A Rotation and a Translation Suffice: Fooling CNNs with Simple Transformations
share
Recent work has shown that neural network-based vision classifiers exhibit a significant vulnerability to misclassifications caused by imperceptible but adversarial perturbations of their inputs. These perturbations, however, are purely pixel-wise and built out of loss function gradients of either the attacked model or its surrogate. As a result, they tend to be contrived and look pretty artificial. This might suggest that such vulnerability to slight input perturbations can only arise in a truly adversarial setting and thus is unlikely to be an issue in more "natural" contexts. In this paper, we provide evidence that such belief might be incorrect. We demonstrate that significantly simpler, and more likely to occur naturally, transformations of the input - namely, rotations and translations alone, suffice to significantly degrade the classification performance of neural network-based vision models across a spectrum of datasets. This remains to be the case even when these models are trained using appropriate data augmentation. Finding such "fooling" transformations does not require having any special access to the model - just trying out a small number of random rotation and translation combinations already has a significant effect. These findings suggest that our current neural network-based vision models might not be as reliable as we tend to assume. Finally, we consider a new class of perturbations that combines rotations and translations with the standard pixel-wise attacks. We observe that these two types of input transformations are, in a sense, orthogonal to each other. Their effect on the performance of the model seems to be additive, while robustness to one type does not seem to affect the robustness to the other type. This suggests that this combined class of transformations is a more complete notion of similarity in the context of adversarial robustness of vision models.
READ FULL TEXT