Dataset

AG News

Threshold Values

Threshold = 0.0001

Threshold = 0.11

Threshold = −0.08

Metrics

Accuracy

Precision

Recall

Accuracy

Precision

Recall

Accuracy

Precision

Recall

Attacks

Models

Self-calibrated Probabilistic Variation MIA

Mistral 7B

74.40

66.14

100.00

50.00

50.00

100.00

86.80

100.00

73.60

LLaMA 7B

93.70

88.81

100.00

50.00

50.00

100.00

81.70

100.00

63.40

LLaMA 13B

89.20

82.24

100.00

50.00

50.00

100.00

95.60

100.00

91.2

Mixtral 8x7B

91.60

85.86

99.60

50.10

50.05

100.00

57.30

100.00

14.60