SMSS: Secure Member Selection Strategy in Federated Learning

Abstract

Data security and user privacy-issue have become an important field. As federated learning (FL) could solve the problems from data security and privacy-issue, it starts to be applied in many different applied machine learning tasks. However, FL does not verify the quality of the data from different parties in the system. Hence, the low-quality datasets with fewer common entities can be cotrained with others. This could result in a huge amount of computing-resources waste, and the attack on the FL model from malicious clients as federal members. To solve this problem, this article proposes a secure member selection strategy (SMSS), which can evaluate the data qualities of members before training. With SMSS, only datasets share more common entities than a certain threshold can be selected for learning, whereas malicious clients with fewer common objects cannot acquire any information about the model. This article implements SMSS, and evaluate its performance via several extensive experiments. Experimental results demonstrate that SMSS is safe, efficient, and effective.