BANKING FRAUD DETECTION USING ARTIFICIAL NEURAL NETWORK (ANN)

Banking fraud causes direct financial loss, erodes customer trust and creates regulatory exposure, and its detection is made difficult by the extreme rarity of fraudulent transactions and by the constantly changing behaviour of fraudsters. Rule-based systems and shallow classifiers struggle to capture the non-linear interactions among transaction attributes. This paper presents a fraud detection framework built around a feedforward Artificial Neural Network (ANN) with three hidden layers (64–32–16 neurons, ReLU activation) and a sigmoid output unit that estimates the probability that a transaction is fraudulent. The framework combines z-score standardisation, stratified 70/15/15 data splitting, Synthetic Minority Over-sampling Technique (SMOTE) applied to the training partition only, Adam-based optimisation with L2 regularisation, early stopping on a validation set, and a tunable decision threshold. The model is evaluated on a 100,000-record, 30- feature synthetic transaction dataset with a fraud rate of 0.89%, and compared with Logistic Regression, Decision Tree and Random Forest baselines. On the held-out test set the ANN reaches an AUC-ROC of 0.958 (the highest among all models), an accuracy of 99.57% and a recall of 0.820, and raising the decision threshold to 0.9 lifts its F1-score from 0.773 to 0.839. An ablation over network depth, a SMOTE on/off comparison and a threshold-sensitivity analysis are reported, together with an honest discussion of limitations: results come from synthetic data, a single split and 133 test-set fraud cases, so they must be confirmed on real bank data before deployment.