Handcrafted Feature Extraction and Classification for Distinguishing Real and AI-Generated Images

The generative machine-learning approaches are extensively used to create artificial images. These images are called AI-generated images. The AI-generated images find applications like rapid content creation, creative design and visualization, rapid prototyping, etc. In spite of many applications, the AI-generated images are leading to issues like bias, copyright issues, deepfakes, etc. Owing to the importance of discriminating between synthetic and real images, many research approaches have been proposed. Among these methods, deep learning approaches are popular. However, the design of common deep learning detection approaches to detect the synthetic images that are generated by different generative models is a challenging task. There are issues with these methods to discriminate images that are subjected to various pre-processing steps. To overcome these issues, machine learning approaches with the handcrafted features are gaining importance. In this context, five types of handcrafted feature extraction and random forest-based classification is proposed to distinguish between AI-generated and real images. The simulation experiments are carried out with AI-generated and real facial images, with each group having 70,000 samples. Out of 70,000 samples from each group, 50,000, 10,000, and 10,000 samples are respectively used for training, validating, and testing the classifier. The performance of the classifier is evaluated for accuracy, precision, recall, and F1 score. The recognition accuracy, precision, recall, and F1 scores of 86.03%, 86.15%, 85.87%, and 86.01%, respectively, are achieved for three feature combinations