AnoDoc: Robust Anomaly-Based Document Forgery Detection

Abstract

Forged official documents threaten institutional trust, especially in regions lacking robust digital verification systems. While supervised methods depend on large labeled datasets and struggle with unseen manipulations, the unsupervised Doc-Patch approach proved the feasibility of detecting anomalies from authentic samples only. Building on this foundation, we propose AnoDoc, a robust anomaly-based framework that enhances unsupervised document forgery detection through adaptive self-attention, document-specific data augmentation, Focal Loss, and dynamic thresholding for improved adaptability and precision. The integration of t-distributed Stochastic Neighbor Embedding (t-SNE) further supports interpretability by visualizing the separation between authentic and forged embeddings. Experiments on a real-world dataset show that AnoDoc achieves 98% F1-score and 0.987 Area Under the Receiver Operating Characteristic (AUROC), surpassing Doc-Patch by over 6% in precision and generalization. These results demonstrate its effectiveness and scalability for deployment in administrative, legal, and financial authentication workflows.