Synthetic Financial Document Generation and Fraud Detection Using Generative AI and Explainable ML

Abstract

Financial document fraud (e.g., falsified invoices or receipts) is a growing challenge requiring automated solutions. However, training data for detecting such fraud is scarce due to privacy and confidentiality concerns¹. In this paper, we propose a system that generates synthetic financial documents using generative models and uses a transformer-based classifier (LayoutLM) for fraud detection, augmented with explainable AI (XAI) techniques for interpretability. Synthetic document generation (with GANs and diffusion models) expands the training dataset while preserving privacy, enabling improved detection of anomalous or fraudulent documents. A LayoutLM-based model is fine-tuned to classify documents as genuine or fraudulent, leveraging both textual content and layout information. We integrate SHAP and LIME explainability tools to highlight the features (e.g. specific text fields or patterns) that drive each fraud prediction, satisfying the "right to explanation" and building user trust. Experiments demonstrate that augmenting training with synthetic documents substantially improves fraud recall and F1-score, while explainability techniques provide insights into the model’s decisions. We also discuss ethical implications, including data privacy, fairness (bias mitigation), and transparency. The results indicate that combining generative data augmentation with explainable deep learning offers a promising approach for financial document fraud detection.