[2510.17480] Unified Privacy Guarantees for Decentralized Learning via Matrix Factorization

Computer Science > Machine Learning arXiv:2510.17480 (cs) [Submitted on 20 Oct 2025 (v1), last revised 27 Feb 2026 (this version, v2)] Title:Unified Privacy Guarantees for Decentralized Learning via Matrix Factorization Authors:Aurélien Bellet, Edwige Cyffers, Davide Frey, Romaric Gaudel, Dimitri Lerévérend, François Taïani View a PDF of the paper titled Unified Privacy Guarantees for Decentralized Learning via Matrix Factorization, by Aur\'elien Bellet and 5 other authors View PDF HTML (experimental) Abstract:Decentralized Learning (DL) enables users to collaboratively train models without sharing raw data by iteratively averaging local updates with neighbors in a network graph. This setting is increasingly popular for its scalability and its ability to keep data local under user control. Strong privacy guarantees in DL are typically achieved through Differential Privacy (DP), with results showing that DL can even amplify privacy by disseminating noise across peer-to-peer communications. Yet in practice, the observed privacy-utility trade-off often appears worse than in centralized training, which may be due to limitations in current DP accounting methods for DL. In this paper, we show that recent advances in centralized DP accounting based on Matrix Factorization (MF) for analyzing temporal noise correlations can also be leveraged in DL. By generalizing existing MF results, we show how to cast both standard DL algorithms and common trust models into a unified formulation...