Generalized Anomaly Detection with Knowledge Exposure:The Dual Effects of Augmentation

Mohammad Akhavan Anvari, Rojina Kashefi, Vahid Reza Khazaie, Mohammad Khalooei, Mohammad Sabokrou

Keywords: OOD gneralziation, OOD detection, Anomaly detection, One class classification, Outlier Expousre

Abstract: Anomaly detection involves identifying samples that deviate from the training data. While previous methods have demonstrated significant performance, our experiments reveal that their generalization ability declines substantially when faced with slight shifts in the test data. This limitation stems from an underlying assumption: these methods generally expect the distribution of normal test samples to closely resemble that of the training set, while anomalies are presumed to be far from this distribution. However, in real-world scenarios, test samples often experience varying degrees of distributional shift while retaining their semantic consistency. The ability to generalize successfully to semantically preserved transformations while accurately detecting normal samples whose semantic meaning has changed as anomalies is critical for a model's trustworthiness and reliability. For instance, while a rotation may alter the semantic meaning of a car in the context of anomaly detection, it typically preserves the meaning of an apple. Yet, current methods, particularly those based on contrastive learning, are likely to detect both as anomalies. This complexity underscores the need for dynamic learning procedures grounded in a deeper understanding of outliers. To address this, we propose a novel approach called Knowledge Exposure (KE), which incorporates external knowledge to interpret concept dynamics and distinguish between transformations that induce semantic shifts. Our approach improves generalization by leveraging insights from a pre-trained CLIP model to assess the significance of anomalies for each concept. Evaluations on datasets such as CIFAR-10, CIFAR-100, SVHN demonstrate superior performance compared to previous methods, validating the effectiveness of our approach.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

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: 7726