12
0

Towards Efficient Secure Aggregation in FL: Partial Vector Freezing for Cost Compression

Abstract

Secure aggregation of user vectors has become a critical issue in the field of federated learning. Many Secure Aggregation Protocols (SAP) face exorbitant computation costs, which severely limit their applicability. We uncover that current endeavors to reduce computation costs tend to overlook a crucial fact: a considerable portion of SAP's computation burden stems from processing each entry in the private vectors. Given this observation, we propose PVF, a portable module for compressing computation costs. PVF is able to ``freeze'' a substantial portion of the private vector through specific linear transformations, only requiring 1λ\frac{1}{\lambda} of the original vector to participate in SAP. Eventually, users can ``thaw'' the public sum of the ``frozen entries" by the result of SAP. To enhance functionality, we introduce extensions that can enforce consistency constraints on users' original vectors, verify aggregated results, and enhance security when a portion of the private vector is known to the server. We demonstrate that PVF can seamlessly integrate with various SAP and prove that it poses no threat to user privacy in the semi-honest and active adversary settings. We select 88 baselines, encompassing 66 distinct types of SAP, and explore the acceleration effects of PVF on these SAP. Empirical investigations indicate that when λ=100\lambda=100, PVF yields up to 99.5×99.5\times speedup and up to 32.3×32.3\times communication reduction, with the potential to approach nearly 1000×1000\times acceleration as λ\lambda increases.

View on arXiv
Comments on this paper