ImageNet Pre-training also Transfers Non-Robustness

TL;DR

Pre-training on ImageNet not only improves generalization but also transfers adversarial non-robustness to fine-tuned models, highlighting a trade-off in transfer learning.

Contribution

The study reveals that ImageNet pre-training transfers non-robust features causing non-robustness in downstream models and proposes a simple robust pre-training method.

Findings

ImageNet pre-training transfers adversarial non-robustness.

Non-robust features are learned and transferred.

A simple robust pre-training approach can mitigate non-robustness.

Abstract

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 ImageNet pre-training also transfers adversarial non-robustness from pre-trained model into fine-tuned model in the downstream classification tasks. We first conducted experiments on various datasets and network backbones to uncover the adversarial non-robustness in fine-tuned model. Further analysis was conducted on examining the learned knowledge of fine-tuned model and standard model, and revealed that the reason leading to the non-robustness is the non-robust features transferred from ImageNet pre-trained model. Finally, we analyzed the preference for feature learning of the pre-trained model, explored the factors influencing robustness, and introduced a simple robust ImageNet pre-training solution. Our code is available at \url{https://github.com/jiamingzhang94/ImageNet-Pretraining-transfers-non-robustness}.