Building Trusted Federated Learning: Key Technologies and Challenges

Federated learning (FL) provides convenience for cross-domain machine learning applications and has been widely studied. However, the original FL is still vulnerable to poisoning and inference attacks, which will hinder the landing application of FL. Therefore, it is essential to design a trustworthy federation learning (TFL) to eliminate users’ anxiety. In this paper, we aim to provide a well-researched picture of the security and privacy issues in FL that can bridge the gap to TFL. Firstly, we define the desired goals and critical requirements of TFL, observe the FL model from the perspective of the adversaries and extrapolate the roles and capabilities of potential adversaries backward. Subsequently, we summarize the current mainstream attack and defense means and analyze the characteristics of the different methods. Based on a priori knowledge, we propose directions for realizing the future of TFL that deserve attention.