Credit Card Fraud Detection using ML & Deep Learning

Overview

Credit card fraud causes billions in annual losses globally. This project performs a comparative analysis of machine learning and deep learning approaches for detecting fraudulent transactions in highly imbalanced datasets — a core challenge in financial fraud detection.

Problem

• Fraud transactions are rare (< 0.2% of all transactions) — severe class imbalance
• Real-time detection requires high accuracy and low latency
• False positives (legitimate transactions blocked) carry high customer cost

Methodology

• Class imbalance handling — SMOTE, undersampling, cost-sensitive learning
• Traditional ML — Logistic Regression, Decision Trees, Random Forest, SVM, k-NN
• Deep learning — Artificial Neural Networks, Autoencoders for anomaly detection
• Evaluation — precision, recall, F1-score, AUC-ROC (accuracy alone is misleading on imbalanced data)

Key Findings

• Ensemble methods (Random Forest) outperformed single classifiers on imbalanced data
• Deep learning autoencoders effective for anomaly-based unsupervised fraud detection
• SMOTE combined with ensemble methods gave best precision-recall tradeoff

Publications

(Gorte et al., 2022)

(Gorte et al., 2023)

Status

Published — Springer ICICC 2022 and AIP Publishing 2023.