notesum.ai

Published at November 25

Unlocking The Potential of Adaptive Attacks on Diffusion-Based Purification

cs.CR

cs.CV

Released Date: November 25, 2024

Authors: Andre Kassis¹, Urs Hengartner¹, Yaoliang Yu¹

Aff.: ¹Cheriton School of Computer Science, University of Waterloo, Canada

Arxiv: http://arxiv.org/abs/2411.16598v1

Refer to caption

Models	Pur.	Gradient Method	Cl-Acc %	Rob-Acc %
WideResNet-50-2	DiffPure [36]	Adjoint (Nie et al. [36])	71.16	44.39
		DiffAttack (Kang et al. [26])	71.16	31.25
		Full-DiffGrad (Ours)	74.22	12.11
DeiT-S	DiffPure [36]	Adjoint (Nie et al. [36])	73.63	43.18
		DiffAttack (Kang et al. [26])	73.63	32.81
		Adjoint-DiffGrad (Ours)	74.22	19.15
		Full-DiffGrad (Ours)	74.22	21.09
	GDMP [49]	Full-DiffGrad (Ours)	69.14	20.70