Klasifikasi Citra Seni Digital Berbasis Kecerdasan Buatan Generatif Dengan EfficientNet-B3

Abstract

The rapid advancement of generative AI technologies has posed new challenges in the field of visual arts, particularly in distinguishing between human-made and AI-generated artworks. This study aims to develop an automatic detection system capable of identifying AI-generated images using the EfficientNet-B3 architecture. The dataset was manually collected from various online platforms to represent modern visual art styles. The model was trained using a transfer learning approach with progressive fine-tuning to mitigate catastrophic forgetting. Experimental results showed that the best configuration was achieved by unfreezing the top 150 layers of the pre-trained model, resulting in a peak accuracy of 95.6%. The model's performance was evaluated using accuracy, precision, recall, and F1-score metrics. While the model performed well overall, it struggled with unique visual styles such as graphic design, extreme camera angles, and images without human subjects. This study demonstrates that EfficientNet-B3 can be an effective approach for detecting AI-generated art, particularly when trained on a diverse and representative dataset.