DOG: Diffusion-based Outlier Generation for Out-of-Distribution Detection

Hao Jiang, Zhen Fang, Xue Jiang, Zhun Zhong, Tongliang Liu, Bo Han

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.

Keywords: OOD Detection, Reliable Machine Learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: \textit{Out-of-distribution} (OOD) detection is essential for neural networks to ensure reliable predictions in open-world machine learning. Previous approaches have shown that suitable surrogate outlier data are helpful in training OOD detection models. However, obtaining appropriate surrogate outliers presents several substantial challenges, including difficulties in collecting surrogate datasets and confusion of selecting the appropriate outlier data. In this paper, we propose a novel framework called \textit{Diffusion-Based Outlier Generation} (\texttt{DOG}), which synthesizes surrogate outlier data using a large-scale pre-trained diffusion model. \texttt{DOG} generates surrogate outliers using only the \textit{in-distribution} (ID) data, which are subsequently used to further fine-tune the OOD detection model. Compared to previous methods that only use visual or text category information to synthesize outliers, our implementation combines both of them to generate outliers for downstream fine-tuning tasks. Specifically, our method reconstructs images with a diffusion model conditioned on the text category, which utilizes the implicit semantic information contained in the visual images, along with explicit textual category information, to synthesize surrogate outliers. In addition, our \texttt{DOG} presents a novel approach for outlier exposure by allowing dynamic adjustment of surrogate outlier data based on the results, leading to an enhancement in OOD detection performance. Extensive experiments across various OOD detection tasks demonstrate that \texttt{DOG} achieves the optimal performance compared to its advanced counterparts.

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: 3181