A Convolution Neural Network Based Classifier for Discerning AI Generated Images and Real Images
Uppala Eswar Basava Kumar, Chintakuntla Vivekreddy, Sujatha Kamepalli · 2024
The exponential advancement of artificial intelligence (AI) has led to the creation of synthetic images that closely resemble real-world photographs. However, the widespread distribution of AI-generated content poses significant challenges in distinguishing between genuine and synthetic imagery, particularly in cybersecurity, content moderation, and image forensics. Traditional image classification methods struggle with this task due to the striking visual similarities between the two types of images. Convolutional Neural Networks (CNNs) have emerged as powerful tools for image classification, leveraging their ability to extract complex features from raw pixel data. In this study, we propose an enhanced CNN methodology to differentiate between AI-generated and real images. Specifically, we focus on three popular CNN architectures: ResNet-50, RexNet-150, and RexNet-200. By augmenting these models with advanced architectural enhancements and training strategies, we aim to improve their discriminative capabilities. Our experimental findings reveal that RexNet-150 surpasses ResNet-50 and RexNet-200, achieving superior accuracy in distinguishing between AI-generated and real images. RexNet-150 attained an impressive 97.9% accuracy on the test dataset, outperforming ResNet-50 (93.1%) and RexNet-200 (90.3%). Moreover, RexNet-150 demonstrated enhanced training convergence and speed, indicating its effectiveness in learning discriminative features for image classification tasks. The superior performance of RexNet-150 holds significant implications for online safety, content moderation, and image authenticity detection. These advancements in CNN technology foster trust in digital imagery and enhance confidence in AI applications across diverse domains.