Toward Cleansing Backdoored Neural Networks in Federated Learning

Malicious clients can attack federated learning systems using compromised data during the training phase, including backdoor samples. The compromised global model will perform well on the validation dataset designed for the task, but a small subset of data with backdoor patterns may trigger the model to make a wrong prediction. In this work, we propose a new and effective method to mitigate backdoor attacks in federated learning after the training phase. Through federated pruning method, we remove redundant neurons and "backdoor neurons", which trigger misbehavior upon recognizing backdoor patterns while keeping silent when the input data is clean. The second optional fine-tuning process is designed to recover the pruning damage to the test accuracy on benign datasets. In the last step, we eliminate backdoor attacks by limiting the extreme values of inputs and neural network neurons' weights. Experiments using our defenses mechanism against the state-of-the-art Distributed Backdoor Attacks on CIFAR-10 show promising results; the averaged attack success rate drops more than 70% with less than 2% loss of test accuracy on the validation dataset. Our defense method has also outperformed the state-of-the-art pruning defense against backdoor attacks in the federated learning scenario.