Exploring techniques to distinguish between real images and those generated using stable diffusion XL.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Text-to-image diffusion models generate realistic synthetic images.
• Concerns exist regarding malicious use of synthetic images.
• Distinguishing AI-generated images from real ones is crucial.

Purpose of the Study:

• To develop and evaluate a method for detecting synthetic images generated by diffusion models.
• To assess the feasibility of discerning between authentic and AI-generated images.
• To contribute a large, publicly accessible dataset of real and synthetic images.

Main Methods:

• A novel convolutional neural network (CNN) was designed and implemented.
• The CNN was trained and tested on a bespoke dataset.
• The dataset comprised authentic ImageNet images and synthetic images from Stable Diffusion XL.
• A ResNet-18 baseline model was also used for comparison.

Main Results:

• The proposed CNN achieved up to 98.38% accuracy in binary classification.
• The ResNet-18 baseline achieved 97.24% accuracy.
• Experiments demonstrated the effectiveness of the developed method in detecting synthetic images.
• The study released the largest publicly available dataset of Stable Diffusion XL images.

Conclusions:

• It is possible to effectively detect synthetic images generated by diffusion models.
• The developed CNN and dataset contribute significantly to research in AI image detection.
• High accuracy in distinguishing real from synthetic images was achieved.