Table 1 Verification accuracy, GAR @ FAR = \(10^{-2}\), and GAR @ FAR = \(10^{-3}\)at test time when employing the CB framework with random matrices \({\textbf {W}}_i \in \textbf{R}^{r \times d}\)and ISTA-Net reconstruction with \(np = 3\), where the feature vector dimensionality is \(d=512\), and np denotes the number of ISTA-Net phases [9]. Results are computed for projection size \(r \in {64, 128, 256}\), corresponding to a compression ratio equal to 1/8, 1/4, and 1/2 of the size of the embeddings (\(d = 512\)), and compared against the unsecured baseline (NOCB)

r = 64

Verification Acc.

99.28

97.08

96.17

94.18

97.73

94.45

GAR@FAR=\(10^{-2}\)

99.53

93.97

92.00

83.25

94.83

82.20

GAR@FAR=\(10^{-3}\)

92.83

82.77

79.03

61.12

85.07

63.53

r = 128

Verification Acc.

99.62

97.78

96.67

95.33

99.13

95.57

GAR@FAR=\(10^{-2}\)

99.80

96.17

93.43

86.70

99.17

87.40

GAR@FAR=\(10^{-3}\)

98.37

90.53

86.77

65.87

96.83

73.67

r = 256

Verification Acc.

99.72

97.85

96.70

95.67

99.18

95.82

GAR@FAR=\(10^{-2}\)

99.83

96.47

93.63

87.63

99.33

88.20

GAR@FAR=\(10^{-3}\)

98.50

91.43

87.57

66.73

97.07

74.07

NOCB

Verification Acc.

99.82

98.16

97.10

95.85

99.43

96.24

GAR@FAR=\(10^{-2}\)

99.80

96.93

94.63

88.53

99.43

89.23

GAR@FAR=\(10^{-3}\)

99.20

92.20

89.50

68.27

97.57

74.70