Fed-Fi: Federated Learning Malicious Model Detection Method Based on Feature Importance

Abstract

Federated learning, as a new security exchange paradigm, is widely used in medical care, driverless cars, finance, and other fields. However, federated learning still faces the problem of sybil attacks common in distributed frameworks. The existing schemes mainly defend against malicious model attacks with distance comparison, neural network, and confidence vote. But they are significantly limited in dealing with collusive sybil attacks. Therefore, we propose a federated learning malicious model detection method based on feature importance (Fed-Fi). Firstly, we screen important features by feature importance reasoning method based on LRP and compare the similarity based on Hamming distance between important features. Then, we adjust model learning rate adaptively to reduce the effect of collusive sybil attacks on the global model. The experimental results indicate that it can effectively resist the attack of collusive sybils in federated learning.