Analysis of Adversarial Attacks on Skin Cancer Recognition

Abstract

Cancer is one of the most detrimental diseases having the highest death rates in recent years. There are various types of cancers; among them, skin cancer is the most familiar one. Early detection and treatment of it can lead to full recovery for the patients. Deep learning based image classification models have been proven to perform exclusively well to classify images. This is also true for images in the domain of medical image analysis. However, in recent years researchers have shown that adding small calculated noises which are not noticeable by human eyes can induce these models to generate wrong answers. These adversarial examples have proven to be harmful with regards to security measures and hamper launching of deep learning based models in to the real world. In this regard, here we performed experiments to improve the robustness of deep learning based models from these type of attacks for skin cancer recognition. We performed adversarial training based on Projected Gradient Descent (PGD) to increase the robustness of two popular deep learning models, namely MobileNet and VGG16, against white-box attacks of PGD and FGSM attacks. We performed our experiments on a dataset of 10015 images and have shown that our models are much robust than standard training ones and achieved almost similar results as them.