A Dense Residual Network-50 Model for Identification of Real and Fake Images
Gurpreet Singh, Kalpna Guleria · 2024
The rapid advancement of artificial intelligence has led to the creation of highly realistic synthetic images, posing significant challenges in distinguishing them from real images. This study addresses this issue by employing a ResNet-50-based model to classify AI-generated and real images. The dataset comprises 120,000 images, equally divided between REAL images from the CIFAR-10 dataset and FAKE images generated using Stable Diffusion version 1.4. The model's architecture integrates batch normalization and dropout layers to enhance training stability and reduce overfitting. Results indicate a steady improvement in training accuracy, reaching 98.87% by the 22nd epoch, with a corresponding decrease in training loss to 0.0457. The validation accuracy stabilized at 95.59% from the 11th epoch onwards, while validation loss fluctuated, settling at 0.1887 by the 22nd epoch. These outcomes demonstrate the model's robust capability to distinguish between real and AI-generated images, highlighting its potential for applications in digital forensics, content verification, and intellectual property protection. The integration of explainable AI techniques further enhances the model's transparency and reliability, ensuring user trust in its classifications. This research contributes to the ongoing efforts to develop reliable methods for synthetic image detection and emphasizes the importance of early detection in maintaining content authenticity.