New AI model beats traditional systems in detecting fraudulent credit card transactions

In a major advancement in financial cybersecurity, researchers have developed a novel hybrid deep learning model that significantly improves the detection of fraudulent credit card transactions. The study, titled “Enhanced Credit Card Fraud Detection Using Deep Hybrid CLST Model” and published in Mathematics in June 2025, introduces a new approach combining Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and fully connected output layers. This architecture, called CLST, demonstrated exceptional performance on highly imbalanced datasets, tackling one of the most persistent challenges in fraud detection.

How can deep learning be better applied to identify fraud?

At study employs a hybrid methodology that leverages the complementary strengths of CNN and LSTM. The CNN component captures spatial patterns within transaction features, while LSTM models sequential dependencies, crucial for understanding time-based transaction behaviors. These two models feed into a dense output layer that synthesizes insights for final classification.

The dataset used was a widely known credit card fraud detection benchmark obtained from Kaggle. It is notoriously imbalanced, with only 0.172% of the transactions being fraudulent. To mitigate this, the researchers utilized the Synthetic Minority Oversampling Technique (SMOTE), which generates synthetic samples for the minority class, thereby balancing the dataset and improving model robustness.

Performance enhancements didn’t stop at architecture design. Hyperparameter tuning via Keras Tuner was applied to refine learning rate, batch size, number of layers, dropout rate, and more. These optimizations pushed the model’s precision for fraud detection from 70% to 83%, and recall from 83% to a near-perfect 99%. The model’s ROC-AUC score jumped from 0.9733 to an outstanding 0.9995, indicating near-flawless discrimination between fraudulent and legitimate transactions.

How does the proposed CLST model compare to traditional algorithms?

The researchers compared the CLST model against conventional machine learning methods like Logistic Regression (LR) and Support Vector Machines (SVM). While LR and SVM showed high accuracy for non-fraudulent transactions (with recall rates of 97% and 98%, respectively), they performed poorly in identifying fraud. LR’s precision for fraudulent cases was just 6%, and SVM’s was only 9%, resulting in a flood of false positives and unreliable results.

In contrast, the optimized CLST model delivered balanced and superior results across all key metrics:

• Precision (fraud): 83%
• Recall (fraud): 99%
• F1-score (fraud): 91%
• Overall accuracy: 99.98%
• ROC-AUC: 0.9995

These improvements are attributed not only to the hybrid architecture but also to SMOTE’s class balancing and the careful calibration of hyperparameters. The model’s robustness was further validated through ten-fold cross-validation, yielding an average accuracy of 99.89% with minimal performance variance across folds.

Moreover, ablation studies confirmed that each component of the architecture contributed significantly to performance. Removing the fusion layer or using only CNN or LSTM individually resulted in substantial drops in precision, recall, and AUC. This underlined the importance of feature fusion and integrated learning across spatial and temporal dimensions.

Can this model be applied in real-time financial environments?

The CLST model was designed not only for accuracy but also for real-world feasibility. Its computational efficiency was thoroughly tested, showing strong scalability for deployment in financial systems. Using an NVIDIA RTX 3080 GPU, the model was able to process over one million transactions per second, making it suitable for real-time fraud detection in high-frequency trading environments or digital banking platforms.

Training the model across 20 epochs required only six minutes, demonstrating that the proposed approach does not compromise speed for accuracy. This level of efficiency positions CLST as a viable solution for financial institutions seeking robust, scalable, and fast fraud detection systems.

The study also situates CLST within the broader landscape of recent advances. Compared to other state-of-the-art models, such as autoencoder-based frameworks or attention-based LSTMs, CLST outperforms in both recall and AUC-ROC metrics. For instance, a 2024 study using attention-based LSTM achieved 97.22% recall and 99.3% AUC-ROC, while CLST attained 99% recall and 99.95% AUC-ROC.

Future directions identified by the authors include extending the CLST model to detect other types of financial fraud, such as insurance fraud or identity theft. There’s also interest in adapting the model for streaming data environments and incorporating federated learning to maintain data privacy across decentralized systems. Additionally, the researchers emphasize the need for explainable AI techniques to build trust and transparency in real-world applications.