Unified Breakdown Analysis for Byzantine Robust Gossip

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Gaucher, Renaud, Dieuleveut, Aymeric, Hendrikx, Hadrien
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915335637041152
author Gaucher, Renaud
Dieuleveut, Aymeric
Hendrikx, Hadrien
author_facet Gaucher, Renaud
Dieuleveut, Aymeric
Hendrikx, Hadrien
contents 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 CS+, such that CS+-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 CS+-RG and highlight the gap with NNA, in particular against a novel attack tailored to decentralized communications.
format Preprint
id arxiv_https___arxiv_org_abs_2410_10418
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Unified Breakdown Analysis for Byzantine Robust Gossip
Gaucher, Renaud
Dieuleveut, Aymeric
Hendrikx, Hadrien
Optimization and Control
Machine Learning
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 CS+, such that CS+-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 CS+-RG and highlight the gap with NNA, in particular against a novel attack tailored to decentralized communications.
title Unified Breakdown Analysis for Byzantine Robust Gossip
topic Optimization and Control
Machine Learning
url https://arxiv.org/abs/2410.10418