<![CDATA[The rapid advancement of generative artificial intelligence, particularly Generative Adversarial Networks and Diffusion Models, has enabled the creation of synthetic images with a visual quality that is increasingly difficult to distinguish from authentic photographs, raising concerns over misinformation, media manipulation, and digital identity misuse. This study implements a Convolutional Neural Network (CNN) to classify real and AI-generated images and compares the performance of two transfer learning architectures, EfficientNetV2-B0 and MobileNetV3-Large, against a CNN trained from scratch. The dataset consists of 10,930 images collected from two Kaggle repositories, comprising 5,508 AI-generated images and 5,422 real images, which were split into 80% training, 10% validation, and 10% testing data, resized to 224x224 pixels, and augmented prior to training using a batch size of 32, a maximum of 20 epochs, a learning rate of 0.001, and the Adam optimizer. The experimental results show that MobileNetV3-Large achieved the best performance with a training accuracy of 96.84%, a validation accuracy of 96.25%, a testing accuracy of 96.71%, and a testing loss of 0.1031, outperforming EfficientNetV2-B0 (94.41% testing accuracy) and the CNN trained from scratch (91.58% testing accuracy). Hyperparameter experiments further confirm that a batch size of 32 combined with 20 training epochs produces the most stable convergence across all three architectures. The best-performing model was subsequently deployed as a REST API using the Flask framework to support real-time image classification. These findings indicate that transfer learning, particularly with the MobileNetV3-Large architecture, provides an effective and computationally efficient approach for detecting AI-generated images]]>
Copyrights © 2026
