Group Signature Based Federated Learning Approach for Privacy Preservation

Federated Learning (FL) is a recently developed machine learning technique for updating the learning parameters in distributed devices. Traditionally in FL, the server receives local updates from the devices in the network and then aggregates them to form a new learning model. This new information is again shared with all the network devices. The actual information is not revealed during the repeated learning process since only the updates are sent from local devices. Hence, the privacy of information is better preserved in FL than the conventional centralized machine learning algorithms. However, the existing algorithms proposed for FL require all client devices to communicate with many other clients to share their secrets and preserve privacy. This increases the computation cost and communication overhead exponentially. Therefore, we propose a group signature-based federated learning process requiring members to sign more efficiently using their group's signatures instead of their signatures. We have shown that our algorithm is safe, and the computational and communication cost is significantly less than existing protocols.