notesum.ai

Published at November 11

Federated Learning Client Pruning for Noisy Labels

cs.LG

cs.AI

cs.CV

cs.DC

Released Date: November 11, 2024

Authors: Mahdi Morafah¹, Hojin Chang¹, Chen Chen², Bill Lin¹

Aff.: ¹University of California San Diego, USA; ²University of Central Florida, USA

Arxiv: http://arxiv.org/abs/2411.07391v1

Refer to caption

	Partitioning	IID				NIID ( $\alpha=0.5$ )
Baseline	Dataset	CIFAR10		CIFAR100		CIFAR10		CIFAR100
	Noise Rate	$\mu=0.5$	$\mu=0.8$	$\mu=0.5$	$\mu=0.8$	$\mu=0.5$	$\mu=0.8$	$\mu=0.5$	$\mu=0.8$
FedAvg [30]	Vanilla	$55.64\%$	$42.18\%$	$42.06\%$	$31.71\%$	$43.37\%$	$34.59\%$	$40.16\%$	$31.35\%$
	+ClipFL	$\bm{69.42\%}$	$\bm{59.14\%}$	$\bm{43.20\%}$	$\bm{41.21}\%$	$\bm{49.56\%}$	$\bm{48.34\%}$	$\bm{42.74\%}$	$\bm{37.02\%}$
	$\Delta$	$+13.78\%$	$+16.96\%$	$+1.14\%$	$+9.50\%$	$+6.19\%$	$+13.75\%$	$+2.25\%$	$+5.67\%$
FedProx [24]	Vanilla	$58.64\%$	$45.11\%$	$\bm{41.16\%}$	$29.42\%$	$46.25\%$	$33.50\%$	$38.48\%$	$29.37\%$
	+ClipFL	$\bm{68.99\%}$	$\bm{68.92\%}$	$39.86\%$	$\bm{45.13\%}$	$\bm{54.85\%}$	$\bm{53.93\%}$	$\bm{40.81\%}$	$\bm{40.16\%}$
	$\Delta$	$+10.35\%$	$+23.81\%$	$-1.30\%$	$+15.71\%$	$+8.60\%$	$+20.43\%$	$+2.33\%$	$+10.79\%$
FedNova [50]	Vanilla	$56.55\%$	$40.62\%$	$43.67\%$	$32.11\%$	$46.86\%$	$34.25\%$	$39.85\%$	$29.56\%$
	+ClipFL	$\bm{67.48\%}$	$\bm{62.00\%}$	$\bm{47.23\%}$	$\bm{39.71\%}$	$\bm{53.03\%}$	$\bm{49.82\%}$	$\bm{43.69\%}$	$\bm{34.72\%}$
	$\Delta$	$+10.93\%$	$+21.38\%$	$+3.56\%$	$+7.60\%$	$+6.17\%$	$+15.57\%$	$+3.84\%$	$+5.16\%$
FedAdam [39]	Vanilla	$53.97\%$	$42.35\%$	$\bm{38.08\%}$	$22.51\%$	$\bm{41.39\%}$	$30.88\%$	$32.33\%$	$23.50\%$
	+ClipFL	$\bm{68.99\%}$	$\bm{61.16\%}$	$35.35\%$	$\bm{34.14\%}$	$40.39\%$	$\bm{49.84\%}$	$\bm{36.27\%}$	$\bm{34.36\%}$
	$\Delta$	$+15.02\%$	$+18.81\%$	$-2.73\%$	$+11.63\%$	$-1.00\%$	$+18.96\%$	$+3.94\%$	$+10.86\%$