Publication

Abstract : Fighting fraud presents a huge problem in financial services, it threatens the trust and security of every party involved. Summary this study introduces a new approach for detecting fraud related transactions by combining synthetic data generation and deep learning. Our design uses a simple feed forward neural network that works well for the classification of transactions based on regime, allowing us to take in features such as amount, time and user/merchant identity into account. To improve inter-pretability of our model, we introduce Shapley Additive explanation (SHAP) analysis that shows how different features impact predictions of fraud and makes the decision-making process more transparent. The synthetic data set used in this study is designed to simulate possible real-life transaction scenarios, thus covering many already described features of the model that needs to be trained and evaluated without prejudice. Results show our proposed method provides a considerable accuracy for fraud identification compared to the existing methods of detection. Additionally, we optimize our SHAP computations by only sampling background data to compute feature importance, saving us that computational overhead while maintaining important color commentary. These results demonstrate the promise of integrating deep learning and explainable AI for fraud detection in robust financial systems. Fixing this issue not only enhances detection, but responding to an increasingly fundamental requirement for transparency in our anonymizing schemes ultimately builds stronger security and trust into financial transactions.