Non-Parametric Stochastic Autoencoder Model for Anomaly Detection

• Main Authors: Alampay, Raphael B, Abu, Patricia Angela R
• Format: text
• Published: Archīum Ateneo 2022
• Subjects: Neural networks; autoencoders; machine learning; anomaly detection; semi-supervised learning; Computer Sciences; Databases and Information Systems; Electrical and Computer Engineering; Physical Sciences and Mathematics
• Online Access: https://archium.ateneo.edu/discs-faculty-pubs/342; https://archium.ateneo.edu/cgi/viewcontent.cgi?article=1342&context=discs-faculty-pubs
• Institution: Ateneo De Manila University

Anomaly detection is a widely studied field in computer science with applications ranging from intrusion detection, fraud detection, medical diagnosis and quality assurance in manufacturing. The underlying premise is that an anomaly is an observation 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 local context, that is, being able to give quantitative measures as to how anomalies are categorized within its own problem domain and cannot be generalized to other domains. Commonly, anomalies are measured according to statistical probabilities relative to the entire dataset with several assumptions such as type of distribution and volume. Second, the performance of a model is dependent on the problem itself. As a machine learning problem, each model has to have parameters optimized to achieve acceptable performance specifically thresholds that are either defined by domain experts of manually adjusted. This study attempts to address these problems by providing a contextual approach to measuring anomaly detection datasets themselves through a quantitative approach called categorical measures that provides constraints to the problem of anomaly detection and proposes a robust model based on autoencoder neural networks whose parameters are dynamically adjusted in order to avoid parameter tweaking on the inferencing stage. Empirically, the study has conducted a relatively exhaustive experiment against existing and state of the art anomaly detection models in a semi-supervised learning approach where the assumption is that only normal data is available to provide insight as to how well the model performs under certain quantifiable anomaly detection scenarios.