Implementation of Fast Gradient Sign Adversarial Attack on Vision Transformer Model and Development of Defense Mechanism in Classification of Dermoscopy Images

In recent years, deep learning has been successfully applied in medical images due to its ability to learn high complex and multidimensional data. However, it is known that deep learning models are vulnerable to adversarial machine learning attacks, which add small imperceptible perturbation to the legitimate input image, causing the models to produce incorrect results. In this study, Fast Gradient Sign Method is applied to Vision Transformer, a basic pre-trained transformer- based model, which is used for binary classification of a publicly available skin lesion dataset and the robustness of the model is analyzed. Then, the adversarial training approach is used to improve the robustness of the model against adversarial attacks. The experimental results show that the classification accuracy is reduced from 90.1% to 27.38% even for a small perturbation, and the adversarial training approach increases the model's robustness with an accuracy value of 96.61%.