Trade-off between Robustness and Accuracy of Vision Transformers
Abstract
Although deep neural networks (DNNs) have shown great successes in computer vision tasks, they are vulnerable to perturbations on inputs, and there exists a trade-off between the natural accuracy and robustness to such perturbations, which is mainly caused by the existence of robust non-predictive features and non-robust predictive features. Recent empirical analyses find Vision Transformers (ViTs) are inherently robust to various kinds of perturbations, but the aforementioned trade-off still exists for them. In this work, we propose Trade-off between Robustness and Accuracy of Vision Transformers (TORA-ViTs), which aims to efficiently transfer ViT models pretrained on natural tasks for both accuracy and robustness. TORA-ViTs consist of two major components, including a pair of accuracy and robustness adapters to extract predictive and robust features, respectively, and a gated fusion module to adjust the trade-off. The gated fusion module takes outputs of a pre-trained ViT block as queries and outputs of our adapters as keys and values, and tokens from different adapters at different spatial locations are compared with each other to generate attention scores for a balanced mixing of predictive and robust features. Experiments on ImageNet with various robust benchmarks show that our TORA-ViTs can efficiently improve the robustness of naturally pretrained ViTs while maintaining competitive natural accuracy. Our most balanced setting (TORA-ViTs with λ= 0.5) can maintain 83.7% a…
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
There exists an inherent trade-off between a model's ability to generalize to unseen data and its robustness against adversarial perturbations.
“Research from the broader computer vision field suggests that there exists an inherent trade-off between a model's ability to generalize to unseen data and its robustness against adversarial perturbations [219-222].”
From Deepfake detection across image, video, and audio: a comprehensive survey with empirical evaluation of generalization and robustness (2026) · cites this paper as [221]
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474 · see citing works
Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., & Guo, B. (2021) · 2021 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv48922.2021.00986 · Cited by 35,534 · see citing works
Masked Autoencoders Are Scalable Vision Learners
He, K., Chen, X., Xie, S., Li, Y., Dollar, P., & Girshick, R. (2022) · 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr52688.2022.01553 · Cited by 8,489 · cited 2× · see citing works
“MAE [12] proposes a masked autoencoder (MAE) built with ViTs for self-supervised learning, which masks random patches of input images.”
MethodSelf-citationPyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions
Wang, W., Xie, E., Li, X., Fan, D.-P., Song, K., Liang, D., Lu, T., Luo, P., & Shao, L. (2021) · 2021 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv48922.2021.00061 · Cited by 5,259 · see citing works
Han, K., Xiao, A., Wu, E., Guo, J., Xu, C., & Wang, Y. (2021) · arXiv · doi:10.48550/arxiv.2103.00112 · Cited by 1,394 · cited 2× · see citing works
“Off-the-shelf vision transformers [6,11,29,56] are often well trained to pursue the natural accuracy through Eq.”
BackgroundSelf-citationBenchmarking Neural Network Robustness to Common Corruptions and Perturbations
Hendrycks, D., & Dietterich, T. (2019) · arXiv · doi:10.48550/arxiv.1903.12261 · Cited by 1,335 · see citing works
AdapterFusion: Non-Destructive Task Composition for Transfer Learning
Pfeiffer, J., Kamath, A., Rücklé, A., Cho, K., & Gurevych, I. (2021) · Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume · doi:10.18653/v1/2021.eacl-main.39 · Cited by 636 · cited 3× · see citing works
“To support multi-task learning (MTL) in NLP, AdapterFusion [38] uses multiple adapters in parallel in a”
BackgroundMethodInfluentialUnderstanding Robustness of Transformers for Image Classification
Bhojanapalli, S., Chakrabarti, A., Glasner, D., Li, D., Unterthiner, T., & Veit, A. (2021) · 2021 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv48922.2021.01007 · Cited by 440 · cited 2× · see citing works
“Although empirical analyses have demonstrated that ViTs demonstrates robustness against various kinds of perturbations [1, 33, 37], there are only a limited number of works [9, 19, 34, 60] focus on improving the robustness.”
BackgroundIntriguing Properties of Vision Transformers
Naseer, M., Ranasinghe, K., Khan, S., Hayat, M., Khan, F. S., & Yang, M.-H. (2021) · arXiv · doi:10.48550/arxiv.2105.10497 · Cited by 431 · cited 1× · see citing works
“[35] modify DeiT and introduce a shape token to encode shape-information.”
BackgroundVision Transformers Are Robust Learners
Paul, S., & Chen, P.-Y. (2022) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v36i2.20103 · Cited by 213 · cited 2× · see citing works
“Although empirical analyses have demonstrated that ViTs demonstrates robustness against various kinds of perturbations [1, 33, 37], there are only a limited number of works [9,19,34,60] focus on improving the robustness.”
Background
Flagged first, then by citations
Cited by
Sorted by citations
Forgery-Aware Adaptive Learning With Vision Transformer for Generalized Face Forgery Detection
Luo, A., Cai, R., Kong, C., Ju, Y., Kang, X., Huang, J., & Kot, A. C. (2025) · IEEE Transactions on Circuits and Systems for Video Technology · doi:10.1109/tcsvt.2024.3522091 · Cited by 41 · see citing works
Nguyen-Le, H.-H., Tran, V.-T., Nguyen, D.-T., & Le-Khac, N.-A. (2026) · Artificial Intelligence Review · doi:10.1007/s10462-026-11608-4 · Cited by 1 · see citing works
- Authors
- Li, Y., & Xu, C.
- Affiliations
- The University of Sydney (AU)
- Venue
- 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
- Publisher
- IEEE
- Published
- 2023-06
- Pages
- 7558-7568
- ISSN
- 1063-6919 (ISSN-L)
- OpenAlex
- W4386065396
- Cited by
- 23
- References
- 10
- Citations by year
- 2026: 4 · 2025: 4 · 2024: 14 · 2023: 1
- FWCI
- 4.08
- Type
- conference paper
- Language
- en
- Open access
- closed
- Locations
- IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings (publishedVersion)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
- Keywords
- vision transformer, robustness-accuracy trade-off, adversarial robustness, predictive features, adapters, gated fusion module, feature mixing, transfer learning, white-box attacks, ImageNet
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
