Table 5. Pre- and post-defense accuracy...

Table 5.

Pre- and post-defense accuracy of attack model on Caltech-UCSD Birds 200

Attack method	Metric	No. defended	Non-member mask	Member mask
Ml-leaks	Accuracy	0.9838	0.5000	0.5950
	Precision	0.9691	0.0000	0.5566
	Recall	1.0000	0.0000	1.0000
	F1-score	0.9843	–	0.7152
BlindMI	Accuracy	0.9801	0.0101	0.4983
	Precision	0.9623	0.0000	0.5143
	Recall	1.0000	0.0000	0.2397
	F1-score	0.9808	–	0.3270

Attack method	Metric	No. defended	Non-member mask	Member mask
Ml-leaks	Accuracy	0.9838	0.5000	0.5950
	Precision	0.9691	0.0000	0.5566
	Recall	1.0000	0.0000	1.0000
	F1-score	0.9843	–	0.7152
BlindMI	Accuracy	0.9801	0.0101	0.4983
	Precision	0.9623	0.0000	0.5143
	Recall	1.0000	0.0000	0.2397
	F1-score	0.9808	–	0.3270

Source: Authors’ own data, using the Ml-leaks described by Salem et al. and BlindMI described by Hui et al.