Improving AI Generated Image Detection through an Interpretable and Enhanced CNN2D Architecture

This extended study enhances the CIFAKE image classification system by integrating additional optimization layers into the existing CNN2D architecture. The original model successfully differentiated real and AI-generated images with 94% accuracy using two convolutional, max-pooling, and dense layers. In the extended concept, Global Average Pooling and Dropout layers are incorporated to refine feature extraction and minimize overfitting by eliminating less significant features. This enhancement enables the CNN2D to focus on the most discriminative patterns, leading to improved accuracy of 95%. The model is further supported by explainable AI through Gradient Class Activation Mapping (Grad-CAM), which visualizes critical image regions influencing the prediction. This framework not only strengthens the reliability of fake image detection but also provides interpretability for decision transparency. The system implementation in a Flask-based web application ensures real-time identification of AI-generated images, contributing to data authenticity and digital trust.