Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Jitendra Sharma, Arthur Carvalho, Suman Bhunia

TL;DR

Generative AI enables realistic images, challenging authenticity; the paper proposes EmbedAIDetect, a training-free detector using ViT-based embeddings, a vector database, and on-chain verification to classify uploaded images as AI-generated or human-created. It evaluates multiple embedding models, demonstrates robustness to perturbations, and compares blockchain-only, vector DB-only, and hybrid frameworks, highlighting trade-offs between verifiability, accuracy, and scalability. The findings show embedding proximity is effective for non-facial content and that blockchain verification enhances trust, while facial forgery detection remains challenging and standardization with decentralized data sharing is needed for broad adoption. The work points toward a scalable, transparent AI image provenance system via future integration with IPFS and industry collaboration.

Abstract

Rapid advancement in generative AI and large language models (LLMs) has enabled the generation of highly realistic and contextually relevant digital content. LLMs such as ChatGPT with DALL-E integration and Stable Diffusion techniques can produce images that are often indistinguishable from those created by humans, which poses challenges for digital content authentication. Verifying the integrity and origin of digital data to ensure it remains unaltered and genuine is crucial to maintaining trust and legality in digital media. In this paper, we propose an embedding-based AI image detection framework that utilizes image embeddings and a vector similarity to distinguish AI-generated images from real (human-created) ones. Our methodology is built on the hypothesis that AI-generated images demonstrate closer embedding proximity to other AI-generated content, while human-created images cluster similarly within their domain. To validate this hypothesis, we developed a system that processes a diverse dataset of AI and human-generated images through five benchmark embedding models. Extensive experimentation demonstrates the robustness of our approach, and our results confirm that moderate to high perturbations minimally impact the embedding signatures, with perturbed images maintaining close similarity matches to their original versions. Our solution provides a generalizable framework for AI-generated image detection that balances accuracy with computational efficiency.

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

TL;DR

Generative AI enables realistic images, challenging authenticity; the paper proposes EmbedAIDetect, a training-free detector using ViT-based embeddings, a vector database, and on-chain verification to classify uploaded images as AI-generated or human-created. It evaluates multiple embedding models, demonstrates robustness to perturbations, and compares blockchain-only, vector DB-only, and hybrid frameworks, highlighting trade-offs between verifiability, accuracy, and scalability. The findings show embedding proximity is effective for non-facial content and that blockchain verification enhances trust, while facial forgery detection remains challenging and standardization with decentralized data sharing is needed for broad adoption. The work points toward a scalable, transparent AI image provenance system via future integration with IPFS and industry collaboration.

Abstract

Rapid advancement in generative AI and large language models (LLMs) has enabled the generation of highly realistic and contextually relevant digital content. LLMs such as ChatGPT with DALL-E integration and Stable Diffusion techniques can produce images that are often indistinguishable from those created by humans, which poses challenges for digital content authentication. Verifying the integrity and origin of digital data to ensure it remains unaltered and genuine is crucial to maintaining trust and legality in digital media. In this paper, we propose an embedding-based AI image detection framework that utilizes image embeddings and a vector similarity to distinguish AI-generated images from real (human-created) ones. Our methodology is built on the hypothesis that AI-generated images demonstrate closer embedding proximity to other AI-generated content, while human-created images cluster similarly within their domain. To validate this hypothesis, we developed a system that processes a diverse dataset of AI and human-generated images through five benchmark embedding models. Extensive experimentation demonstrates the robustness of our approach, and our results confirm that moderate to high perturbations minimally impact the embedding signatures, with perturbed images maintaining close similarity matches to their original versions. Our solution provides a generalizable framework for AI-generated image detection that balances accuracy with computational efficiency.
Paper Structure (21 sections, 3 equations, 7 figures, 5 tables)

This paper contains 21 sections, 3 equations, 7 figures, 5 tables.

Table of Contents

  1. Introduction
  2. System Design
  3. System Components
  4. Embedding Model
  5. Vector Database
  6. Blockchain Setup
  7. Web Application
  8. System Implementation
  9. Framework 1: Blockchain only
  10. Framework 2: Vector Database only
  11. Framework 3: Hybrid Approach
  12. Evaluation
  13. Vector similarity based classification
  14. Benchmarking different embedding models
  15. Smart Contract Gas Usage Estimation
  16. ...and 6 more sections

Figures (7)

  • Figure 1: System architecture for EmbedAIDetect
  • Figure 2: Application setup and data flowchart for blockchain only framework of EmbedAIDetect
  • Figure 3: Application setup and data flowchart for the vector database only framework of EmbedAIDetect
  • Figure 4: Application setup and data flowchart for the hybrid framework of EmbedAIDetect
  • Figure 5: Sample images from human-art dataset and AI-generated image dataset with various modifications
  • ...and 2 more figures