ACORN: Efficient, Input Validated Secure Aggregation
Baiyu Li
Video
Abstract:
Secure aggregation allows a server to learn the sum of client inputs in a privacy-preserving way, without learning anything about individual client's input except what can be inferred by the sum. It has been used in federated learning and federated aggregation with success. In this talk, I will present a more computation-efficient secure aggregation protocol using Ring Learning With Errors (RLWE) encoding. In addition, to prevent malicious clients from gaining disproportionate influence on the aggregation results, our protocol has been extended with input validation capability. This extension allows the client to prove, in zero knowledge to the server, that its input satisfies constraints such as encoding validity, L_0, L_2, and L_\infty bounds.
Our evaluation shows that the RLWE-based secure aggregation protocol improves the computational efficiency of the state-of-the-art (Bell et al., CCS 2020) by 2-8X in client computation in practical scenarios, while maintaining comparable communication cost. Our input validation extension improves on prior work by more than 30X in terms of client communication while keeping comparable computation costs. Compared to the base protocols without input validation, the extended protocols incur only 0.1X additional communication, and can process binary indicator vectors of length 1M, or 16-bit dense vectors of length 250K, in under 80s of computation per client.
Bio:
Baiyu Li is a research scientist at Google. He works on lattice-based cryptography, especially its use in homomorphic encryption and other cryptographic primitives. Prior to Google, Baiyu completed his PhD in Computer Science at UCSD, advised by Daniele Micciancio.