Adaptive Fraud Detection in Digital Banking Using Deep Learning: A Hybrid CNN-RNN Approach

Financial fraud in digital banking has evolved to a point that is quite complex, meaning that there are greater needs for a more effective and efficient system for detection of financial fraud. The traditional rule-based system used to detect fraudulent use of financial instruments such as credit cards and banking systems experience quite high rates of false positive. Also, relatively low rate of effectiveness in detecting the fast-evolving fraud pattern. The result, of course, is that financial losses to both the institutions that maintain such systems and the clients themselves. This paper presents a model for detecting fraud in a digital banking system designed on deep learning approaches including a sequence of recurrent neural networks and a variety of convolutional neural networks for learning patterns on real-time transactions data. The model generates highly accurate values of 95.2%, as well as precisions and recalls of highly accurate values of 91.0% and 89.3%, respectively. Furthermore, the model generates better results in comparison to the systems used in existing applications, which produce the average accuracy of about %. The processing time of the model is relatively short, which takes about 15 ms, and in the meantime, the model generates low levels of false positives and relatively high rates of real fraud detection, which are assessed at 65%. Overall, the research topic presents an efficient and effective solution for fraud detection, which is highly adaptive and scalable to detect fraud in an environment where no weaknesses in the rule-based legacy systems are experienced.