Robust Out-of-Distribution Detection Based on Effective Points Select

Out-of-distribution (OOD) detection plays a crucial role in the security of deep learning systems deployed in open scenarios. In recent years, research has shown that OOD detectors are vulnerable to adversarial examples, which further threatens the security of the systems. To address this challenge, some methods combine outlier exposure with adversarial training, generating adversarial examples using an auxiliary OOD dataset and enhancing the model's robustness through adversarial training. However, these methods adopt a random selection strategy when choosing the auxiliary OOD data, which leads to the introduction of a large amount of irrelevant information and results in less compact decision boundaries. Additionally, the effectiveness of adversarial training methods in OOD detection is also unsatisfactory. To overcome these limitations, we introduce an effective point selection method. This method uses the Mahalanobis distance to screen out effective auxiliary OOD data in the aggregated low-dimensional embedding space and preferentially uses the effective points during training to reduce the introduction of irrelevant information. To ensure adversarial robustness, we integrate the latest advancements in the field of neural ordinary differential equations (ODEs) to achieve robust OOD detection and use the selected effective points to train the neural ODE module. Comprehensive experiments conducted on multiple benchmark datasets verify the effectiveness and superiority of our method, demonstrating its capability in enhancing the robustness of OOD detection.