FedADSN: Anomaly detection for social networks under decentralized federated learning

Abstract

With the rapid development of distributed machine learning, federated learning (FL) has attracted plentiful attention and has become a paradigm. Compared with general machine learning, its outstanding advantage is that it can greatly lock the privacy of user data, allowing data to be calculated without leaving the local. We break through the data fortress in social networks with FL. However, it can be enormously difficult to identify a wholly trustworthy server in a social network, which is an obvious deficiency of the FL. Therefore, the decentralized FL is imperative to eliminate the inherent evil of server. In this article, we borrow the core idea of FL and propose an anomaly detection for decentralized federated learning under social network (FedADSN) to detect users with anomalous behaviours. We incorporate Graph Anomaly Detection (GAD) into decentralized FL framework to find malicious attackers in social networks. Our experiments can be performed to calculate and find outliers for malicious users, which demonstrates the feasibility of our framework and claims our algorithms are the state of the art.