Clinique des données
17 mars 2025 @ 14 h 00 min – CBGP, Grande salle de réunion du CBGP (Présentiel) La Clinique des données est un service mis en place par l’ISDM afin de porter assistance à la communauté scientifique sur des thématique liées aux données. Ainsi, toute personne ayant une problématique, une question, un bug est la bienvenue lors des permanences de ce service. Vous serez […]
Unified Breakdown Analysis for Byzantine Robust Gossip
16 mai 2025 @ 11 h 00 min – 12 h 00 min –
Inria Montpellier, St-Priest Campus, Building 5, Room 02/022
Machine Learning in Montpellier, Theory & Practice
In decentralized machine learning, different devices communicate in a peer-to-peer manner to collaboratively learn from each other’s data. Such approaches are vulnerable to misbehaving (or Byzantine) devices. We introduce F-RG, a general framework for building robust decentralized algorithms with guarantees arising from robust-sum-like aggregation rules F. We then investigate the notion of breakdown point, and show an upper bound on the number of adversaries that decentralized algorithms can tolerate. We introduce a practical robust aggregation rule, coined CSours, such that CSours-RG has a near-optimal breakdown. Other choices of aggregation rules lead to existing algorithms such as ClippedGossip or NNA. We give experimental evidence to validate the effectiveness of CSours-RG and highlight the gap with NNA, in particular against a novel attack tailored to decentralized communications.
MachineLearning, LabéliséHallesIA, IA&Expert
Unified Breakdown Analysis for Byzantine Robust Gossip
16 mai 2025 @ 14 h 00 min –
Room 02.249, Building 5, St Priest campus
Machine Learning in Montpellier, Theory & Practice – Hadrien Hendrikx (Inria)
In decentralized machine learning, different devices communicate in a peer-to-peer manner to collaboratively learn from each other's data. Such approaches are vulnerable to misbehaving (or Byzantine) devices. We introduce F-RG, a general framework for building robust decentralized algorithms with guarantees arising from robust-sum-like aggregation rules F. We then investigate the notion of breakdown point, and show an upper bound on the number of adversaries that decentralized algorithms can tolerate. We introduce a practical robust aggregation rule, coined CSours, such that CSours-RG has a near-optimal breakdown. Other choices of aggregation rules lead to existing algorithms such as ClippedGossip or NNA. We give experimental evidence to validate the effectiveness of CSours-RG and highlight the gap with NNA, in particular against a novel attack tailored to decentralized communications
Machine Learning in Montpellier, Theory & Practice