A highly efficient, confidential, and continuous federated learning backdoor attack strategy

Federated learning is a kind of distributed machine learning. Researchers have conducted extensive research on federated learning's security defences and backdoor attacks. However, most studies are based on the assumption federated learning participant's data obey iid (independently identically distribution). This paper will evaluate the security issues of non-iid federated learning and propose a new attack strategy. Compared with the existing attack strategy, our approach has three innovations. The first one, we conquer foolsgold [1] defences through the attacker's negotiation. In the second one, we propose a modified gradient upload strategy for fedsgd backdoor attack, which significantly improves the backdoor attack's confidentiality on the original basis. Finally, we offer a bit Trojan method to realize continuous on non-iid federated learning. We conduct extensive experiments on different datasets to illustrate our backdoor attack strategy is highly efficient, confidential, and continuous on non-iid federated learning.