[TensorFlow] GANomaly: Semi-Supervised Anomaly Detection via Adversarial Training

TensorFlow implementation of GANomaly with MNIST dataset.
PyTorch Version is also implemented.

Summary

GANomaly architecture

Simplified GANomaly architecture.

Graph in TensorBoard

Graph of GANomaly.

Problem Definition

'Class-1' is defined as normal and the others are defined as abnormal.

Results

Training Procedure

Loss graph in the training procedure.
Each graph shows encoding loss, reconstruction loss, adversarial loss, and total (target) loss respectively.

Restoration result by GANomaly.

Test Procedure

Box plot with encoding loss of test procedure.

Normal samples classified as normal.

Abnormal samples classified as normal.

Normal samples classified as abnormal.

Abnormal samples classified as abnormal.

Environment

• Python 3.7.4
• Tensorflow 1.14.0
• Numpy 1.17.1
• Matplotlib 3.1.1
• Scikit Learn (sklearn) 0.21.3

Reference

[1] S Akcay, et al. (2018). Ganomaly: Semi-supervised anomaly detection via adversarial training.. arXiv preprint arXiv:1805.06725.