A Security-Friendly Privacy Solution for Federated Learning

Abstract Federated learning is a privacy-aware collaborative machine learning method, but it needs other privacy enhancing technologies to prevent data leakage from local model updates. However, privacy-enhancing technologies may not allow the usage of security mechanisms against some security attacks to model training such as poisoning and backdoor attacks. The reason is that the server that aggregates the local model updates may not be able to analyze them to detect anomalies resulting from these attacks. Solutions that satisfy both privacy and security at the same time are needed for federated learning. Another way could be introducing new privacy solutions that allow the server to execute some analysis on the local model updates without violating privacy. In this paper, we introduce a security-friendly privacy solution for federated learning, which is based on multi-hop communication to hide identities of clients but ensures that the clients in the middle points in the path between clients and the server cannot execute malicious activities such as altering model updates of other clients and sending more than one update