DeepDefense: A Steganalysis-Based Backdoor Detecting and Mitigating Protocol in Deep Neural Networks for AI Security

Table 6

Effectiveness performance comparison of defensing protocols under different backdoor attacks.

Dataset	Attack type	FP [15]		ABL [19]		Ours
		ASR %	CA %	ASR %	CA %	ASR %	CA %
CIFAR10	None Attack	0.00	91.88	0.00	92.75	0.00	93.79
	BadNets	99.81	90.37	0.42	93.14	0.21	90.54
	Blend Attack	100.00	93.43	0.48	76.56	0.15	60.43
	SSBA	99.90	93.09	0.50	93.17	0.43	90.81
VGGFACE2 subset	None Attack	0.00	72.62	0.00	82.96	0.00	86.73
	BadNets	11.79	77.26	0.00	14.90	0.32	83.36
	Blend Attack	14.89	71.46	0.00	9.72	0.46	78.67
	SSBA	11.47	72.23	0.00	7.97	0.17	84.27

For different attacks, bold values represents the best defense effect among the three defense schemes.