Deep Learning for Fraud Detection: A CNN-SVM Model with Temporal Feature Optimization and Statistical Insights

As digital financial transactions expand at a rapid pace the detection of fraudulent activities has become vitally important due to advanced internet-based criminal activities. Researchers developed a superior fraud detection framework combining Sequential Feature Analysis (SFA) with Convolutional Neural Network (CNN) and Support Vector Machine (SVM) to achieve better performance results. The proposed model uses SFA for temporal pattern detection and CNN together with SVM for extracting spatial features to classify transactions between fraudulent and legitimate categories. The proposed model exhibits better performance than conventional methods when assessing the publicly available credit card fraud data. The CNN-SVM combination displayed an accuracy level of 97.1% which exceeded both logistic regression (91.5%) and random forest (94.3%). The detection abilities of the model were confirmed by its precision rate of 95.6% together with recall of 94.3% and F1-score of 94.9%. The model proved its ability to detect fraudulent patterns through the ROC curve analysis which delivered an AUC value of 0.98. The training and validation accuracy increased from 60% to 98% and 58% to 95% while the loss decreased from 1.2 to 0.2 and 1.3 to 0.3 throughout 20 epochs. The best detection outcomes emerged from using 0.001 as the learning rate value within a 3-layer CNN architecture structure because suitable hyperparameter choice matters. The proposed model proves reliable in both performance and scalability for running real-time fraud detection systems.