Congratulations to Dimitri Lereverend, Davide Frey, and François Taiani for their work “Low-Cost Privacy-Preserving Decentralized Learning” accepted at PoPETS 2025.
The paper explores the use of correlated noise in decentralized learning to achieve an efficient privacy-utility tradeoff. Decentralized learning allows machine learning while keeping individual data local and private, but it remains vulnerable to privacy attacks that can expose sensitive information. Differentially private solutions address this by injecting noise, but excessive noise often degrades model utility. This work reduces the negative impact on utility and convergence speed, both theoretically and empirically, by introducing locally correlated noise. This approach protects user data with significantly lower costs, making decentralized learning more practical and efficient.
This is a joint work with Romaric Gaudel of the LACODAM Inria team, Sayan Biswas, Anne-Marie Kermarrec, Rafael Pires, and Rishi Sharma, members of the SaCS team at EPFL.
A preprint version of the work is available on HAL.
