[Paper Review] A Survey on GANs for Anomaly Detection

Federico Di Mattia, Paolo Galeone+2 more|arXiv (Cornell University)|Jun 27, 2019

Anomaly Detection Techniques and Applications6 references114 citations

TL;DR

This paper surveys principal GAN-based anomaly detection methods (AnoGAN, BiGAN/EGBAD, GANomaly), provides empirical evaluation across datasets, and releases an open-source anomaly detection toolbox.

ABSTRACT

Anomaly detection is a significant problem faced in several research areas. Detecting and correctly classifying something unseen as anomalous is a challenging problem that has been tackled in many different manners over the years. Generative Adversarial Networks (GANs) and the adversarial training process have been recently employed to face this task yielding remarkable results. In this paper we survey the principal GAN-based anomaly detection methods, highlighting their pros and cons. Our contributions are the empirical validation of the main GAN models for anomaly detection, the increase of the experimental results on different datasets and the public release of a complete Open Source toolbox for Anomaly Detection using GANs.

Motivation & Objective

• Motivate anomaly detection as a problem where normal behavior is modeled via adversarial training.
• Summarize and compare GAN-based architectures for anomaly detection (AnoGAN, EGBAD, GANomaly).
• Provide empirical validation across multiple datasets and configure a public toolbox for replication.

Proposed method

• Explain GAN foundations and key extensions (conditional GANs, BiGAN).
• Describe AnoGAN as mapping from data to latent space and back for anomaly scoring.
• Present BiGAN/EGBAD to enable encoder learning for faster inference and inverse mapping.
• Detail GANomaly which combines autoencoder-like structure with GAN training and multiple loss components.
• Publicly release a TensorFlow-based anomaly detection toolbox and reproduce multiple configurations.

Experimental results

Research questions

• RQ1 How can GAN-based frameworks model normal data for effective anomaly detection?
• RQ2 What are the strengths and limitations of AnoGAN, BiGAN/EGBAD, and GANomaly for anomaly detection?
• RQ3 Can a unified TensorFlow implementation replicate and compare these GAN-based approaches across datasets?
• RQ4 How do different loss formulations and architectures affect detection performance and inference speed?

Key findings

• GAN-based methods can be effective for anomaly detection across image and tabular data.
• BiGAN/EGBAD and GANomaly provide faster inference and more interpretable anomaly scores than AnoGAN in practice.
• GANomaly uses an encoder during training and an autoencoder-like generator to speed up learning and improve interpretability.
• On KDD, BiGAN/EGBAD achieved higher precision, recall, and F1 than GANomaly (0.941174, 0.956155, 0.948605 vs 0.830256, 0.841112, 0.835648).
• The authors released an open-source anomaly detection toolbox implemented in TensorFlow for replication and extension.