Non-Parametric Stochastic Autoencoder Model for Anomaly Detection
Date of Award
5-2022
Document Type
Dissertation
Degree Name
Doctor of Philosophy in Computer Science
First Advisor
Patricia Angela R. Abu, PhD
Abstract
Anomaly detection is a widely studied field in computer science with ap- plications ranging from intrusion detection, fraud detection, medical diagnosis and quality assurance. The underlying premise is that an anomaly is an ob- servation that does not conform to what is considered to be normal. This study addresses two major problems in the field. First, anomalies are defined in a lo- cal context, that is, being able to give quantitative measures as to how anoma- lies are categorized within its own domain problem and cannot be generalized to other domains. Currently this has been measured often using statistical probabilities of anomalies existing relative to the entire dataset with several assumptions such as distribution type and volume. Second, the performance of a model is dependent on a per problem basis. Each detection model has to have parameters properly selected to achieve acceptable performance such as the thresholds defined by domain experts or by manually adjusting the values. As such, models that attempt to solve anomaly detection are highly special- ized or engineered on a specific problem. To address these, the study explores the creation of a model whose parameters, is applied in an adaptive manner through the means of a neural network based autoencoder model. The study also contributes to a quantiative description of anomaly detection datasets and provide insights as to how different models behave in different situations.
Recommended Citation
Raphael, Alampay B., (2022). Non-Parametric Stochastic Autoencoder Model for Anomaly Detection. Archīum.ATENEO.
https://archium.ateneo.edu/theses-dissertations/744
