A Little Robustness Goes a Long Way: Leveraging Robust Features for Targeted Transfer Attacks

Adversarial examples for neural network image classifiers are known to be\ntransferable: examples optimized to be misclassified by a source classifier are\noften misclassified as well by classifiers with different architectures.\nHowever, targeted adversarial examples -- optimized to be classified as a\nchosen target class -- tend to be less transferable between architectures.\nWhile prior research on constructing transferable targeted attacks has focused\non improving the optimization procedure, in this work we examine the role of\nthe source classifier. Here, we show that training the source classifier to be\n"slightly robust" -- that is, robust to small-magnitude adversarial examples --\nsubstantially improves the transferability of class-targeted and\nrepresentation-targeted adversarial attacks, even between architectures as\ndifferent as convolutional neural networks and transformers. The results we\npresent provide insight into the nature of adversarial examples as well as the\nmechanisms underlying so-called "robust" classifiers.\n