A Curious Case of Remarkable Resilience to Gradient Attacks via Fully Convolutional and Differentiable Front End with a Skip Connection

We experimented with front-end enhanced neural models where a differentiable and fully convolutional model with a skip connection is added before a frozen backbone classifier. By training such composite models using a small learning rate for about one epoch, we obtained models that retained the accuracy of the backbone classifier while being unusually resistant to gradient attacks-including APGD and FAB-T attacks from the AutoAttack package-which we attribute to gradient masking. Although gradient masking is not new, the degree we observe is striking for fully differentiable models without obvious gradient-shattering-e.g., JPEG compression-or gradient-diminishing components. The training recipe to produce such models is also remarkably stable and reproducible: We applied it to three datasets (CIFAR10, CIFAR100, and ImageNet) and several modern architectures (including vision Transformers) without a single failure case. While black-box attacks such as the SQUARE attack and zero-order PGD can partially overcome gradient masking, these attacks are easily defeated by simple randomized ensembles. We estimate that these ensembles achieve near-SOTA AutoAttack accuracy on CIFAR10, CIFAR100, and ImageNet (while retaining almost all clean accuracy of the original classifiers) despite having near-zero accuracy under adaptive attacks. Adversarially training the backbone further amplifies this front-end "robustness". On CIFAR10, the respective randomized ensemble achieved 90.8$\pm 2.5\%$ (99\% CI) accuracy under the full AutoAttack while having only 18.2$\pm 3.6\%$ accuracy under the adaptive attack ($\varepsilon=8/255$, $L^\infty$ norm). We conclude the paper with a discussion of whether randomized ensembling can serve as a practical defense. Code and instructions to reproduce key results are available. https://github.com/searchivarius/curious_case_of_gradient_masking