Generative-Model-Based AI Image Forgery Forensics and Efficient Detection Framework
TL;DR
This work proposes an efficient image forensics framework based on generative models to identify and trace AI-forged images, and introduces feature trajectory prediction and multimodal feature fusion to enhance the detection of subtle forgery traces in low-quality images.
Abstract
: The rapid advancement of generative artificial intelligence has brought significant convenience to image processing but has also led to a surge in AI-forged images. In e-commerce, malicious actors exploit these tools to fabricate "damaged goods" images for refund fraud. To address this, we propose an efficient image forensics framework based on generative models to identify and trace AI-forged images. By integrating reverse forensics of generative models, this framework leverages a "magic defeats magic" approach. We introduce feature trajectory prediction and multimodal feature fusion to enhance the detection of subtle forgery traces in low-quality images. Furthermore, an efficient batch detection system using a fast-screening and fine-detection cascade is developed to meet the real-time processing demands of large-scale e-commerce platforms. The framework provides not only binary classification but also explainable forensics via heatmaps and frequency anomaly visualizations. Our approach demonstrates strong robustness and high accuracy, offering a scalable technical path for cross-domain applications including social media content verification, judicial authentication, and copyright protection.