The rapid advancement of generative artificial intelligence has led to the widespread creation of highly realistic synthetic images, making it increasingly challenging to distinguish authentic visual content from AI-generated media. This research presents a comprehensive AI-powered fake image detection system designed to accurately classify images using advanced deep learning techniques. The proposed approach leverages transfer learning with the EfficientNet-B4 architecture, chosen for its optimal balance between parameter efficiency and feature extraction depth. To ensure high generalization, the model was trained on a massive, diverse dataset comprising 330,335 images, including human portraits, natural landscapes, architectural structures, and anime-style illustrations. The deep learning pipeline, implemented in PyTorch, incorporates Mixed Precision Training (MPT) for computational efficiency and a OneCycleLR learning rate scheduler for stable convergence. Experimental evaluations demonstrate that the system achieves high classification accuracy across multiple domains, significantly outperforming standard convolutional neural network baselines. Furthermore, the system integrates Grad-CAM (Gradient-weighted Class Activation Mapping) to provide visual explanations of the model’s decision-making process, highlighting the specific regions of an image that contribute to a
Gadekar et al. (2026) studied this question.