A Comparative Multi-Model Framework for Explainable Detection of AI-Generated Images using Grad-CAM

The rapid advancement of generative artificial intelligence, especially Generative Adversarial Networks (GANs) and diffusion models, has greatly improved the realism of synthetic images, raising concerns about misinformation, identity manipulation, and digital fraud. In this paper, we propose a comparative multi-model framework for explainable detection of AI generated images using three deep learning architectures, namely Baseline CNN, ResNet50 and EfficientNet-B0. The models are trained on a balanced dataset of 50,000 images combining CIFAKE and Tiny GenImage to ensure diversity across multiple generative techniques and to improve generalization capability. To improve interpretability, Gradientweighted Class Activation Mapping (Grad-CAM) is also integrated to visualize the decisioncritical regions. Experimental results show that EfficientNet-B0 achieves the highest performance of 97.09% accuracy which is better than ResNet50 (96.78%) and Baseline CNN (92.89%) and 91.3% accuracy on unseen generators. The system is implemented as a full-stack application with FastAPI and React which enables real-time prediction, model switching and visual analytics. The proposed framework presents an accurate, scalable, and interpretable solution, while also pointing out the challenges of cross-generator generalization in AIgenerated image detection.