Table 7. BlindMI in models with...

Table 7.

BlindMI in models with different overfitting degrees and defense methods

Training accuracy	Test accuracy	Attack accuracy with dropout	Attack accuracy with member mask	Attack accuracy with non-member mask
0.5234	0.3024	0.7652	0.5009	0.5035
0.7792	0.3255	0.8580	0.5009	0.4905
0.9554	0.3542	0.8641	0.5027	0.5010

Training accuracy	Test accuracy	Attack accuracy with dropout	Attack accuracy with member mask	Attack accuracy with non-member mask
0.5234	0.3024	0.7652	0.5009	0.5035
0.7792	0.3255	0.8580	0.5009	0.4905
0.9554	0.3542	0.8641	0.5027	0.5010

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