Hybrid Deep Learning for Image Authenticity: Distinguishing Between Real and AI-Generated Images

Abstract

The increasing use of artificially generated images raises significant concerns about the authenticity of digital content. This study introduces a hybrid deep learning model for binary classification of real and generated images by combining spatial and relational features. The central idea is to integrate a convolutional backbone adapted from ResNet18 for visual feature extraction with a graph representation based on nearest-neighbor relations to capture inter-image similarities. The objective is to evaluate whether this dual-feature approach improves classification performance compared to single-feature baselines. Using a balanced dataset of 1,256 images (744 real and 512 generated), the model was trained on 70% of the data and tested on the remaining 30%. Experimental findings demonstrate that the model achieved an overall accuracy of 88%, with precision of 0.91 and recall of 0.89 for real images, and precision of 0.85 and recall of 0.87 for generated images. The corresponding F1 scores were 0.90 and 0.86, yielding a macro average F1 of 0.88. Confusion matrix analysis shows balanced misclassification across both classes, while stable performance across epochs indicates reliable learning behavior. Results confirm that the hybrid model achieves stronger classification effectiveness than convolution-only or graph-only baselines. The novelty of this work lies in demonstrating that the integration of spatial and relational learning provides a more robust framework for detecting synthetic images than single-modality approaches. The contribution of this research is both methodological, in proposing a hybrid architecture that unifies convolutional and graph-based learning, and practical, in providing empirical evidence that such integration enhances the reliability of image authenticity verification. While the absence of a validation set limited hyperparameter optimization and early stopping, the findings indicate that this hybrid design offers a promising direction for improving the robustness and generalizability of synthetic image detection.