A Hybrid Approach for AI-Generated Image Detection System Integrating Convolutional Neural Network and Digital Forensic Error Level Analysis

Abstract

The rapid advancement of generative AI, particularly diffusion models, poses a significant threat to digital media authenticity, necessitating robust detection tools. This study aims to develop and evaluate a high-accuracy, explainable system for identifying AI-generated images. The materials include a large-scale benchmark dataset from Kaggle with over 250,000 real and synthetic images across diverse subjects. The proposed method is a novel hybrid framework that integrates Error Level Analysis (ELA), a classic digital forensic technique, with a lightweight Convolutional Neural Network (CNN). ELA is used to expose subtle compression artifacts indicative of synthetic generation, and the CNN classifies the resulting ELA map. To ensure forensic interpretability, the system incorporates Gradient-weighted Class Activation Mapping (Grad-CAM) to provide visual explanations for its classifications. The results show exceptional performance, with the system achieving 96.96% detection accuracy, 95.70% recall, and an AUC of 0.95 on the test set. The Grad-CAM visualizations successfully highlighted the specific image regions influencing the model's decisions. This work primarily validates a hybrid forensic-deep learning approach that is both effective and transparent. For future work, we recommend expanding the evaluation to include datasets from the latest diffusion models, testing "in-the-wild" robustness against social media transformations, and exploring architectural enhancements such as Vision Transformers or ensemble learning to further improve performance and generalization.