Fig. 1
From: On the role of deep learning model complexity in adversarial robustness for medical images
Average accuracy and standard deviation of FGSM and PGD attacks on medical image DNNs. Column 1 & 2 are FGSM and PGD attacks, respectively. Row 1, 2, & 3 are Chest X-ray, Dermoscopy & OCT datasets, respectively. All networks exhibit similar performance on unperturbed data for a given dataset. The magnitude of the perturbation was increased by \(\epsilon\) ∈ [0.01,10]. Models of reduced complexity exhibit greater adversarial robustness