Image Authentication: Differentiating Camera-Captured and AI-Generated Images with Provenance Verification using Vision Transformer (ViT)

Abstract

The rapid growth of generative artificial intelligence significantly affects the development of digital image creation. Recently, advances in Generative Adversarial Networks (GAN) and Diffusion-based Generative Networks have made it possible to create synthetic images that are hard to tell apart from real ones. Therefore, verifying images produced by artificial intelligence is essential. This poses a major challenge for many fields, including digital forensics, security, and image verification. There has been a rise in the misuse of AI-generated images to spread false news, impersonate people, and manipulate images. Traditional methods for verifying image authenticity, such as human observation and image metadata analysis, are now unreliable. AI-generated images can be easily altered, and human observation alone cannot confirm an image's authenticity. As a result, there is a pressing need to develop an effective image authentication system to tell apart real images from those created by AI.
This paper proposes an automated image authentication system. It utilizes a Vision Transformer model for classifying images as either camera-captured or AI-generated. The system employs a pre-trained model for feature extraction, which is then fine-tuned for classification. Unlike conventional convolutional neural networks, the Vision Transformer treats an image as a sequence of patches and uses self-attention to capture global dependencies. This method helps to identify subtle differences in AI-generated images. Additionally, the proposed system incorporates a confidence level represented by Softmax probabilities, which helps understand the reliability of the system's results. An explainability feature is also included, using Explainable Artificial Intelligence techniques to highlight areas in the image that influence the results. This system provides a strong solution for the current challenges in image authentication. It can be implemented as a web-based application using the Flask framework. Experimental results demonstrate that the system achieves high accuracy in classifying AI-generated images.