DisCoNet: Rethinking Adversarial Networks for Discriminator-Driven Distribution Modeling

Francisco Caetano, Christiaan Viviers, Peter H.N. de With, Fons van der Sommen

Keywords: Covariate Shift, Deep Learning, Generative Adversarial Networks, Generative Modelling, Out-of-Distribution Detection

Abstract: Out-of-distribution (OOD) detection holds significant importance across various applications. While semantic and domain-shift OOD problems are well-documented, this work focuses on the nuances of covariate shifts, which entail subtle perturbations or variations in the data distribution. These disturbances have proven to negatively impact machine learning performance. We have found that existing OOD detection methods often struggle to effectively distinguish covariate shifts from in-distribution instances, emphasizing the need for specialized solutions. Therefore, we propose DisCoNet, an Adversarial Variational Autoencoder (VAE) that rethinks the Generative Adversarial Networks paradigm. Instead of prioritizing the generator as the network's core, we focus on the discriminator, using the generator as a supporting training tool. DisCoNet uses the VAE's suboptimal outputs as negative samples to train the discriminator, thereby improving its ability to delineate the boundary between in-distribution samples and covariate shifts. By tightening this in-distribution boundary, DisCoNet achieves state-of-the-art results in public OOD detection benchmarks. The proposed model not only excels in detecting covariate shifts, achieving 98.9% AUROC on ImageNet-1K(-C), but also outperforms all prior methods on public semantic OOD benchmarks. With a model size of $\leq$ 25MB, it is highly effective on Far-OOD (OpenImage-O (99.4%) and iNaturalist (100.0%)) and Near-OOD (SSB-hard (99.9%) and NINCO (99.7%)) detection. The code will be made publicly available.

Primary Area: generative models

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.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 9714