Transfer Learning-Based Detection of AI-Generated Image
TL;DR
This study investigates the automatic classification of real and AI-generated flower images using fine-tuned transfer learning models and shows that Swin Transformer-Tiny achieved the best overall performance, reaching an F1-score of 88.64% and outperforming the other architectures.
Abstract
Recent advances in artificial intelligence–based image generation have significantly increased the number of AI-generated images that closely resemble real visual content. This rapid growth has introduced new challenges for visual authenticity verification, particularly in natural object domains characterized by highly complex textures and subtle structural variations. Flower imagery represents a particularly challenging benchmark setting, where fine-grained petal textures, delicate color transitions, and complex organic forms can be realistically mimicked by generative models. To address this problem, this study investigates the automatic classification of real and AI-generated flower images using fine-tuned transfer learning models. For this purpose, a novel flower image dataset containing both real and DALL·E-generated synthetic samples was constructed and used as a benchmark for comparative evaluation. Four widely used transfer learning architectures, EfficientNet-B2, EfficientNet-B3, ConvNeXt-Tiny, and Swin Transformer-Tiny, were fine-tuned and evaluated. The results showed that Swin Transformer-Tiny achieved the best overall performance, reaching an F1-score of 88.64% and outperforming the other architectures.
