[Paper Review] Explainable Deep One-Class Classification

Philipp Liznerski, Lukas Ruff+4 more|arXiv (Cornell University)|Jul 3, 2020

Anomaly Detection Techniques and Applications61 references94 citations

TL;DR

FCDD combines fully convolutional networks with a one-class objective to produce anomaly heatmaps as explanations, achieving competitive anomaly detection and state-of-the-art pixel-wise explanations on MVTec-AD, with benefits from semi-supervised tuning using ground-truth anomaly maps.

ABSTRACT

Deep one-class classification variants for anomaly detection learn a mapping that concentrates nominal samples in feature space causing anomalies to be mapped away. Because this transformation is highly non-linear, finding interpretations poses a significant challenge. In this paper we present an explainable deep one-class classification method, Fully Convolutional Data Description (FCDD), where the mapped samples are themselves also an explanation heatmap. FCDD yields competitive detection performance and provides reasonable explanations on common anomaly detection benchmarks with CIFAR-10 and ImageNet. On MVTec-AD, a recent manufacturing dataset offering ground-truth anomaly maps, FCDD sets a new state of the art in the unsupervised setting. Our method can incorporate ground-truth anomaly maps during training and using even a few of these (~5) improves performance significantly. Finally, using FCDD's explanations we demonstrate the vulnerability of deep one-class classification models to spurious image features such as image watermarks.

Motivation & Objective

• Motivate the need for explainable deep anomaly detection in safety-, security-, and decision-support contexts.
• Develop a deep one-class classifier whose outputs are interpretable anomaly heatmaps.
• Assess performance on standard benchmarks (Fashion-MNIST, CIFAR-10, ImageNet) and manufacturing data (MVTec-AD).
• Demonstrate how ground-truth anomaly maps can improve training and how explanations reveal model behavior.

Proposed method

• Use a fully convolutional network (FCN) to map images to a low-resolution anomaly heatmap via spatially preserved features.
• Adopt a hypersphere classifier objective (HSC) variant to push nominal data toward a center and anomalies away from it, using a pseudo-Huber loss.
• Define anomaly heatmaps A(X) as the FCN output transformed by a robust loss, and use the L1 norm of A(X) as the anomaly score.
• Upsample the low-resolution heatmaps to full resolution using a principled receptive-field based Gaussian upsampling scheme (strided transposed convolution with a fixed Gaussian kernel).
• Support learning from Outlier Exposure (OE) with synthetic or real anomalies, and enable semi-supervised training using pixel-level anomaly maps when available.
• Provide an explanation mechanism intrinsically tied to the anomaly score, enabling inspection of which image regions contribute to the anomaly.

Experimental results

Research questions

• RQ1 Can a one-class deep learning model produce interpretable, spatial anomaly heatmaps tied to the anomaly score?
• RQ2 How does FCDD perform on standard anomaly detection benchmarks compared to existing methods, both with and without outlier exposure?
• RQ3 Does incorporating ground-truth anomaly maps (even sparsely) improve detection and explanation quality in a semi-supervised setting?
• RQ4 Are the produced explanations robust to spurious features (Clever Hans effects) and can they reveal such vulnerabilities?
• RQ5 How do heatmaps compare to gradient-based and reconstruction-based explanations in terms of noise and structure?

Key findings

• FCDD achieves competitive anomaly detection performance close to state-of-the-art on CIFAR-10 and ImageNet, while providing interpretable heatmaps as explanations.
• On MVTec-AD, FCDD sets a new state-of-the-art for unsupervised pixel-wise anomaly localization, with further gains in a semi-supervised setting when ground-truth anomaly maps are used.
• FCDD heatmaps are less noisy and more structured than gradient-based or autoencoder explanations, improving interpretability of detections.
• The method can reveal vulnerabilities to spurious features (Clever Hans effect) in deep one-class models, illustrating the value of transparent explanations.
• FCDD supports effective use of Outlier Exposure and synthetic anomalies, achieving strong explanations even with minimal labeled anomaly data.