Research on robustness of federated learning based on pruning optimization
lei zhao, Yitong Chen · 2024
Federated learning has more flexible data ownership participants, therefore its data (sample feature vector or label) is more likely to be changed, and it is more vulnerable to data poisoning by malicious users, resulting in the final global model not getting the expected effect. This paper focuses on this data poisoning defense problem and applies the traditional centralized machine learning pruning optimization method to each client of federated learning. Each client needs to execute before each global iteration. Pruning optimization algorithm to remove abnormal data. The experimental results indicate that when the discrepancy between abnormal and normal samples is significant, the pruning optimization algorithm effectively eliminates the outliers, thereby minimizing their impact on the final federated learning model.