SeaFlame: Communication-Efficient Secure Aggregation for Federated Learning against Malicious Entities

Abstract

Secure aggregation is a popular solution to ensuring privacy for federated learning. However, when considering malicious participants in secure aggregation, it is difficult to achieve both privacy and high efficiency. Therefore, we propose SeaFlame, a communication-efficient secure aggregation protocol against malicious participants. Inspired by the state-of-the-art work, ELSA, SeaFlame also utilizes two non-colluding servers to ensure privacy against malicious entities and provide defenses against boosted gradients. Crucially, to improve communication efficiency, SeaFlame uses arithmetic sharing together with arithmetic-to-arithmetic share conversion to reduce client communication, and uses the random linear combination to reduce server communication.
Security analysis proves that our SeaFlame guarantees privacy against malicious clients colluding with one malicious server. Experimental evaluation demonstrates that, compared to ELSA, SeaFlame optimizes communication by up to 10.5, 6.00, and 8.17 times for a client, a server, and all entities, at the expense of 1.25-1.86 times additional end-to-end runtime.