[Paper Review] Anomaly Detection with Adversarial Dual Autoencoders

Ha Son Vu, Daisuke Ueta+4 more|arXiv (Cornell University)|Feb 19, 2019

Anomaly Detection Techniques and Applications10 references35 citations

TL;DR

The paper introduces Adversarial Dual Autoencoders (ADAE), a GAN-based anomaly detection framework using a dual-autoencoder setup forGenerator andDiscriminator to improve training stability, with anomaly scores derived from the discriminator’s reconstruction error.

ABSTRACT

Semi-supervised and unsupervised Generative Adversarial Networks (GAN)-based methods have been gaining popularity in anomaly detection task recently. However, GAN training is somewhat challenging and unstable. Inspired from previous work in GAN-based image generation, we introduce a GAN-based anomaly detection framework - Adversarial Dual Autoencoders (ADAE) - consists of two autoencoders as generator and discriminator to increase training stability. We also employ discriminator reconstruction error as anomaly score for better detection performance. Experiments across different datasets of varying complexity show strong evidence of a robust model that can be used in different scenarios, one of which is brain tumor detection.

Motivation & Objective

• Motivate semi-supervised/unsupervised anomaly detection with GANs and address training stability issues.
• Introduce a dual-autoencoder GAN framework to stabilize training.
• Leverage discriminator reconstruction error as a meaningful anomaly score.
• Demonstrate applicability across datasets of varying complexity, including brain tumor detection.

Proposed method

• Implement two autoencoders as generator and discriminator within a GAN framework to enhance training stability.
• Use the discriminator’s reconstruction error as the anomaly score for detection.
• Leverage adversarial training to improve the quality of representation for anomaly characterization.
• Validate ADAE across multiple datasets with varying complexity to assess robustness.

Experimental results

Research questions

• RQ1 Can Adversarial Dual Autoencoders improve training stability in GAN-based anomaly detection compared to single-autoencoder setups?
• RQ2 Does using the discriminator reconstruction error yield effective anomaly scores across datasets of varying complexity?
• RQ3 How does ADAE perform relative to other GAN-based anomaly detection methods on tasks like brain tumor detection?

Key findings

• ADAE demonstrates strong evidence of robustness across datasets with different levels of complexity.
• Discriminator reconstruction error provides a meaningful anomaly score that enhances detection performance.
• The dual-autoencoder setup contributes to improved training stability in adversarial anomaly detection.
• Experiments indicate ADAE can be applied effectively to real-world tasks such as brain tumor detection.