Fig. 4
From: On the role of deep learning model complexity in adversarial robustness for medical images
Saliency maps of adversarially trained (AT) Resnet8 (a, c, e) and Resnet50 (b, d, f) for Chest X-ray, Dermoscopy & OCT datasets. Unperturbed images with predicted labels and corresponding saliency maps were visualized on row 1 (no attack) for (a–f), respectively. Imperceptibly perturbed images with predicted labels and corresponding saliency maps were visualized on row 2 (PGD attack) for (a–f), respectively. Saliency maps of Resnet50 are more concentrated on the regions of interest whereas Resnet8 include attention regions that are more sporadic on areas that do not contribute to the classification of the disease