Reproduce “Layout-Aware Representation Learning for Open-Set ID Fraud Discovery” on OpenResearch

About this paper

Jinxing Li, Nicholas Ren, Cathy Chang, Hongkai Pan, Daniel George

Persona Identities researchers developed a layout-aware representation learning model for open-set ID fraud discovery, leveraging self-supervised and metric learning to effectively identify novel fraud campaigns. The system demonstrated robust generalization across diverse ID layouts and uncovered previously undetected large-scale fraud operations in production traffic.

Key insights

• Self-supervised fine-tuning of a foundation model on document images is crucial for capturing rich, layout-specific contextual and spatial representations before supervised metric learning.
• Combining multiple deep metric learning losses effectively creates a highly structured embedding space with strong intra-class compactness and inter-class separability, superior to individual losses.
• The learned visual embeddings enable discovery of 'off-manifold' fraud clusters and similarity-based expansion from fraud seeds, uncovering campaigns undetectable by conventional metadata.

Summary by alphaXiv.