A Systematic Ablation-Based Optimisation Protocol for Convolutional Neural Networks in Electronic Banking Fraud Detection

Abstract

Reliable deep learning for electronic banking fraud detection remains constrained by extreme class imbalance. Standard evaluation methods frequently struggle because convolutional network choices are often assessed alongside simultaneous methodological modifications, making it difficult to determine the independent contribution of architectural design decisions. To address this problem, this study introduces the Systematic Ablation-Based Optimisation Protocol (SAOP) as a controlled evaluation framework. The protocol examines predefined CNN variations in convolutional depth and dropout regularisation under identical preprocessing, resampling, training, and evaluation conditions. Empirical testing relied on the Credit Card Fraud Detection benchmark dataset, which contains 284,807 transactions with an explicit fraud prevalence of 0.172%. Within this heavily skewed search space, internal ablation results indicate that dropout regularisation variations exerted a more measurable influence on performance than changes in convolutional depth. The proposed CNN was benchmarked against traditional machine learning models using stratified five-fold cross-validation. This architecture achieved a mean accuracy of 0.9988 ± 0.0003, a Matthews Correlation Coefficient (MCC) of 0.7081, a Receiver Operating Characteristic–Area Under the Curve (ROC-AUC) of 0.9659, and a Log Loss of 0.0104. The CNN recorded both the highest MCC and the lowest Log Loss among all evaluated models, showing stable minority-class discrimination alongside reliable probability calibration. These results indicate that architectural design choices remain an important consideration in CNN-based fraud detection under severe class imbalance.