Fedward: A Flexible Federated Backdoor Defense Framework for Non-IID Data

Researchers proposed a federated backdoor defense framework, Fedward, aimed at thoroughly eliminating federated backdoor attacks (FBA) under Non-IID data. The framework decomposes FBA into multiple types of attacks, addressing them respectively with AmGrad sparsification and AutoOPTICS adaptive clustering. On MNIST, FMNIST, and CIFAR10, the defense effect improved by 33%~75% over clustering-based defense methods, and under extreme Non-IID the average FBA success rates were 96.98%, 90.74%, and 89.8%, respectively.