Research on Detecting AI-Generated Forged Handwritten Signatures via Data-Efficient Image Transformers

Handwritten signature verification, a behavioral physiological biometric, is widely accepted in legal and social practices. However, the advent of artificial intelligence (AI) and its application in handwriting synthesis, known as AI-generated content (AIGC), has introduced a new security threat: AI-generated forged handwritten signatures. To address this challenge, a series of AI-generated forged handwritten signatures was constructed using various font-creation software. A model based on a pre-trained Data-Efficient Image Transformer (DeiT) was proposed to identify forged handwritten signatures. The results show that for AI-generated forged handwritten signatures constructed using Huawei and Baidu font software, the DeiT model achieved a high identification accuracy of nearly 100%. Signatures generated by the Shouji font software also showed a high recognition rate (97.76%). The overall accuracy across all the datasets was 98.46%. Hence, in our experimental states, the DeiT model shows better performance in forged signature identification than other models, including the Vision Transformer (ViT), VGG16, ResNet, Convolutional Neural Networks (CNN), and XGBoost. The DeiT model demonstrated superior performance with AUC values of 100% for the Huawei and Baidu datasets, 99.99% for the combined Huawei and Baidu datasets, 100% for the AI & Handwriting dataset, and 99.72% for the Shouji dataset. These results strongly prove the status of the DeiT model as a state-of-the-art (SOTA) solution for the identification of AI-generated handwritten signature verification.