FedSecure: A Privacy-Preserving Federated Learning Algorithm

The federated learning (FL) paradigm effectively distributes the training burden among several units, each possessing local datasets. This paper presents a novel privacy-preserving FL algorithm, FedSecure. Our approach is distinct from the state of the art as it redefines the conventional distributed optimization problem inherent in FL. Unlike traditional frameworks that assume a common weight vector as the global decision variable, our method introduces an equivalent constrained problem. Each agent maintains its own weight vector as the local decision variable under the constraint that these local weight vectors must be equal. This enables the dual decomposition method to solve the distributed optimization problem. A key advantage of FedSecure is its ability to eliminate the necessity of sharing both the initial weights and the updated network weight values of each agent with the server. This feature ensures that the information related to agents’ training samples remains impervious to potential state-of-the-art cyber espionage attempts, underscoring the robust security measures of our algorithm. We validate FedSecure on the MNIST and CIFAR-10 datasets and compare it to a differential privacy algorithm, in which artificial noise is added to parameters at the clients’ side before aggregating, namely, noising before model aggregation FL (NbAFL).