Catching Both Gray and Black Swans: Open-set Supervised Anomaly Detection (2203.14506v1)

Abstract: Despite most existing anomaly detection studies assume the availability of normal training samples only, a few labeled anomaly examples are often available in many real-world applications, such as defect samples identified during random quality inspection, lesion images confirmed by radiologists in daily medical screening, etc. These anomaly examples provide valuable knowledge about the application-specific abnormality, enabling significantly improved detection of similar anomalies in some recent models. However, those anomalies seen during training often do not illustrate every possible class of anomaly, rendering these models ineffective in generalizing to unseen anomaly classes. This paper tackles open-set supervised anomaly detection, in which we learn detection models using the anomaly examples with the objective to detect both seen anomalies (gray swans') and unseen anomalies (black swans'). We propose a novel approach that learns disentangled representations of abnormalities illustrated by seen anomalies, pseudo anomalies, and latent residual anomalies (i.e., samples that have unusual residuals compared to the normal data in a latent space), with the last two abnormalities designed to detect unseen anomalies. Extensive experiments on nine real-world anomaly detection datasets show superior performance of our model in detecting seen and unseen anomalies under diverse settings. Code and data are available at: https://github.com/choubo/DRA.

Open-set Supervised Anomaly Detection: Addressing the Challenge of Unseen Anomalies

The paper, "Catching Both Gray and Black Swans: Open-set Supervised Anomaly Detection," presents a novel framework, DRA (Disentangled Representations of Abnormalities), that advances the capacity of anomaly detection models to effectively identify both detected ("gray swans") and undetected anomalies ("black swans"). Given the increasing importance of detecting anomalies in various domains such as industrial quality inspection, medical image analysis, and autonomous driving, the paper addresses a critical gap in existing supervised anomaly detection methodologies—the challenge of generalizing to unseen anomaly categories not represented in the training data.

The authors recognize that whilst labeled anomaly samples are often available, they typically represent a narrowed set of anomalies encountered historically. Thus, models trained on such data tend to overfit to these seen anomalies, significantly underperforming when faced with unseen anomalies that deviate from known patterns. The DRA model seeks to circumvent this limitation by leveraging three types of learned abnormality representations: seen, pseudo, and latent residual abnormalities.

Core Contributions

1. Disentanglement of Abnormality Representations: The model proposes learning disentangled representations specific to different categories of abnormalities. By doing so, it encourages the learning of broad and generalized features that extend beyond the scope of anomalies seen in the training data.
2. Innovative Multi-head Network Architecture: DRA employs a multi-head architecture, where separate network branches are responsible for different types of anomalies. This design enhances the model's ability to capture diverse anomaly patterns, specifically addressing overfitting issues.
3. Latent Residual Abnormality Learning: A critical innovation in DRA is the use of latent residuals, which are differences in feature space representations between anomalies and normal data. This allows the model to detect subtle anomalies that might not be apparent within the original feature space.
4. Comprehensive Evaluation: The paper reports extensive experiments over nine real-world datasets, outperforming current state-of-the-art models in both seen and open-set, unsupervised contexts. The performance gains, particularly in scenarios with minimal labeled anomalies, emphasize the model's robustness and promise.

Numerical Insights and Performance

The advancement brought by DRA is quantitatively substantiated by superior performance metrics such as AUC scores across diverse datasets — spanning from industrial inspection datasets like MVTec AD and AITEX to medical datasets like Hyper-Kvasir. The ability of DRA to outperform not only supervised but also unsupervised baselines even when trained with as few as one labeled anomaly demonstrates its strong generalization capabilities.

Theoretical and Practical Implications

The success of DRA signifies a shift in anomaly detection paradigms, suggesting a more resilient approach to generalize well in unsupervised environments. Practically, this means industries can leverage DRA even when varying anomaly examples exist. Theoretically, it invites further exploration into disentanglement techniques in other machine learning applications, promoting robustness in open-set tasks.

Future Directions

This research opens several avenues for further investigation. Enhancing the adaptability of such models to varying scales of anomalies and refining feature extraction processes could augment their effectiveness. Moreover, extending the DRA framework to other complex domains, such as network security, could test its efficacy in environments characterized by rapidly evolving anomaly patterns.

In conclusion, the proposed DRA framework presents an influential contribution to the ongoing development of anomaly detection methodologies, illustrating a thoughtful integration of disentangled learning principles to meet the evolving challenges of open-set scenarios.