Exploring anomaly detection methods using features extracted with a neural network

Yeom, Jaewook

Permalink

https://hdl.handle.net/2142/112952 Copy

Description

Title

Exploring anomaly detection methods using features extracted with a neural network

Author(s)

Yeom, Jaewook

Issue Date

2021-05-26

Director of Research (if dissertation) or Advisor (if thesis)

Bresler, Yoram

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Anomaly detection
• outlier detection
• neural networks
• computer vision
• information theory
• dimensionality reduction
• random projection
• coding rate

Abstract

Anomaly (or outlier) detection is a problem with the goal of detecting outliers—or anomalous samples—within a dataset or distribution. Anomaly detection has important applications in computer vision, cybersecurity, biomedical imaging, and more. As the data becomes bigger (more samples collected) and more complex (more variables observed), it becomes increasingly difficult to detect outliers within a dataset. In this thesis, we explore the problem of outlier detection in computer vision applications, in which the input images can be interpreted as vectors with hundreds of dimensions. While an effective anomaly detection method is important to achieve good performance in detecting outliers, it is also often important to first extract useful features from the data. We explore both extracting useful features from the data (using neural networks) and the use of different anomaly detection methods that use the extracted features. We first train a neural network on the training set (containing only inlier images) with the goal of image classification, and then obtain feature vectors for both the training and test sets by performing a forward pass of the training and test images through the trained network. We then use an anomaly detection method—such as nearest subspace distance or Mahalanobis distance—that makes use of the extracted features in order to obtain an anomaly score for each image in the test set, which can then be used to detect the outliers in the test set. We show that this outlier detection scheme (i.e. using an anomaly detection method on features extracted using a neural network) often outperforms the use of the maximum softmax probability (MSP) as an anomaly score. We also demonstrate that the features extracted using neural networks often lead to better anomaly detection performance than the features obtained by simply reducing the dimensions of the input images (e.g. using principal component analysis).

Graduation Semester

2021-08

Type of Resource

Thesis

Permalink

http://hdl.handle.net/2142/112952

Copyright and License Information

Copyright 2021 Jaewook Yeom

Owning Collections