Enhancing Privacy and Utility in Federated Learning: A Hybrid P2P and Server-Based Approach with Differential Privacy Protection

Federated Learning has been recently adopted in several contexts as a solution to train a Machine Learning model while preserving users’ privacy. Even though it avoids data sharing among the users involved in the training, it is common to use it in conjunction with a privacy-preserving technique like DP due to potential privacy issues. Unfortunately, often the application of privacy protection strategies leads to a degradation of the model’s performance. Therefore, in this paper, we propose a framework that allows the training of a collective model through Federated Learning using a hybrid architecture that enables clients to mix within the same learning process collaborations with (semi-)trusted entities and collaboration with untrusted participants. To reach this goal we design and develop a process that exploits both the classic Client-Server and the Peerto-Peer training mechanism. To evaluate how our methodology could impact the model utility we present an experimental analysis using three popular datasets. Experimental results demonstrate the effectiveness of our approach in reducing, in some cases, up to 32% the model accuracy degradation caused by the use of DP.