Model Similarity-Based Defense Scheme Against Backdoor Attacks on Federated Learning

Federated learning is a deep learning model that helps thousands of participants train data without compromising their local data. For example, multiple smartphones can jointly learn and train the predicted value of a word under the keyboard without revealing what the user typed. However, federated learning is still fragile, and due to a large number of participating clients, it is impossible to guarantee that participants will engage in malicious behavior in the training process. Malicious actors will embed backdoors in their data, malicious clients will converge and show good accuracy in their primary task, and in a particular task, the attacker will embed backdoor input in a specific way that the attacker wants. Normally, the parameters sent by the abnormal client are completely different from those sent by the normal client. This paper proposes a new method to filter dishonest participants. The scheme mainly looks for the similarity between each client, identifies benign clients and abnormal clients by looking for the largest cluster, and then ignores the abnormal local model when the global model is updated.