Communication-efficient Verifiable and Oblivious Aggregation with Client Dropouts

Federated learning (FL) allows each client to train data locally and share only model parameters with an aggregation server. A critical component of FL is secure aggregation (SA), which protects user privacy during the server-side aggregation of client model parameters. However, SA-based FL still faces several challenges from a malicious server, affecting both performance and security. To address these issues, we propose the first 2-round-trip verifiable and oblivious aggregation protocol. Specifically, our protocol incorporates: (1) a consistent temporary key disclosure mechanism for round-trip-efficient communication with client dropout resilience; (2) a message authentication code to verify aggregation results; and (3) a multi-round oblivious aggregation scheme that conceals both individual and global model parameters. These three key techniques collectively enable efficient communication while preventing a malicious server from tampering with aggregation results and accessing any individual or global parameters. We conducted a comprehensive evaluation to demonstrate the practicality and efficiency of our design compared to existing schemes.