Skip to main content

Table 1 Model performance at \(\epsilon\) that produce the largest margin between least and most robust networks. Models of reduced complexity exhibit greater performance on perturbed medical images compared to larger, overly complex networks while maintaining comparable performance on unperturbed data

From: On the role of deep learning model complexity in adversarial robustness for medical images

Attack

CBR-LargeT

Resnet-8

Resnet-20

Resnet-32

Resnet-50

(a) Chest X-Ray Accuracy(%), \(\epsilon\) = 1

No Attack

96.43 + − 1.84

97.43 + − 1.01

97.46 + − 1.24

97.41 + − 0.98

96.90 + − 1.26

FGSM

88.83 + − 2.14

24.63 + − 11.12

15.13 + − 10.68

5.77 + − 3.30

9.37 + − 5.82

PGD

88.37 + − 2.31

15.07 + − 9.95

7.10 + − 9.26

0.43 + − 0.62

0.83 + − 1.23

(b) Dermoscopy Accuracy (%), \(\epsilon\) = 0.1

No Attack

96.40 + − 0.95

95.53 + − 1.02

95.03 + − 1.16

95.07 + − 1.48

95.20 + − 0.85

FGSM

92.80 + − 1.48

66.30 + − 12.37

49.37 + − 5.70

53.23 + − 13.13

49.10 + − 7.28

PGD

92.63 + − 1.66

53.30 + − 10.61

38.17 + − 7.74

41.83 + − 15.77

43.63 + − 14.91

(c) OCT Accuracy (%), \(\epsilon\) = 2

No Attack

96.30 + − 0.67

95.53 + − 1.08

95.03 + − 0.81

95.07 + − 1.05

95.20 + − 2.71

FGSM

88.58 + − 2.12

63.85 + − 6.56

63.00 + − 3.92

67.88 + − 3.94

68.65 + − 5.36

PGD

78.35 + − 10.70

55.85 + − 8.60

35.23 + − 9.30

37.98 + − 8.23

36.13 + − 11.56