Provenance of AI-Generated Images: A Vector Similarity and Blockchain-based Approach

Abstract

Rapid advancement in generative AI and large language models (LLMs) has enabled the creation of highly realistic and context-aware digital content. Multimodal systems such as ChatGPT with DALL-E integration and diffusion models like Stable Diffusion can now generate images that are often indistinguishable from human-created media, posing significant challenges for content authenticity and provenance. Ensuring the integrity and origin of digital data is critical for preserving trust, accountability, and legal compliance in digital ecosystems. This paper presents an embedding-based framework for detecting AI-generated images using vector similarity analysis. The proposed approach is based on the hypothesis that AI-generated images exhibit closer embedding proximity to other synthetic content, whereas human-created images cluster within distinct semantic regions. We validate this hypothesis by developing a prototype system that processes a diverse dataset of AI- and human-generated images using five benchmark embedding models. Furthermore we investigated whether perturbation with image, such as blurring or putting patches, would allow AI-generated images to confuse the system. Experimental results confirm the robustness of the method, showing that image perturbations have minimal impact on the embedding space, with altered images retaining high similarity to their originals.

Ähnliche Arbeiten