Toward Secure and Verifiable Hybrid Federated Learning
Runmeng Du, Xuru Li, Daojing He, Kim‐Kwang Raymond Choo
Abstract
Reducing computation cost and ensuring update integrity, are key challenges in federated learning (FL). In this paper, we present a secure and verifiable hybrid FL system for training, namely SVHFL. SVHFL enables training models on both plaintext and encrypted data simultaneously. Furthermore, we propose a mutual verification scheme for the integrity of updates in FL. It is a general and efficient scheme that can eliminate malformed updates from clients and enforce the integrity checks of the aggregation results from the server. The training and verification schemes of SVHFL have reduced the computation cost from a quadratic cost to a linear cost. The experimental results demonstrate the practicality of SVHFL.