MPC-Based Privacy-Preserving Serverless Federated Learning

Federated learning (FL) enables multiple users to collaboratively train a global model by keeping their data sets local. Since a single server is used, traditional FL faces the single point of failure problem. An available approach is to adopt a serverless architecture. However, most of existing serverless FL schemes fail to protect gradient privacy, except for a few schemes that adopt differential privacy (DP) where the global model accuracy will decrease. To address these problems, we propose a privacy-preserving serverless FL scheme based on secure multiparty computation (MPC). Combining multiple cryptographic primitives (e.g., key agreement and symmetric encryption), our scheme protects gradient privacy in FL and it is accuracy-lossless. By secret sharing, our scheme supports users to quit an FL task in each round during training.