Countering the Evolution of Forgery: Detecting Generative Handwriting Synthesis via Hybrid Architecture and Boundary-Aware Perturbation Learning

Abstract

AI-generated handwriting poses a growing threat to cybersecurity. It enables anonymous synthetic scripts to bypass existing verification systems, potentially compromising legal, financial, and administrative workflows. However, detection remains challenging due to the extreme information sparsity and localized sub-pixel stroke artifacts of synthetic handwriting. Existing frameworks often rely heavily on global image features, which leads to high computational overhead and poor robustness against noise or adversarial perturbations when processing sparse handwriting data. To fill this research gap, we propose a Texture-semantic Discriminator (TSD) that integrates hybrid CNN-Transformer features for classification, while extracting tri-scale intermediate embeddings to facilitate regularized adversarial augmentation. To further enhance robustness, we propose the Boundary-Aware Deformation Network (BAD-Net) to adaptively perturb these representations toward the decision boundary, forcing the model to distill more invariant and discriminative features. In addition, we construct a comprehensive benchmark dataset containing over 110,000 word-level images generated by 13 state-of-the-art synthesis models to facilitate rigorous evaluation. Evaluations against six state-of-the-art methods across public and proposed datasets demonstrate that our framework achieves SOTA performance in both inference speed and detection accuracy, particularly excelling in robustness on challenging samples. This establishes a robust and real-time solution for authenticating digitized handwriting in the generative AI era.