Intelligent Artwork Authentication Using Transfer Learning and Explainable AI in Cultural Heritage Systems

Abstract

Museums and heritage institutions continue to face difficulties in reliably authenticating cultural artifacts. Artificial intelligence can help, but curators need precise and trustworthy tools. The authors propose a deep learning framework that combines vision transformers, gradient-weighted class activation mapping, and SHapley additive exPlanations as an explainable methodology to support artwork authentication in real-world museum environments. The authors fine-tune a pre-trained vision transformer model on the museum identification challenge dataset, which comprises more than 11,000 images across four artwork categories. Post hoc explanations of model decisions were obtained by merging gradient weighted class activation mapping, SHapley additive exPlanations, and cultural heritage. The proposed model achieves 91.4% accuracy and is robust to occlusions and lighting distortions. Attribution maps were aligned with regions identified by the experts in 85% of cases. The results indicate that explainable deep learning is significant in high-stakes, interdisciplinary applications, such as cultural heritage verification.