SS-ANFIS: a semi-supervised neuro-fuzzy model for offline signature verification

Abstract

Signature verification remains a critical authentication mechanism in academic and administrative environments, yet manual verification is vulnerable to forgery and subjective judgment. This study proposes SS- ANFIS, a semi-supervised neuro-fuzzy model for offline signature verification under limited labeled data conditions. The proposed model integrates pseudo-label-based self-training into a Takagi-Sugeno-Kang adaptive neuro-fuzzy inference system (ANFIS). Static image-based features were extracted from offline signature images and transformed using principal component analysis (PCA) before classification. Experiments were conducted on 800 offline signature samples collected from Dipa University Makassar, consisting of 400 genuine and 400 forged signatures. The proposed model achieved an accuracy of 90.5%, precision of 98.8%, recall of 82.0%, and F1-score of 90.0%. The high precision indicates that SS- ANFIS is effective in minimizing false positive predictions, which is important for academic document verification. The results show that the proposed model provides a practical, interpretable, and computationally efficient approach for offline signature verification, particularly in institutional settings with limited labeled data.