[Paper Review] Explainable Deep Few-shot Anomaly Detection with Deviation Networks

Guansong Pang, Choubo Ding+2 more|arXiv (Cornell University)|Aug 1, 2021

Anomaly Detection Techniques and Applications68 references43 citations

TL;DR

Presents DevNet, a deviation-based, end-to-end deep framework for few-shot anomaly detection that directly learns anomaly scores guided by a Gaussian prior and MIL-based deviation loss, enabling open-set detection and explanation.

ABSTRACT

Existing anomaly detection paradigms overwhelmingly focus on training detection models using exclusively normal data or unlabeled data (mostly normal samples). One notorious issue with these approaches is that they are weak in discriminating anomalies from normal samples due to the lack of the knowledge about the anomalies. Here, we study the problem of few-shot anomaly detection, in which we aim at using a few labeled anomaly examples to train sample-efficient discriminative detection models. To address this problem, we introduce a novel weakly-supervised anomaly detection framework to train detection models without assuming the examples illustrating all possible classes of anomaly. Specifically, the proposed approach learns discriminative normality (regularity) by leveraging the labeled anomalies and a prior probability to enforce expressive representations of normality and unbounded deviated representations of abnormality. This is achieved by an end-to-end optimization of anomaly scores with a neural deviation learning, in which the anomaly scores of normal samples are imposed to approximate scalar scores drawn from the prior while that of anomaly examples is enforced to have statistically significant deviations from these sampled scores in the upper tail. Furthermore, our model is optimized to learn fine-grained normality and abnormality by top-K multiple-instance-learning-based feature subspace deviation learning, allowing more generalized representations. Comprehensive experiments on nine real-world image anomaly detection benchmarks show that our model is substantially more sample-efficient and robust, and performs significantly better than state-of-the-art competing methods in both closed-set and open-set settings. Our model can also offer explanation capability as a result of its prior-driven anomaly score learning. Code and datasets are available at: https://git.io/DevNet.

Motivation & Objective

• Motivate and address few-shot anomaly detection where only a few labeled anomalies are available.
• Learn a discriminative and robust model of normality while handling unseen anomalies.
• Provide end-to-end anomaly score learning rather than indirect feature learning.
• Enable anomaly explanation via gradient-based localization tied to the learned scores.

Proposed method

• DevNet directly learns an anomaly scoring function that maps inputs to scalar anomaly scores.
• A Gaussian prior-based reference score is used to guide end-to-end learning of anomaly scores.
• A MIL-based top-K deviation loss enforces that anomaly scores deviate from the prior in the upper tail for anomalies while keeping normal scores near the prior.
• An end-to-end network combines a fine-grained feature learner with a simple linear anomaly score head.
• Anomaly explanation is achieved via gradient back-propagation to localize contributing features.

Experimental results

Research questions

• RQ1 Can few labeled anomalies guide end-to-end learning of anomaly scores for both seen and unseen anomaly classes?
• RQ2 Does leveraging a Gaussian prior and MIL-based deviation loss improve sample efficiency and generalization in FSAD?
• RQ3 Can the model provide faithful anomaly localization/explanation tied to the learned scores?
• RQ4 Is the approach robust to anomaly contamination in normal data and effective across different backbones and datasets?

Key findings

• DevNet achieves improved sample efficiency and robustness over state-of-the-art methods in nine real-world image anomaly benchmarks.
• The framework yields better generalization to unseen anomaly classes in open-set settings.
• Anomaly scores are interpretable due to the prior-driven learning and can be localized to input regions via gradient-based explanations.
• Top-K MIL deviation learning enhances learning by focusing on the most informative anomalous patches, reducing false detections.
• DevNet provides accurate anomaly localization supporting explanation of detected anomalies.