notesum.ai

Published at October 30

Byzantine-Robust Federated Learning: An Overview With Focus on Developing Sybil-based Attacks to Backdoor Augmented Secure Aggregation Protocols

cs.LG

cs.CR

Released Date: October 30, 2024

Authors: Atharv Deshmukh¹

Aff.: ¹Not specified

Arxiv: http://arxiv.org/abs/2410.22680v1

Refer to caption

Existing Robust FL Frameworks
Solution	Category	Target Attack	Max Attackers	Model Accuracy	Data Distribution	Time Complexity
Multi-Krum [6]	Distance based	Data/model poisoning	50%	Medium	IID	$O(K^{2}d)$
FABA [9]	Distance based	Data/model poisoning	50%	Medium	IID	$O(K^{2}d)$
Sniper [10]	Distance based	Data poisoning	50%	Medium	IID	$O(K^{2}d)$
FoolsGold [11]	Distance based	Data/model poisoning	No limit	Medium High (with Multi-Krum)	IID/non-IID	$O(K^{2}d)$
Wan et al. [12]	Distance based	Model Poisoning	50%	High	IID	$O(K^{2}d)$
MAB-RFL [13]	Distance based	Data/model poisoning	50%	High	IID/non-IID	$O(K^{2}d)$
Li et al. [14]	Performance based	Data/model poisoning	50%	High	IID/non-IID	$O(Kd)$
Zeno [15]	Performance based	Data/model poisoning	No limit	High	IID/non-IID	$O(Kd)$
Cao et al. [16]	Performance based	Data/model poisoning	No limit	High	IID	$O(Kd)$
FLTrust [17]	Performance based	Data/model poisoning	No limit	High	IID/non-IID	$O(Kd)$
AFA [18]	Statistic based	Data/model poisoning	50%	High	IID	$O(K^{2}d)$
GeoMed MarMed Trimmedmean [19, 20]	Statistic based	Data/model poisoning	50%	High	IID/non-IID	$O(KdlogK)$
Bulyan [21]	Statistic based	Data/model poisoning	50%	Medium	IID	$O(K^{2}d)$