Privacy-Enhanced Federated Learning with Weighted Aggregation

Abstract

The pervasive adoption of machine learning (ML) techniques by social network operators has led to a growing concern in the personal data privacy of their customers. ML inevitably accesses and processes users’ personal data, which could potentially breach the relevant privacy protection regulations if not performed carefully. In this backdrop, Federated Learning (FL) is an emerging area that allows ML on distributed data without the data leaving their stored location. Typically, FL starts with an initial global model, with each datastore uses its local data to compute the gradient based on the global model, and uploads their gradients (instead of the data) to an aggregation server, at which the global model is updated and then distributed to the local datastores iteratively. However, depending on the nature of the services operated by social networks, data captured at different locations may carry different significance to the business operation, hence a weighted aggregation will be highly desirable for enhancing the quality of the FL model. Furthermore, to prevent the data leakage from aggregated gradients, cryptographic mechanisms are needed to allow secure aggregation of FL. As such, this paper proposes a privacy-enhanced FL scheme, based on cryptographic mechanisms that allow both the data significance evaluation and weighted aggregation of local models in a privacy-preserving manner. Experimental results show that our scheme is practical and secure.

J. Guo and Z. Liu—Both authors contributed equally to this research.