Test Time Augmentations are Worth One Million Images for Out-of-Distribution Detection

Zhiyuan He, Yijun YANG, Pin-Yu Chen, Qiang Xu, Tsung-Yi Ho

Keywords: Out-of-distribution, Test time augmentation, OOD Detection

Abstract: Out-of-distribution (OOD) detection is a major threat for deploying machine learning models in safety-critical scenarios. Data augmentations have been proven to be beneficial to OOD detection by providing diverse features. However, previous methods have only focused on the role of data augmentation in the training phase, overlooking its impact on the testing phase. In this paper, we present the first comprehensive study of the impact of test-time augmentation (TTA) on OOD detection. We find aggressive TTAs can cause distribution shifts on OOD scores of In-distribution (InD) data, whereas mild TTAs do not, resulting in the effectiveness of mild TTAs on OOD Detection. Based on the above observations, we propose a detection method that performs a K-nearest-neighbor (KNN) search on mild TTAs instead of InD data. With only 25 TTAs, our method outperforms state-of-the-art methods using the entire training set (1.2 million images) on IMAGENET for OOD detection. Moreover, our approach is compatible with various model architectures and robust to adversarial examples.

Primary Area: societal considerations including fairness, safety, privacy

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/2024/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: 7025