[Paper Review] Deep Anomaly Detection Using Geometric Transformations

Izhak Golan, Ran El‐Yaniv|arXiv (Cornell University)|May 28, 2018

Anomaly Detection Techniques and Applications34 references291 citations

TL;DR

The paper introduces a discriminative anomaly detection method that trains a multi-class classifier to identify geometric transformations applied to normal images, and uses a Dirichlet-based normality score at test time to detect out-of-distribution samples.

ABSTRACT

We consider the problem of anomaly detection in images, and present a new detection technique. Given a sample of images, all known to belong to a "normal" class (e.g., dogs), we show how to train a deep neural model that can detect out-of-distribution images (i.e., non-dog objects). The main idea behind our scheme is to train a multi-class model to discriminate between dozens of geometric transformations applied on all the given images. The auxiliary expertise learned by the model generates feature detectors that effectively identify, at test time, anomalous images based on the softmax activation statistics of the model when applied on transformed images. We present extensive experiments using the proposed detector, which indicate that our algorithm improves state-of-the-art methods by a wide margin.

Motivation & Objective

• Motivate robust anomaly detection for vision systems under distributional shift and noise.
• Propose a non-generative, discriminative framework that leverages geometric transformations to learn informative features from normal data.
• Develop a probabilistic scoring approach to quantify normality without reconstruction.
• Evaluate the method against state-of-the-art anomaly detection baselines on multiple image datasets.

Proposed method

• Create a self-labeled dataset by applying a set of geometric transformations T to all normal training images and train a k-class classifier to predict the applied transformation index.
• Obtain a normality score n_S(x) for any test image by averaging the log-likelihoods of the softmax outputs across all transformed versions of x, under Dirichlet parameter estimates derived from the normal training set.
• Estimate Dirichlet parameters alpha_i for each transformation i using a fixed-point iteration on the transformed training outputs.
• Use a simplified variant n_hat_S(x) that averages the softmax outputs and, in preliminary experiments, still achieves strong performance.
• Employ a Wide Residual Network trained on the augmented self-labeled set, with standard cross-entropy loss, optimized via Adam.

Experimental results

Research questions

• RQ1 Can learning to discriminate among many geometric transformations on normal images yield features effective for detecting anomalies?
• RQ2 How can transformed-image predictions be aggregated into a reliable normality score for out-of-distribution detection?
• RQ3 What is the empirical performance of the transformation-based anomaly detector compared to existing deep methods across multiple datasets?
• RQ4 Does the method extend to identifying out-of-distribution samples in labeled multi-class datasets?
• RQ5 What is the impact of the transformation set choice on anomaly detection performance?

Key findings

• On CIFAR-10, the method improves the top baseline AUROC by a substantial margin across 10 experiments (average increase reported as 32%).
• On CatsVsDogs, the method achieves a large improvement over the best baseline (67% AUROC) in the reported setting.
• Across datasets, the approach often outperforms reconstruction-based and other deep anomaly detection methods, especially with larger image sizes.
• In CIFAR-100, performance varies by normal class due to intra-class diversity, with some classes showing weaker results.
• On fashion-MNIST, most methods perform well and the proposed method shows a slight advantage.
• The approach can also improve out-of-distribution detection in labeled multi-class settings, surpassing some established baselines on reported tasks.