A Socially-Aware, Privacy-Preserving, and Scalable Federated Learning Protocol for Distributed Online Social Networks

Abstract

Online Social Networks (OSNs) have been gaining tremendous growth by attracting billions of users from all over the world. Such massive growth leads to scalability and data privacy concerns. Decentralized solutions still are not able to solve privacy and scalability problems efficiently. Hence, recently fully Distributed Online Social Networks (DOSNs) have been proposed. However, despite solving scalability and privacy issues, fully DOSNs impose difficulties in executing data mining and machine learning services which are vital for social networks. In the fully DOSN, each user has only one feature vector and these vectors cannot move to any central storage or other users in a raw form due to privacy issues. In addition, users can directly communicate only with their immediate neighbours/friends in a social network/graph. To cope with these problems, we propose a novel Federated learning algorithm for DOSNs based on the Gossip protocol. We propose a two layer protocol in which the underlying layer is a socially-aware gossip sampling protocol and the upper layer is a push-based merging gossip protocol. The former is responsible for creating a socially-aware random overlay network while the latter, utilizing the sampling protocol, does the training model. We implement our algorithm and through extensive experiments show that the algorithm trains the model up to 88% accuracy compared to the centralized approach.