Synthetic Adversary Modeling via In-Silico Transaction Experimentation: A Multimodal Foundation-Model Framework with Source-Dropout and Universal-Consumer Baselines

Key Points

• The aim is to create a framework that improves payment fraud detection through synthetic data generation.
• In-silico transaction experimentation framework created based on multimodal transaction data.
• Synthetic fraud candidates generated using source dropout and a universal-consumer baseline.
• Experiments conducted on the public ULB credit-card benchmark to assess model performance.
• Directional gains observed in fraud detection efficacy compared to traditional methods.
• Framework shows potential for future validation with real fraud data.
• Improvements highlight need for further exploration in various transaction contexts.