{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/rethinking-the-design-principles-of-robust","title":"Towards Robust Vision Transformer","arxiv_id":"2105.07926","date":"2021-05-17","proceeding":"CVPR 2022 1","authors":["Xiaofeng Mao","Gege Qi","Yuefeng Chen","Xiaodan Li","Ranjie Duan","Shaokai Ye","Yuan He","Hui Xue"],"abstract":"Recent advances on Vision Transformer (ViT) and its improved variants have shown that self-attention-based networks surpass traditional Convolutional Neural Networks (CNNs) in most vision tasks. However, existing ViTs focus on the standard accuracy and computation cost, lacking the investigation of the intrinsic influence on model robustness and generalization. In this work, we conduct systematic evaluation on components of ViTs in terms of their impact on robustness to adversarial examples, common corruptions and distribution shifts. We find some components can be harmful to robustness. By using and combining robust components as building blocks of ViTs, we propose Robust Vision Transformer (RVT), which is a new vision transformer and has superior performance with strong robustness. We further propose two new plug-and-play techniques called position-aware attention scaling and patch-wise augmentation to augment our RVT, which we abbreviate as RVT*. The experimental results on ImageNet and six robustness benchmarks show the advanced robustness and generalization ability of RVT compared with previous ViTs and state-of-the-art CNNs. Furthermore, RVT-S* also achieves Top-1 rank on multiple robustness leaderboards including ImageNet-C and ImageNet-Sketch. The code will be available at \\url{https://github.com/alibaba/easyrobust}.","url_abs":"https://arxiv.org/abs/2105.07926v4","url_pdf":"https://arxiv.org/pdf/2105.07926v4.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"rethinking-the-design-principles-of-robust","repo_url":"https://github.com/alibaba/easyrobust","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"pytorch","reach":null},{"paper_slug":"rethinking-the-design-principles-of-robust","repo_url":"https://github.com/vtddggg/Robust-Vision-Transformer","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok"}}],"tasks":[{"task_slug":"domain-generalization","task_name":"Domain Generalization"},{"task_slug":"image-classification","task_name":"Image Classification"},{"task_slug":"robust-design","task_name":"Robust Design"}],"methods":[{"method_slug":"absolute-position-encodings","method_name":"Absolute Position Encodings"},{"method_slug":"adam","method_name":"Adam"},{"method_slug":"attention","method_name":"Attention"},{"method_slug":"bpe","method_name":"BPE"},{"method_slug":"convolution","method_name":"Convolution"},{"method_slug":"dense-connections","method_name":"Dense Connections"},{"method_slug":"dropout","method_name":"Dropout"},{"method_slug":"label-smoothing","method_name":"Label Smoothing"},{"method_slug":"layer-normalization","method_name":"Layer Normalization"},{"method_slug":"linear-layer","method_name":"Linear Layer"},{"method_slug":"multi-head-attention","method_name":"Multi-Head Attention"},{"method_slug":"position-wise-feed-forward-layer","method_name":"Position-Wise Feed-Forward Layer"},{"method_slug":"residual-connection","method_name":"Residual Connection"},{"method_slug":"softmax","method_name":"Softmax"},{"method_slug":"transformer","method_name":"Transformer"},{"method_slug":"vision-transformer","method_name":"Vision Transformer"}],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/domain-generalization-on-imagenet-a","task":"Domain Generalization","dataset":"ImageNet-A","model":"RVT-B*","rank_in_archive_order":28,"of":39,"metrics":{"Top-1 accuracy %":"28.5"},"uses_additional_data":false},{"leaderboard":"/sota/domain-generalization-on-imagenet-a","task":"Domain Generalization","dataset":"ImageNet-A","model":"RVT-S*","rank_in_archive_order":29,"of":39,"metrics":{"Top-1 accuracy %":"25.7"},"uses_additional_data":false},{"leaderboard":"/sota/domain-generalization-on-imagenet-a","task":"Domain Generalization","dataset":"ImageNet-A","model":"RVT-Ti*","rank_in_archive_order":30,"of":39,"metrics":{"Top-1 accuracy %":"14.4"},"uses_additional_data":false},{"leaderboard":"/sota/domain-generalization-on-imagenet-c","task":"Domain Generalization","dataset":"ImageNet-C","model":"RVT-B*","rank_in_archive_order":24,"of":47,"metrics":{"mean Corruption Error (mCE)":"46.8"},"uses_additional_data":false},{"leaderboard":"/sota/domain-generalization-on-imagenet-c","task":"Domain Generalization","dataset":"ImageNet-C","model":"RVT-S*","rank_in_archive_order":26,"of":47,"metrics":{"mean Corruption Error (mCE)":"49.4"},"uses_additional_data":false},{"leaderboard":"/sota/domain-generalization-on-imagenet-c","task":"Domain Generalization","dataset":"ImageNet-C","model":"RVT-Ti*","rank_in_archive_order":33,"of":47,"metrics":{"mean Corruption Error (mCE)":"57.0"},"uses_additional_data":false},{"leaderboard":"/sota/domain-generalization-on-imagenet-r","task":"Domain Generalization","dataset":"ImageNet-R","model":"RVT-B*","rank_in_archive_order":26,"of":39,"metrics":{"Top-1 Error Rate":"51.3"},"uses_additional_data":false},{"leaderboard":"/sota/domain-generalization-on-imagenet-r","task":"Domain Generalization","dataset":"ImageNet-R","model":"RVT-S*","rank_in_archive_order":28,"of":39,"metrics":{"Top-1 Error Rate":"52.3"},"uses_additional_data":false},{"leaderboard":"/sota/domain-generalization-on-imagenet-r","task":"Domain Generalization","dataset":"ImageNet-R","model":"RVT-Ti*","rank_in_archive_order":31,"of":39,"metrics":{"Top-1 Error Rate":"56.1"},"uses_additional_data":false},{"leaderboard":"/sota/image-classification-on-imagenet","task":"Image Classification","dataset":"ImageNet","model":"RVT-B*","rank_in_archive_order":516,"of":1060,"metrics":{"GFLOPs":"17.7","Number of params":"91.8M","Top 1 Accuracy":"82.7%"},"uses_additional_data":false},{"leaderboard":"/sota/image-classification-on-imagenet","task":"Image Classification","dataset":"ImageNet","model":"RVT-S*","rank_in_archive_order":597,"of":1060,"metrics":{"GFLOPs":"4.7","Number of params":"23.3M","Top 1 Accuracy":"81.9%"},"uses_additional_data":false},{"leaderboard":"/sota/image-classification-on-imagenet","task":"Image Classification","dataset":"ImageNet","model":"RVT-Ti*","rank_in_archive_order":770,"of":1060,"metrics":{"GFLOPs":"1.3","Number of params":"10.9M","Top 1 Accuracy":"79.2%"},"uses_additional_data":false}],"syntology":{"syntology_url":"https://syntology.ai/paper/2105.07926","atlas_url":"https://app.syntology.ai/?focus=2105.07926","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}