CryptoDB
OPA: One-shot Private Aggregation with Single Client Interaction and its Applications to Federated Learning
| Authors: |
|
|---|---|
| Download: | |
| Conference: | CRYPTO 2025 |
| Abstract: | Our work minimizes interaction in secure computation, addressing the high cost of communication rounds, especially with many clients. We introduce One-shot Private Aggregation $\mathsf{OPA}$, enabling clients to communicate only once per aggregation evaluation in a single-server setting. This simplifies dropout management and dynamic participation, contrasting with multi-round protocols like Bonawitz et al. (CCS'17) (and subsequent works) and avoiding complex committee selection akin to YOSO. $\mathsf{OPA}$'s communication behavior \emph{closely} mimics learning-in-the-clear where each client party speaks only once. $\mathsf{OPA}$, built on LWR, LWE, class groups, and DCR, ensures single-round communication for all clients while also achieving sub-linear overhead in the number of clients, making it asymptotically efficient and practical. We achieve malicious security with abort and input validation to defend against poisoning attacks, which are particularly relevant in Federated Learning, where adversaries attempt to manipulate the gradients to degrade model performance or introduce biases. We build two flavors of $\mathsf{OPA}$ (1) from (threshold) key homomorphic PRF and (2) from seed homomorphic PRG and secret sharing. The threshold Key homomorphic PRF addresses shortcomings observed in previous works that relied on DDH and LWR in the work of Boneh~\textit{et al.}(CRYPTO, 2013), marking it as an independent contribution to our work. Our other contributions include new constructions of (threshold) key-homomorphic PRFs and seed-homomorphic PRGs that are secure under the LWE, DCR Assumption, and other Class Groups of Unknown Order. |
BibTeX
@inproceedings{crypto-2025-35822,
title={OPA: One-shot Private Aggregation with Single Client Interaction and its Applications to Federated Learning},
publisher={Springer-Verlag},
author={Harish Karthikeyan and Antigoni Polychroniadou},
year=2025
}