Verification Challenges of AI-Generated Identity Documents: Blockchain Technology as a Trust Layer and AI as a Supporting Signal

The rapid advancement of generative AI has revolutionized digitalization, while simultaneously introducing new security challenges. As AI models are increasingly integrated into cameras to enhance or modify images, a fundamental question arises for verification systems: whether captured images retain authentic camera fingerprints, such as Photo-Response Non-Uniformity (PRNU), Color Filter Array (CFA) patterns, physically random sensor noise, and lens distortions, or are heavily altered or fully generated by AI. Modern generative AI models create images that are highly similar to those produced by cameras, increasing the risk of document forgery and verification challenges. To address these challenges, this research proposes blockchain technology as a foundational trust layer for digital identity, enabling secure and tamper-proof evidence recording through an immutable ledger and cryptographic mechanisms. The proposed system integrates blockchain with a layered microservices architecture, separating user management, blockchain interaction, and audit logging into independent services. Communication between services uses gRPC with clearly defined Protocol Buffer schemas for efficient communication. The API layer is implemented using FastAPI for authentication, authorization, and request routing with high performance and automatic documentation. Data is stored in MongoDB, including user profiles, authentication records, verification results, and audit logs, which ensures flexibility and high availability. AI is used as a supporting signal rather than a definitive decision-maker. Experimental evaluation was conducted on 4,550 handwritten signatures, created using real ink pens but not belonging to any specific individual, and 4,550 AI-generated signatures were created using OpenAI's GPT image models, Nano Banana 2, and Qwen image generation models. ResNet50 was used to compute the signal score and achieved an F1 score of 0.996 on the classification task. The proposed method is designed to generalize well across a wide range of document and image domains.