Securing Federated Learning: Enhancing Defense Mechanisms against Poisoning Attacks
Benjamin Birchman, Geethapriya Thamilarasu · 2024
In recent years, Federated Learning (FL) has emerged as a powerful paradigm, revolutionizing machine learning by harnessing distributed computational resources. FL enables collaborative model training across multiple devices or servers without centralizing data, thereby enhancing model inference quality. However, these distributed systems are not immune to security threats, particularly poisoning attacks leading to compromised model performance and potential privacy breaches. In this paper, we develop defense strategies to mitigate poisoning attacks in federated learning systems. Specifically, we address the challenge posed by distributed backdoor poisoning attacks against existing poisoning defenses in federated learning systems. We propose a defense solution, known as Area Similarity FoolsGold (ASF), that builds upon the existing cosine similarity solution by integrating Triangle Area Similarity and Sector Area Similarity algorithms. Our solution aims to enhance the identification of maliciously compromised clients and mitigate the impact of distributed backdoor attacks in federated learning systems.