Springer, Jacob M.; Melanie Mitchell and Garrett T. Kenyon
Adversarial examples for neural network image classifiers are known to be transferable: examples optimized to be misclassified by a source classifier are often misclassified as well by classifiers with different architectures. However, targeted adversarial examples—optimized to be classified as a chosen target class—tend to be less transferable between architectures. While prior research on constructing transferable targeted attacks has focused on improving the optimization procedure, in this work we examine the role of the source classifier. Here, we show that train- ing the source classifier to be “slightly robust”—that is, robust to small-magnitude adversarial examples—substantially improves the transferability of targeted at- tacks, even between architectures as different as convolutional neural networks and transformers. We argue that this result supports a non-intuitive hypothesis: on the spectrum from non-robust (standard) to highly robust classifiers, those that are only slightly robust exhibit the most universal features—ones that tend to overlap with the features learned by other classifiers trained on the same dataset. The results we present provide insight into the nature of adversarial examples as well as the mechanisms underlying so-called “robust” classifiers.