A Neural Architecture Dataset for Adversarial Robustness

Bowen Zheng, Ran Cheng, Shihua Huang, Zhichao Lu, Vishnu Boddeti

Keywords: Adversarial robustness, neural architecture design

Abstract: Robustness to adversarial attacks is critical for practical deployments of deep neural networks. However, pursuing adversarial robustness from the network architecture perspective demands tremendous computational resources, thereby hampering progress in understanding and designing robust architectures. In this work, we aim to lower this barrier-to-entry for researchers without access to large-scale computation by introducing the first comprehensive neural architecture dataset under adversarial training, dubbed NARes, for adversarial robustness. NARes comprises 15,625 WRN-style unique architectures adversarially trained and evaluated against four adversarial attacks (including AutoAttack). With NARes, researchers can query the adversarial robustness of various models immediately, along with more detailed information, such as fine-grained training statistics, empirical Lipschitz constant, stable accuracy, etc. In addition, four checkpoints are provided for each architecture to facilitate further fine-tuning or analysis. For the first time, the dataset provides a high-resolution architecture landscape for adversarial robustness, enabling quick verifications of theoretical or empirical ideas. Through NARes, we offered some new insight and identified some contradictions in statements of prior studies. We believe NARes can serve as a valuable resource for the community to advance the understanding and design of robust neural architectures.

Primary Area: datasets and benchmarks

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 13215