Fast Secure Aggregation With High Dropout Resilience for Federated Learning
Shisong Yang, Yuwen Chen, Zhen Yang, Bowen Li, Huan Liu
Abstract
Federated learning has been a paradigm for privacy-preserving machine learning, but recently gradient leakage attacks threaten privacy in federated learning. Secure aggregation for federated learning is an approach to protect users’ privacy from these attacks. Especially in federated learning with large-scale mobile devices, e.g., smartphones, secure aggregation should be dropout-resilient and have low communication overhead in addition to protecting privacy. However, existing studies still suffer from performance degradation and security risk, since the number of the dropped users increases. To address these problems, we propose an effective and high dropout-resilience secure aggregation protocol based on homomorphic Pseudorandom Generator and Paillier, which can guarantee privacy while tolerating up to almost 50% of users dropping out in both the honest but curious and actively malicious settings, and the performance of aggregation in computation and communication is independent to the dropped users. To further improve performance, we reduce the number of the ciphertexts through a homomorphic Pseudorandom Generator in the multiplicative group of integers, and decrease the running time of the server-side aggregation by computing discrete logarithms fast in Paillier. Experimental evaluation shows that the proposed protocol reduces the communication overhead by <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$3\times $ </tex-math></inline-formula> while achieving <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$2\times $ </tex-math></inline-formula> computation speedup over the prior dropout-resilience scheme.