KRYP71CGH057/AUTO-ENCODER-MODEL

Development of an Auto-encoder Model for Predicting Financial Crimes in Nigeria Banking Sector # Development of an Auto-encoder Model for Predicting Financial Crimes in Nigeria Banking Sector **Author:** Henry Isaac Owoche **For:** Mbang Blessing **Department:** Computer Science, University of Calabar **Degree:** Bachelor of Science (B.Sc.) in Computer Science **Academic Session:** 2025/2026 --- ## Table of Contents 1. Project Overview 2. Research Motivation 3. Research Objectives 4. Dataset 5. System Architecture 6. Project Structure 7. Installation 8. Usage 9. Results 10. Output Files 11. Limitations 12. References --- ## 1. Project Overview This project develops and evaluates an **auto-encoder-based anomaly detection system** for predicting financial crimes in the Nigerian banking sector. The system trains an auto-encoder neural network exclusively on legitimate (non-fraudulent) banking transactions using **unsupervised learning**, then identifies fraudulent transactions as those whose reconstruction error exceeds a statistically derived threshold. The core advantage of this approach is that it requires **no labeled fraud examples during training** — a critical practical benefit in the Nigerian banking context, where labeled fraud datasets are scarce, sensitive, and difficult to obtain. > *"Nigerian banks lost approximately ₦12.2 billion to electronic payment fraud in the first half of 2023 alone."* > — Amuge (2023), Business AM Live --- ## 2. Research Motivation The Nigerian banking sector faces an escalating financial fraud problem driven by three compounding factors: - **Rapid digitisation** of banking services creating new attack vectors - **Regulatory gaps** in fintech oversight that fraudsters exploit (Enwe, 2021) - **Inadequacy of existing rule-based systems** that cannot adapt to evolving fraud tactics Traditional supervised machine learning approaches are further constrained by: - Severe **class imbalance** — fraudulent transactions represent only ~0.13% of all transactions - **Scarcity of labeled fraud data** due to privacy concer …