ARCHIVES

The rapid growth of artificial intelligence has made it possible to create images that look very real. These images are produced using technologies like Generative Adversarial Networks, diffusion models and text-to-image systems. While these tools offer advantages, they also bring new problems when it comes to verifying the authenticity of digital images. Traditional detection methods often struggle when tested on images from unseen generators or when images have been compressed or altered. This paper introduces a deep learning system to detect AI-generated images. It uses two types of features: features and frequency-domain features. A convolutional neural network is used to capture details while a branch that uses Fourier transforms captures frequency patterns. These two sets of features are combined before the final classification decision is made—either real or AI-generated. To make the system more reliable data augmentation is added. This helps the model stay effective when images have been resized, compressed or blurred. The framework is built to work well on datasets like GenImage and across different generators. Performance is measured using accuracy, precision, recall, F1-score and AUROC.The main goal of this work is to make the system better at working with images that come from generators it has never seen before. This is a problem with existing detection methods. The results, from our system will be compared to CNN models and models that only use frequency features. The comparison will show how better the new method.