Robust out-of-distribution detection based on non-parametric adversarial virtual outlier generation

Out-of-distribution (OOD) detection plays a crucial role in deploying deep learning systems in open-world environments. Despite notable advancements in OOD detection, research has shown that adversarial examples continue to pose a serious challenge to the security of these models. To tackle this challenge, recent defense strategies have been developed that integrate outlier exposure with adversarial training. These approaches utilize auxiliary OOD datasets to provide models with insights into OOD characteristics. However, collecting and cleaning diverse OOD datasets can be challenging in some cases. Furthermore, using specific auxiliary datasets to represent OOD samples may cause the model to bias toward particular OOD instances, reducing its ability to generalize to unseen OOD data during inference. To address these limitations, we introduce a non-parametric method that synthesizes virtual outliers in a robust feature space. Our method employs adversarial propagation and sampling within a robust latent space to generate OOD embeddings, enabling the model to effectively differentiate between in-distribution and out-of-distribution inputs. To ensure adversarial robustness, we integrate the latest advancements in neural ordinary differential equations (ODEs) for robust OOD detection and employ synthesized virtual outliers to train neural ODE modules. To further enhance the model's OOD detection capability, we employ a logit-norm layer for normalization. Comprehensive experiments conducted on multiple benchmarks validate the effectiveness and superiority of our approach, showcasing its capability to enhance robustness in OOD detection.