notesum.ai

Published at November 26

BadScan: An Architectural Backdoor Attack on Visual State Space Models

cs.CV

Released Date: November 26, 2024

Authors: Om Suhas Deshmukh¹, Sankalp Nagaonkar¹, Achyut Mani Tripathi¹, Ashish Mishra²

Aff.: ¹IIT Dharwad; ²HPE lab, Bangalore

Arxiv: http://arxiv.org/abs/2411.17283v1

Refer to caption

Attacks	Model	ImageNet-1K			CIFAR-10
Attacks	Model	CTA	TTA	TAR	CTA	TTA	TAR
BadNets [11]	ResNet-18	72.03	46.27	1.56	85.43	46.77	1.83
	ResNet-50	76.00	47.80	1.59	83.97	15.13	5.55
	MLP-mixer	70.40	50.83	1.38	95.17	7.00	13.60
	ViT-S	70.37	46.67	1.51	94.40	50.33	1.88
	VMamba	66.13	49.80	1.33 $\uparrow$ (14.17)	93.40	66.53	1.40 $\uparrow$ (4.87)
WanNet [23]	ResNet-18	72.13	54.07	1.33	84.47	52.23	1.61
	ResNet-50	75.20	53.53	1.40	84.57	17.53	4.82
	MLP-mixer	66.67	38.93	1.71	91.93	71.37	1.29
	ViT-S	73.17	60.20	1.22	94.43	74.00	1.28
	VMamba	67.38	59.20	1.14 $\uparrow$ (14.39)	93.47	86.97	1.07 $\uparrow$ (5.20)
Refool [21]	ResNet-18	72.87	41.37	1.76	87.73	35.23	2.49
	ResNet-50	75.40	48.97	1.54	83.73	8.37	10.01
	MLP-mixer	69.10	38.90	1.78	94.30	32.57	2.90
	ViT-S	72.63	49.47	1.47	96.47	38.20	2.53
	VMamba	65.73	54.80	1.20 $\uparrow$ (14.30)	94.93	50.73	1.87 $\uparrow$ (4.40)
BadScan	VMamba	93.00	6.00	15.50	94.00	15.00	6.27
	MiM	90.00	6.00	15.00	94.00	8.00	11.75
	EF-Mamba	90.00	5.00	18.00	98.00	11.00	8.91