Hybrid intelligence approach for detecting synthetic art

Abstract This study presents a hybrid intelligence approach for detecting synthetic digital art, combining human-in-the-loop strategies with active learning. The goal is to develop a robust AI model that distinguishes synthetic images regardless of generative techniques or image perturbations. The motivation behind this research is to address the rapid development of new generative techniques that render existing classifiers obsolete. Additionally, the paper addresses the use of perturbations designed to break or bypass classifiers. Using a pre-trained ResNet-152 model as the baseline, the methodology involves iterative training with human feedback, integrating active learning and transfer learning techniques. The dataset includes authentic and synthetic artworks from models like Stable Diffusion and Latent Diffusion and perturbed images from the Glaze tool. The model achieved 98.65% accuracy, outperforming benchmarks with fewer labeled samples and demonstrating resilience against image perturbations. Initially, accuracy dropped to 75% with unseen generative techniques but recovered to 98% through active learning. This research lays the groundwork for future synthetic image detection, enhancing transparency, and protecting intellectual property in digital art.