clients deliberately send harmful model updates to the server, thereby compromising the integrity of the global model. In this paper, we perform a comparative analysis of 30 different security strategies or defense mechanisms aimed at mitigating the negative impact of malicious nodes. This analysis is the first to shed light on several research gaps in addressing malicious or compromised clients within a federated learning system. We conclude the paper by outlining a set of research directions aimed at developing FL-based defense systems that can address the identified gaps.