A Multimodal Deep Learning Framework For AI-Generated Images Detection using Frequency and Noise Analysis

Abstract

The accelerated development of generative artificial intelligence allowed creating incredibly realistic artificial images, which create immense difficulties in determining visual authenticity in various realms of practice ,including journalism, digital forensics, and social media networks. The proposed paper presents a powerful deep learning architecture that will enable it to detect the level of discrimination between AI- generated and real-life photographs. This system uses the spatial feature extraction approach of a convolutional neural network with the ResNet50 architecture together with frequency-domain and noise-based representations to extract the fine-grained artifacts of generative artworks .A balanced dataset of more than 19,000 images is used, including real and GAN-generated ones. More sophisticated preprocessing strategies, such as data augmentation, Fast Fourier Transform (FFT) and noise residual drive up model generalization among various image distributions. The multimodal fusion method combines the RGB, frequency and noise features, which enhances detection much better than the traditional single-stream methods. As shown in the experimental results, the model proposed has reached the accuracy of around 97.8% and high ROC-AUC scores and low misclassification rates .In addition, explainable AI methods, e.g., Grad-CAM, can give insights into the model decisions visually, which can improve transparency and trust. The suggested system is a scalable and interpretable deepfake detector that has a significant application in a real-life verification system.