f-AnoGAN: Fast unsupervised anomaly detection with generative adversarial networks

TensorFlow implementation of f-AnoGAN with MNIST dataset [1].
The base model WGAN is also implemented with TensorFlow.

Summary

f-AnoGAN architecture

The architecture of f-AnoGAN [1].

The logic for calculating anomaly score [1].

Graph in TensorBoard

Graph of f-AnoGAN.

Problem Definition

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

Results

Training Phase-1 (WGAN Training)

Training graph of Phase-1

The rear half of the graph represents the state of the training phase 2.

|Term Real|Term Fake|
|:—-:|:—-:|
|||

|Loss D (Discriminator)|Loss G (Generator)|
|:—-:|:—-:|
|||

Result of Phase-1

|z:2|z:2 (latent space walking)|
|:—-:|:—-:|
|||

|z:64|z:128|
|:—-:|:—-:|
|||

Training Phase-2 (izi Training)

Training graph of Phase-2

The front half of the graph represents the state of the training phase 1.

|Term izi|Term ziz|Loss E (Encoder)|
|:—-:|:—-:|:—-:|
||||

Result of Phase-2

Restoration result by f-AnoGAN.

Test Procedure

Box plot with encoding loss of test procedure.

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] Schlegl, Thomas, et al (2019). f-AnoGAN: Fast unsupervised anomaly detection with generative adversarial networks. Medical image analysis 54 (2019): 30-44.