Out-of-distribution Detection with Diffusion-based Neighborhood

Luping Liu, Yi Ren, Xize Cheng, Zhou Zhao

Keywords: OOD detection, diffusion model

Abstract: Out-of-distribution (OOD) detection is an important task to ensure the reliability and safety of deep learning and the discriminator models outperform others for now. However, the feature extraction of such models must compress the data and lose certain information, leaving room for bad cases and malicious attacks. However, despite effectively fitting the data distribution and producing high-quality samples, generative models lack suitable indicator scores to match with discriminator models in the OOD detection tasks. In this paper, we find that these two kinds of models can be combined to solve each other's problems. We introduce diffusion models (DMs), a kind of powerful generative model, into OOD detection and find that the denoising process of DMs also functions as a novel form of asymmetric interpolation. This property establishes a diffusion-based neighborhood for each input data. Then, we perform discriminator-based OOD detection based on the diffusion-based neighborhood instead of isolated data. In this combination, the discriminator models provide detection metrics for generation models and the diffusion-based neighborhood reduces the information loss of feature extraction. According to our experiments on CIFAR10 and CIFAR100, our new methods successfully outperform state-of-the-art methods. Our implementation is put in the supplementary materials.

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.

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: Yes

Please Choose The Closest Area That Your Submission Falls Into: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)

TL;DR: We design a general strategy to combine a diffusion model and a Resnet to do OOD detection.