SignatureGuard: hybrid CNN–transformer model for signature verification and identification across Arabic and English datasets

Abstract Offline signature verification has a persistent Latin-script bias: most systems are built and evaluated on English datasets, while Arabic and other non-Latin scripts are largely absent from the benchmarking literature. SignatureGuard is a three-task framework that evaluates six hybrid CNN–transformer architectures on two offline signature benchmarks (one Arabic, ASVAR; one English, CEDAR) under a single shared preprocessing and training pipeline, enabling direct architectural comparison across writing systems. The three tasks are binary forgery detection, multi-class biometric identification, and forgery source identification. To address the Arabic data gap, we publicly released ASVAR: 3471 images (1712 genuine, 1759 forged) from 70 individuals. Hybrid pairings of EfficientNetB7 or ResNet50 with the Vision Transformer (ViT-B/16) achieve test accuracies of 98.2% and 98.4% on forgery detection, macro-F1 above 0.97, and Cohen’s $$\kappa$$ above 0.96; 95% Wilson confidence intervals ( $${\pm }$$ 1.5 pp) confirm these are not artefacts of finite test-set size. MobileNetV2-based hybrids trail by at most 0.7 percentage points. Architectural rankings are broadly consistent across both datasets. All results are obtained under a seen-writer, image-level 80/10/10 split—a closed-set protocol that supports reproducible architectural comparison but overestimates real-world deployment performance; a signer-disjoint evaluation is identified as the primary follow-on experiment. An information-theoretic argument demonstrates that the hybrid classification head cannot perform worse than either frozen component in isolation. Confusion-matrix analysis, multi-seed validation, and backbone fine-tuning are recommended as extensions.