Abnormality Detection Using Variational Autoencoder: An Application in Banking System

This internship project focuses on Variational Autoencoders (VAEs) for the precise identification of fraudulent activities within the banking sector, notably on check fraud detection. The literature shows that VAEs are very powerful in extracting principal features and components of a given dataset. This project meticulously outlines a comprehensive methodology where VAE is utilized for analyzing check images and extracting critical features such as signature styles, printed patterns, and unique identifiers to transform complex, high-dimensional data into an encoded, simplified representation. This project proposes an innovative integration of machine learning techniques to address challenges in financial fraud detection. By utilizing VAEs, we aim to extract and decode complex patterns, optimizing the detection of anomalies in the banking system. The methodology involves the application of VAEs for feature extraction from high-dimensional data, subsequently clustering these features to identify irregularities and predict long-term performance. The project leverages publicly available datasets, such as RVL-CDIP (Ryerson Vision Lab Complex Document Information Processing) and NIST Special Database 19, for checks to refine models capable of handling diverse scenarios.

Faculty Supervisor:

Jahrul Alam;Kevin Pope

Student:

Partner:

NASDAQ Canada Inc

Discipline:

Mathematics

Sector:

Finance and Insurance; Artificial Intelligence

University:

Memorial University of Newfoundland

Program: