Title

Anomaly Detection with Generative Adversarial Active Learning with Gradient Penalty

Abstract

Abstract

Anomaly detection has become a very important topic with the advancing machine learning techniques and is used in many different application areas. In this study, we approach differently than the anomaly detection methods performed on standard generative models and describe anomaly detection as a binary classification problem. However, in order to train a highly accurate classifier model, the number of anomaly data in data-sets is very limited, and with synthetic data produced using generative models, it can be brought to a usable level to train the model. In our model like GANs while Generator produces potential informative anomaly data, the Discriminator tries to determine whether the generated data is fake or real. In addition to these, we have added the Critic network to our model in order to enable the Generator to produce more realistic and informative data. In this way, we designed our model the Discriminator to be trained with the data produced by the Generator which is improved by the Critic network . Therefore, after sufficient training, the Discriminator turns into a natural anomaly detection classification tool. Since the Generator produce more realistic data in each epoch during the training phase, it produce more informative potential anomaly data for the Discriminator, which will allow the algorithm to develop with more informative data with active learning logic. Our novelty is a generative adversarial active learning (GAAL) structure designed over the Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) instead of just applying this method over the standard GAN model. In this way, both our Generator model can produce more realistic and more informative data than before, and at the same time, it prevents the mode collapse problem, which is one of the biggest problems of the standard GAN model. In this way, we have obtained a model that can detect anomalies with higher accuracy. Improved version of Wasserstein Generative Adversarial Active Learning (WGAAL-GP) has been tested on different data sets and the results obtained are presented in this study by comparing them with previous studies.

Supervisor(s)

Supervisor(s)

HASAN ALI DURAN

Date and Location

Date and Location

2021-09-08 13:30:00

Category

Category

MSc_Thesis