Electronic Thesis and Dissertation Repository

Thesis Format

Monograph

Degree

Master of Science

Program

Computer Science

Supervisor

Yalda Mohsenzadeh

Abstract

Anomaly Detection (AD) is to identify samples that differ from training observations in some way. Those samples that do not follow the distribution of normal data are called outliers or anomalies. In this thesis, we examined two different challenges related to deep learning-based anomaly detection methods. The first challenge is the generalizability to outliers. A wide range of unsupervised anomaly detection methods use deep autoencoders as a foundation. However, a notable limitation of deep autoencoders is that they generalize to outliers and reconstruct them with low error. In order to overcome this issue, we propose an adversarial framework consisting of two competing components, an adversarial distorter, and an autoencoder. During training, the adversarial distorter produces perturbations that are applied to the encoder’s latent space to maximize the reconstruction error. The autoencoder attempts to neutralize the effects of these perturbations to minimize the reconstruction error. Another challenge is the high computational cost, complexity, and unstable training procedures of deep anomaly detection methods. Despite being successful at anomaly detection, deep neural networks are difficult to deploy in real-world applications because of this challenge. We overcome this problem by using a simple learning procedure that trains a lightweight convolutional neural network. We propose to solve anomaly detection as a supervised regression problem. We label normal and anomalous data using two separable distributions of continuous values. As a way to compensate for the lack of anomalous samples during training, we use straightforward image augmentation techniques to create a distinct set of anomalous samples. An augmented set has a distribution that is similar to normal data, but deviates slightly from it, while real anomalies should have a further distribution. Consequently, training a regressor on normal and these augmented samples will result in more distinct distributions of labels for normal and real anomalous data points. In several image and video anomaly detection benchmarks, our methods outperform cutting-edge approaches.

Summary for Lay Audience

The goal of anomaly detection is to recognize samples that differ in some way from the regular observations. Those samples that are out of the distribution of normal data are called anomalies or outliers. Anomaly detection problems have been largely solved well by deep neural networks. However, there are some challenges to developing deep learning-based anomaly detection methods. In this thesis, we have touched upon some critical aspects of anomaly detection methods that have been neglected. The first challenge is generalizability to outliers, while the second challenge is deployability. To address the aforementioned challenges, we propose two different solutions in chapters three and four. The proposed methods achieve state-of-the-art results when applied to anomaly detection tasks.

Creative Commons License

Creative Commons Attribution-Noncommercial 4.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License

Download

Share

COinS