Layout-Aware Representation Learning for Open-Set ID Fraud Discovery

A team of researchers (Jinxing Li, Nicholas Ren, Cathy Chang, Hongkai Pan, and Daniel George) has introduced a novel approach to identity-document fraud detection that moves beyond traditional binary classification. Their paper, published on arXiv, tackles the challenge of adaptive attackers who constantly modify templates and fabrication pipelines, making historical fraud labels stale. The solution is Layout-Aware Representation Learning for open-set fraud discovery, which uses a DINOv3 backbone fine-tuned with context-aware SimMIM (masked image modeling) and supervised metric learning with a composite loss function.

Trained exclusively on U.S. IDs, the model produces layout-aware document embeddings that generalize remarkably well. Using a lightweight MLP and softmax classifier, it achieves 99.83% layout classification accuracy on Canadian IDs. On a dataset of 20,448 Canadian IDs, embedding-space analysis uncovered 276 adaptive physical-fraud cases, 222 of which were not surfaced by incumbent detectors. The embeddings also support similarity-based expansion from a single confirmed seed to related cases not linked by conventional metadata graphs, making this approach production-aligned for discovering novel and campaign-scale fraud under distribution shift.