1 of 52

Anomaly Detection

5/21

2 of 52

Outline

What is Anomaly Detection
Classic Method

With Classifier
GMM (Gaussian Mixture Model)
Auto-Encoder
PCA
Isolation Forest
Summary

Anomaly Detection on image

AnoGAN
EGBAD
GANomaly
Summary

Anomaly Detection on Audio

GMGAN

3 of 52

What is Anomaly Detection

4 of 52

What is Anomaly

Training Data

Anomaly

5 of 52

Classic Method

6 of 52

With Classifier

7 of 52

With Classifier

8 of 52

GMM (Gaussian Mixture Model)

1-dim Gaussian Mixtures

http://speech.ee.ntu.edu.tw/DSP2019Autumn/Slides/2.0.pptx

9 of 52

GMM (Gaussian Mixture Model)

2-dim Gaussian

http://speech.ee.ntu.edu.tw/DSP2019Autumn/Slides/2.0.pptx

10 of 52

GMM (Gaussian Mixture Model)

N-dim Gaussian Mixtures

http://speech.ee.ntu.edu.tw/DSP2019Autumn/Slides/2.0.pptx

11 of 52

GMM (Gaussian Mixture Model)

12 of 52

GMM (Gaussian Mixture Model)

13 of 52

GMM (Gaussian Mixture Model)

https://towardsdatascience.com/a-simple-way-to-detect-anomaly-3d5a48c0dae0

14 of 52

GMM (Gaussian Mixture Model)

https://towardsdatascience.com/a-simple-way-to-detect-anomaly-3d5a48c0dae0

15 of 52

GMM (Gaussian Mixture Model)

https://towardsdatascience.com/a-simple-way-to-detect-anomaly-3d5a48c0dae0

16 of 52

Auto-Encoder

17 of 52

Auto-Encoder

Examples:

https://www.kaggle.com/tikedameu/anomaly-detection-with-autoencoder-pytorch

https://towardsdatascience.com/anomaly-detection-with-autoencoder-b4cdce4866a6

18 of 52

PCA

http://speech.ee.ntu.edu.tw/~tlkagk/courses/ML_2017/Lecture/PCA%20(v3).pdf

19 of 52

PCA

https://www.oreilly.com/library/view/hands-on-unsupervised-learning/9781492035633/ch04.html

20 of 52

PCA

https://www.oreilly.com/library/view/hands-on-unsupervised-learning/9781492035633/ch04.html

21 of 52

PCA

https://www.oreilly.com/library/view/hands-on-unsupervised-learning/9781492035633/ch04.html

22 of 52

PCA

n_components = 30 n_components = 27

https://www.oreilly.com/library/view/hands-on-unsupervised-learning/9781492035633/ch04.html

23 of 52

Isolation Forest

Use tree like structure to split data

24 of 52

Isolation Forest

https://towardsdatascience.com/outlier-detection-with-isolation-forest-3d190448d45e

25 of 52

Isolation Forest

https://towardsdatascience.com/outlier-detection-with-isolation-forest-3d190448d45e

26 of 52

Isolation Forest

https://towardsdatascience.com/outlier-detection-with-isolation-forest-3d190448d45e

27 of 52

Isolation Forest

https://towardsdatascience.com/outlier-detection-with-isolation-forest-3d190448d45e

28 of 52

Summary of classic method

With Classifier
GMM (Gaussian Mixture Model)
Auto-Encoder
PCA
Isolation Forest

29 of 52

Anomaly detection on image

30 of 52

Typical GANs

http://speech.ee.ntu.edu.tw/~tlkagk/courses/MLDS_2018/Lecture/GAN%20(v2).pdf

31 of 52

Typical GANs

https://arxiv.org/pdf/1906.11632.pdf

32 of 52

Example

33 of 52

AnoGAN

Train a standard GAN only on positive samples

https://arxiv.org/pdf/1906.11632.pdf

34 of 52

AnoGAN

Anomaly Score

for 𝛾 = 1, 2, … 𝜞 find proper z

https://arxiv.org/pdf/1905.13147.pdf

35 of 52

AnoGAN

Pros

Showed that GANs can be used for anomaly detection
Introduced a new mapping scheme from latent space to input data space.
Used the same mapping scheme to define an anomaly score.

Cons

Requires Γ optimization steps for every new input: bad test-time performance
The GAN objective has not been modified to take into account the need for the inverse mapping learning.
The anomaly score is difficult to interpret, not being in the probability range.

https://arxiv.org/pdf/1906.11632.pdf

36 of 52

EGBAD (Efficient GAN-Based Anomaly Detection)

Train a Bi-GAN only on positive samples

https://arxiv.org/pdf/1906.11632.pdf

37 of 52

EGBAD (Efficient GAN-Based Anomaly Detection)

Pros

The Encoder E can learns how to encode image while adversarial training.
Therefore, it can bypass Γ optimization steps of AnoGAN while calculating anomaly score.

38 of 52

GANomaly

Generator is composed of encoder G_E, decoder G_D, and encoder E
Trained on only normal data.

Anomaly score

https://arxiv.org/pdf/1906.11632.pdf

https://arxiv.org/pdf/1905.13147.pdf

39 of 52

GANomaly

Pros

An encoder is learned during the training process, so it can bypass the Γ optimization.
Using an autoencoder like architecture (no use of noise prior) makes the entire learning process faster.
The contextual loss can be used to localize the anomaly.

Cons

Defines a new anomaly score.
It allows to detect anomalies both in the image space and in the latent space, but the results couldn’t match:

a higher anomaly score, that’s computed only in the latent space, can be associated with a generated sample with a low contextual loss value and thus very similar to the input - and vice versa.

https://arxiv.org/pdf/1906.11632.pdf