AI-driven forgery detection in offline handwriting signatures: advances, challenges, and the role of Generative Adversarial Networks

Handwriting-based authentication continues to be a critical element in forensic analysis, particularly in the context of document fraud and signature forgery. Although deep learning (DL) techniques have shown promising results, there are still obstacles associated with the availability of limited datasets, the generalization of models, and their robustness. This review conducts a systematic examination of recent developments in DL methods for signature forgery detection. It employs the PRISMA protocol and retrieves literature from four well-established databases: Scopus, ACM Digital Library, Web of Science, and IEEE Xplore. Following a rigorous screening procedure, a total of 15 primary studies published between 2019 and 2025 were selected from an initial 115 records that were filtered by Computer Science subject area, English language, and original research articles. Five publicly accessible datasets: CEDAR, BHSig260, ICDAR 2011 SigComp, Kaggle signature verification dataset by RobinReni, and Kaggle handwritten signatures by Divyansh Rai were identified and analysed. The review indicates that Siamese networks dominate the DL architecture for signature forgery detection tasks, while alternative methods either employed fine-tuned pre-trained models (i.e., VGG16) or a hybrid of autoencoders and Convolutional Neural Networks (CNNs). An accuracy of 100% has been achieved through utilization of Siamese network leveraging the CEDAR dataset. This result is reasonable since CEDAR has the advantages of clean and balanced dataset. In response to the persisting limitations, this review emphasizes Generative Adversarial Networks (GANs) as the powerful data augmentation technique and a potential solution to enrich training datasets, simulate diverse forgery patterns, and enhance the robustness of models. Finally, a generative-aware conceptual framework is proposed at the end of the review to inform future research on the development of offline handwriting signature forgery detection system that is more resilient and forensic-ready.