Do Adversarially Robust ImageNet Models Transfer Better? Andrew Ilyas

Transfer learning is a widely-used paradigm in which models pre-trained on standard datasets can efficiently adapt to downstream tasks. Typically, better pretrained models yield better transfer results, suggesting that initial accuracy is a key aspect of transfer learning performance. In this work, we identify another such aspect: we find that adversarially robust models, while less accurate, often perform better than their standard-trained counterparts when used for transfer learning. Specifically, we focus on adversarially robust ImageNet classifiers, and show that they yield improved accuracy on a standard suite of downstream classification tasks.