Synthetic Handwritten Signature Generation Using Deep Convolutional GANs: A Security Evaluation Framework

Handwritten signatures remain a widely used biometric modality for authentication and forensic verification, yet the security implications of generative adversarial network (GAN)-based signature synthesis have received limited direct evaluation compared to other biometric modalities such as faces and fingerprints. This paper presents a generate-then-detect security evaluation framework applied to handwritten signatures. A Deep Convolutional GAN (DCGAN) was trained on the CEDAR genuine signature corpus under free-tier cloud GPU constraints, following an initial, ultimately infeasible attempt using StyleGAN2-ADA. The resulting synthetic signatures exhibited recognizable stroke-level structure but showed evidence of partial mode collapse, yielding a Fréchet Inception Distance (FID) of 248.95 and a mean Structural Similarity Index (SSIM) of 0.512 (± 0.077) relative to genuine samples. A ResNet-18 classifier fine-tuned to distinguish genuine from synthetic signatures achieved perfect test-set performance (Accuracy, F1-score, and AUC-ROC of 1.0000; Equal Error Rate of 0.0000), a result attributed to the limited diversity and fidelity of the generated samples rather than an inherently unbreakable detection boundary. These findings provide, to the authors’ knowledge, the first generate-then-detect security evaluation conducted specifically on the handwritten signature modality, and offer a practical methodological note on generative architecture selection under resource-constrained research conditions.