Adversarial Robustness Enhancement through Deep Feature Representations-Based CIFAR-10 Image Classification

Abstract

Adversarial perturbations pose a significant challenge to the reliability of deep neural networks in image classification. The conventional feature aggregation approaches are still vulnerable to adversarial noise, which inspires the improvement of more powerful feature presentations. In this study, the adversarial robustness of a ResNet34-based architecture is investigated on the CIFAR-10 dataset, where conventional global average pooling (GAP) is compared with a proposed framework combines Global Average Pooling (GAP) and Fractal Average Pooling (FAP) within a ResNet-34. Multiple perturbation budgets were tested on the models via the white-box fast gradient sign method (FGSM) and projected gradient descent (PGD) attacks. The test accuracy of the proposed model was 97.52% on clean data whereas the test accuracy of the GAP baseline was 95.69%, which is a 1.83% improvement. In the FGSM, the adversarial accuracy increased from 59.56% to 81.38% at e = 1/255 and from 15.81% to 70.81% at e = 8/255. With the PGD, the improvement was not significant; the accuracy increased from 0.06% to 14.99% at e = 1/255 but reached only 0.32% versus 0.03% at  e = 8/255. The distortion characteristics of any of the above methods are highly similar to the perturbation-quality analysis results of the PSNR, SSIM, SNR, and MSE. Overall, the proposed representation enhances the ability to perform clean classification tasks while also being robust to FGSM attacks, while stronger iterative PGD attacks are still a major challenge