yeonghyeon / adae-tf

TensorFlow implementation of Anomaly Detection with Adversarial Dual Autoencoders (ADAE) with MNIST dataset.

[TensorFlow] Anomaly Detection with Adversarial Dual Autoencoders

TensorFlow implementation of Anomaly Detection with Adversarial Dual Autoencoders (ADAE) with MNIST dataset.
The Keras implementation is provided as the following link.
https://github.com/kjm1559/ADAE_LSTM_Autoencoder

Architecture

Objective Functions

The objective functions (losses) for training ADAE [1].

ADAE architecture

The architecture of ADAE.

Graph in TensorBoard

Graph of ADAE.

Problem Definition

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

Results

Training Procedure

Loss graphs in the training procedure.
Each graph shows the generative loss, and the two terms that make loss-G.

Loss graphs in the training procedure.
Each graph shows the discriminative loss, and the two terms that make loss-G.

Restoration result by ADAE.

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] Ha Son Vu et al. (2019). Anomaly Detection with Adversarial Dual Autoencoders. arXiv preprint arXiv:1902.06924.