Janus: Dual-server Multi-Round Secure Aggregation with Verifiability for Federated Learning
Keywords: federated learning, secure aggregation, privacy enhancement
Abstract: Secure Aggregation (SA) in federated learning is essential for preserving user privacy by ensuring that model updates are masked or encrypted and remain inaccessible to servers. Although the advanced protocol Flamingo (S\&P'23) has made significant strides with its multi-round aggregation and optimized communication, it still faces several critical challenges: (i) $\textit{Dynamic User Participation}$, where Flamingo struggles with scalability due to the complex setups required when users join or leave the training process; (ii) $\textit{Model Inconsistency Attacks}$ (MIA), where a malicious server could infer sensitive data, which poses severe privacy risks; and (iii) $\textit{Verifiability}$, as most schemes lack an efficient mechanism for clients to verify the correctness of server-side aggregation, potentially allowing inaccuracies or malicious actions. We introduce Janus, a generic privacy-enhanced multi-round SA scheme through a dual-server architecture. A new user can participate in training by simply obtaining the servers' public keys for aggregation, eliminating the need for complex communication graphs. Our dual-server model separates aggregation tasks, ensuring that neither server can successfully launch a MIA without controlling at least $n-1$ clients. Additionally, we propose a new cryptographic primitive, $\textit{Separable Homomorphic Commitment}$, integrated with our dual-server approach to ensure the verifiability of aggregation results. Extensive experiments across various models and datasets show that Janus significantly boosts security while enhancing efficiency. It reduces per-client communication and computation overhead from logarithmic to constant scale compared to state-of-the-art methods, with almost no compromise in model accuracy.
Primary Area: alignment, fairness, safety, privacy, and societal considerations
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8730
Loading