Anomaly-Based Financial Fraud Detection Using Autoencoder A Case Study on the Kaggle Credit Card Dataset

Financial fraud remains a critical challenge for banking systems and digital payment platforms worldwide. With the rapid growth of electronic transactions, effective fraud detection mechanisms are essential to ensure security and user trust. This study explores the application of an unsupervised deep learning model—Autoencoder—for anomaly-based financial fraud detection. Utilizing the publicly available Kaggle Credit Card Fraud Detection dataset, which comprises 284,807 transactions including 492 fraudulent cases, the model is trained exclusively on legitimate transactions to learn typical behavioral patterns. Prior to training, the dataset underwent feature anonymization using Principal Component Analysis (PCA), and numerical columns such as "Amount" and "Time" were normalized using Min-Max Scaling. The Autoencoder architecture includes three encoder and decoder layers with ReLU activations, and is optimized using the Adam optimizer with Mean Squared Error (MSE) as the loss function. Experimental results show that the model achieves a classification accuracy of 94% and an AUC score of 0.931, indicating strong potential for detecting anomalies. However, the precision for identifying fraudulent transactions remains relatively low (5%), reflecting the challenges posed by imbalanced datasets. Despite this, the study demonstrates that Autoencoder offers a promising foundation for fraud detection systems, with further improvements possible through model integration and hybrid ensemble techniques