Malicious Client Detection with Differential Privacy Constraints Based on Model Update Slicing and K-Means Clustering

Federated learning (FL) is a decentralized machine learning technique that protects data privacy by trainng models locally on devices and sharing only model updates instead of raw data. However, FL also faces multiple security challenges. Clientside poisoning attacks can compromise global model training by injecting malicious updates. Additionally, the server may infer private information from client model updates. Most existing studies address these two risks separately, lacking a systematic analysis and response to their combined threats. This paper proposes a method called Sliced-Clustering Guard (SCGuard), which slices gradients and performs clustering analysis on each segment, enabling the server to more accurately detect malicious updates and effectively differentiate malicious client updates with differential privacy (DP) protection. Experimental results demonstrate that SCGuard can mitigate poisoning attacks while effectively safeguarding clients' sensitive information.