school assignment 3

[kiw8]

This code is designed to train a ResNet18 model on the CIFAR-10 image classification dataset and evaluate both the model’s general performance (accuracy on clean inputs) and its adversarial robustness (accuracy on adversarial inputs) by applying Projected Gradient Descent (PGD), a type of adversarial attack.

PGD is a representative iterative attack method based on the L∞ norm, which adds small, imperceptible perturbations to input images in order to mislead the model into making incorrect predictions. Through this experiment, we explore how vulnerable deep learning classifiers are to subtle input modifications and highlight the necessity of adversarial robustness training as a countermeasure.