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

Payment fraud detection is reactive because supervised models learn from previously observed fraud. This preprint proposes an in-silico transaction experimentation framework that uses a multimodal transaction foundation model to generate synthetic fraud candidates through source dropout, a universal-consumer baseline, and divergence-graded selection. Experiments on the public ULB credit-card benchmark show directional gains and motivate future validation on real source-grouped fraud data.