The FGSM Attack on Image Classification Models and Distillation as Its Defense

This work introduces the concept of adversarial attacks on image classification models that rely on convolutional neural networks (CNNs). CNNs are commonly used for image classification tasks in deep learning. However, when these networks are exposed to adversarial attacks, even highly effective pre-trained CNN models that perform well on image datasets may exhibit poor performance. This study focuses on exploring one prominent adversarial attack, the fast gradient sign method (FGSM), and its adverse effects on image classification model performance. The FGSM attack is applied to three pre-trained image classifier CNN architectures: DenseNet 201, Inception V3, and RegNetY 400MF, using randomly selected images from the ImageNet dataset. The accuracy of these models is measured both in the absence and presence of the attack to illustrate the negative impact of the attack on their performance. Finally, a defense mechanism based on a distillation approach is presented to counter the attack.