A Pre-trained EfficientNetV2B0 Model for the Accurate classification of Fake and Real Images
Gurpreet Singh, Kalpna Guleria, Shagun Sharma · 2024
This study discovers the use of the EfficientNetV2-B0 model to discriminate between AIgenerated and real images. We utilized a comprehensive dataset of images created by Generative Adversarial Networks (GANs) and real-world photos from publicly available sources like CIFAR-10 and ImageNet. The model was trained with these datasets using advanced preprocessing techniques and data augmentation to enhance its generalization capabilities. Over 20 epochs, the model demonstrated significant improvements in both training and validation accuracy, peaking at $\mathbf{9 9. 6 7 \%}$ training accuracy and 98.56% validation accuracy. Despite some fluctuations, the validation loss showed an overall downward trend, reflecting the model’s increasing proficiency in minimizing classification errors. The results indicate that the EfficientNetV2-B0 model is highly effective in identifying AI-generated images, achieving high accuracy, precision, and recall. This research underscores the potential of advanced deep learning models in addressing the growing challenge of differentiating between synthetic and real images, with implications for enhancing security and authenticity in digital media.