Reconstructing Fraudulent Strategies in Online Payments via Interpretable Inverse Reinforcement Learning

| Source: arXiv AI

Tags: PaySeqGen, Inverse Reinforcement Learning, Fraud Detection, Ant International, Data Augmentation

PaySeqGen uses interpretable inverse reinforcement learning to reconstruct fraudster behavior from payment trajectories, outperforming generative augmentation baselines in tests at Ant International.

Details

Fraud detectors miss complex behaviors, and augmenting training data with synthetic fraud helps only if the generated sequences keep the links between transaction choices, payment feedback and the fraudster's next move. PaySeqGen learns fraud behavior policies from observed fraudster payment trajectories with inverse reinforcement learning. It builds candidate actions for plausible sequence generation and learns interpretable rewards that show fraudster preferences across transaction contexts, alternating reward learning and policy optimization. In real payment scenarios at Ant International, it beat state-of-the-art generative augmentation methods, and reward analysis helped analysts locate suspicious segments. The authors report evaluation in a deployed online payment system, though the abstract gives no numbers.