ImageNet Pre-training Also Transfers Non-robustness

ImageNet pre-training has enabled state-of-the-art results on many tasks. In spite of its recognized contribution to generalization, we observed in this study that also transfers adversarial non-robustness from pre-trained model into fine-tuned the downstream classification We first conducted experiments various datasets and network backbones uncover model. Further analysis was examining learned knowledge standard model, revealed reason leading is non-robust features transferred Finally, analyzed preference for feature learning explored factors influencing robustness, introduced a simple robust solution. Our code available at https://github.com/jiamingzhang94/ImageNet-Pretraining-transfers-non-robustness.