Visual Veracity: Advancing AI-Generated Image Detection with Convolutional Neural Networks
Achal Shankar Gupta, Krishan Pratap Shreneter, Smriti Sehgal
Abstract
The rapid advancements of image generation models, especially prompt-based image generation models, have started a new chapter in how we understand visual information, blurring the lines between what is real and what is synthetic. Moreover, the propagation of AI-generated images across industries, coupled with the easier accessibility of generative models, raises concerns about misinformation and ethical implications. This study delves into the critical need for a robust image classification model to differentiate between real and AI-generated images. Using the CIFAKE dataset, a comprehensive collection of AI-generated and real images and employing transfer learning with various convolutional neural network (CNN) architectures, this study aims to further advance AI-generated image detection to new heights. Training of models was conducted in two phases; one involved using pre-trained weights and freezing the base model layers, while the other involved fine-tuning the base model. The most optimal model with EfficientNet as the base model achieved a validation accuracy of 97.29%. In a world where the authenticity of visual content is increasingly vital, this study holds promise for applications spanning content moderation, cybersecurity, and digital forensics.