Enhanced Synthetic Images Classification with Vision Transformers and Effective Data Augmentation Techniques

Synthetic images have made significant advancements, resulting in highly realistic and high-quality visuals generated at an unprecedented speed. However, the proliferation of synthetic images raises growing concerns regarding their potential misuse, ethical implications, and associated risks in various sectors. This study focuses on leveraging data augmentation techniques to enhance the accuracy and reliability of synthetic image detection. In our study, we fine-tuned a Vision Transformer (ViT) on the CIFAKE dataset, comprising real and synthetic images, augmented using various techniques such as transformations, photometric adjustments, geometric changes, and noise addition. The best results were obtained with ViT fine-tuned by distortions and noise augmentation, achieving an accuracy rate of 98.66%, demonstrating that sophisticated augmentation techniques improve performance. Data augmentation plays a crucial role in handling the increasing realism and sophistication of synthetic images, contributing to more effective content moderation, digital forensics efforts, and the mitigation of potential risks associated with synthetic media.