Toward Real-world Text Image Forgery Localization: Structured and Interpretable Data Synthesis

In the rapidly evolving field of digital forensics, the detection of manipulated text images remains a significant challenge. Existing datasets often focus on visible artifacts, neglecting the complex, invisible parameters underlying real-world forgery behaviors. This limits the ability of deep learning models to generalize effectively across diverse scenarios. To address this, the authors introduce Fourier series-based Tampering Synthesis (FSTS), a novel framework that models the hidden distribution of tampering parameters by analyzing extensive real-world operation logs. By decomposing these parameters into interpretable basis functions, the method captures recurring behavioral patterns at both individual and population levels. This hierarchical frequency domain model enables the generation of highly diverse and realistic tampered images, which serve as superior training data. Extensive experiments across multiple benchmarks demonstrate that models trained with FSTS data outperform traditional synthetic datasets, achieving over 15% improvement in F1 scores and 10% in AUC on real-world datasets. The approach effectively narrows the distribution gap, significantly enhancing the robustness and generalization of forgery localization systems. This work marks a substantial step toward practical, scalable solutions for digital document security, with promising avenues for future multi-modal behavior modeling and automated data collection. Despite its strengths, the method faces challenges in scalability and modeling highly nonlinear behaviors, suggesting directions for ongoing research and refinement.