Implementation and Comparison of CNN Models on CIFAR-10 Dataset

This study investigates the vulnerability of Convolutional Neural Network (CNN) models to adversarial attacks, focusing on the Fast Gradient Sign Method (FGSM). We implemented and compared CNN models with different depths, including ResNet8, ResNet18, ResNet34, and ResNet50, on the CIFAR-10 dataset. Our results show that while deeper networks achieved higher accuracy on clean data, they exhibited significantly lower robustness under FGSM perturbations. In contrast, shallower models maintained relatively better performance under adversarial conditions despite lower clean accuracy. For example, ResNet34 achieved 82.2% accuracy on clean data but dropped to 2.5% under attack, whereas ResNet8 maintained 25.8% accuracy under attack. These findings highlight a trade-off between model depth and adversarial robustness. The study suggests that improving the robustness of deep learning models against adversarial examples is essential for their safe deployment in real-world applications. Future work should focus on integrating robust defense techniques such as adversarial training, defensive distillation, and regularization strategies to enhance model security.