Investigo. Repositorio Institucional de la Universidade de Vigo
Publicación
Artículo científico (JournalArticle)
Privacy-aware Berrut Approximated Coded Computing for Federated Learning
Federated Learning (FL) is a machine learning framework that enhances privacy compared to normal FL, since model training is done across multiple nodes by exchanging only local parameters instead of raw data. Despite this advantage, FL is still vulnerable to some privacy attacks that have been addressed with techniques such as Differential Privacy (DP), Homomorphic Encryption (HE), or Secure Multi-Party Computation (SMPC). These methods need further assumptions that narrow their range of application, exhibiting problems to work with non-linear functions, to perform large matrix multiplications and having high communication and computational costs to manage semi-honest nodes. In this paper, we propose a solution to guarantee privacy in FL schemes that simultaneously solves these shortcomings. Our proposal is based on the Berrut Approximated Coded Computing (BACC), a technique from the coded computing paradigm, adapted here to provide input privacy to FL in a scalable way. In our Private BACC, controlled randomness is introduced into the rational Berrut approximation of a target function so that the input values are hidden and the output function values can still be computed with good approximation. We provide a theoretical bound to the privacy leakage level of private BACC, and further describe its application for computing arbitrary non-linear functions, particularly distributed matrix multiplications. Our numerical results demonstrate that adding privacy in this form outperforms DP in model quality, since it has negligible cost in the precision. Additionally, it beats HE in terms of computational cost, and requires lower computation and communications costs than with SMPC algorithms. Finally, the proposed method supports distributed matrix multiplications, so it can be integrated into a wide class of machine learning systems., Agencia Estatal de Investigación | Ref. PID2020-113795RB-C33, Agencia Estatal de Investigación | Ref. PID2023-148716OB-C31, Instituto Nacional de Ciberseguridad (INCIBE), Xunta de Galicia | Ref. ED431B 2024/41, Universidade de Vigo/CISUG
Proyecto:
AEI / Plan Estatal de Investigación Científica y Técnica y de Innovación 2017-2020 PID2020-113795RB-C33
,
AEI / Plan Estatal de Investigación Científica y Técnica y de Innovación 2021-2023 PID2023-148716OB-C31
,
EC / HE 101070038
DOI: http://hdl.handle.net/11093/11638, https://linkinghub.elsevier.com/retrieve/pii/S1084804525001778
HANDLE: http://hdl.handle.net/11093/11638, https://linkinghub.elsevier.com/retrieve/pii/S1084804525001778