Federated Learning Security Threats and Defense Approaches

Abstract

Artificial intelligence technology has developed rapidly. As a new technology, Federated learning can keep all parties' data locally and train the global model together with all data parties. Therefore, it can solve the problem of "data islands" while protecting privacy, so Federated learning is widely used. However, the existing Federated learning system still has many loopholes. For example, when uploading a local model, an attacker may mix in models with incorrect data. This requires corresponding defensive measures. Before beginning this article, we learned about the previous work related to the security threats and defense measures of Federated learning. This paper first explains the concept, advantages, and disadvantages of Federated learning. Secondly, it summarizes five common security threats in Federated learning and explains and compares various threats. Then it summarizes four defense approaches commonly used in Federated learning and explains each approach in principle. Finally, this paper looks forward to the follow-up development of defense methods in Federated learning.