Input Integrity and Authentic Results: Towards Trustworthy Aggregation in Federated Learning

Federated learning (FL) is a collaborative machine learning approach that allows multiple clients to train a model jointly while keeping their data local, thus enhancing privacy. A critical component of FL is secure aggregation (SA), which protects user privacy during the server-side aggregation of client updates. However, current state-of-the-art (SOTA) SA schemes, such as LERNA and Flamingo, have significant limitations. They require at least 3 round-trip communications to handle client dropouts, making them inefficient. Additionally, they are vulnerable to malicious attacks from both clients (input data poisoning) and the server (aggregation manipulation). To address these issues, we propose the first 2-round-trip trustworthy aggregation scheme, called HyperSA, which guarantees both input integrity and result authenticity. HyperSA achieves its goal with two key techniques: a consistent authenticated decryption mechanism and a three-step chain proof procedure. The former, implemented through consistent threshold key recovery, reduces the communication round trips from 3 to 2. The latter, comprising three zero-knowledge proofs chained together, enables misbehavior detection, ensuring that all clients receive authentic aggregation results. We implemented HyperSA and conducted a comprehensive evaluation. Both theoretical analysis and experimental results demonstrate the practicality and efficiency of our design compared to existing SOTA schemes.