Advanced Image Forensics: Detecting Tampered and AI-Generated Images with Adversarial Learning

ICLR 2026 Conference Submission16038 Authors

Keywords: Adversarial learning, tri-branch, image forensics, AIGC detection

Abstract: Detecting image tampering and Artificial Intelligence Generated Images are vital challenges in the fields of computer vision. The primary difficulty in identifying tampered images lies in uncovering minute evidence of manipulation, while AIGC image detection struggles with the increasingly lifelike quality of generated images. Existing solutions often focus on either tampered or AIGC images, yet both can coexist in various contexts. To address this limitation, we propose an advanced framework for social media image forensics that utilizes adversarial learning to identify both tampered and AIGC images. This framework employs a tri-branch architecture that combines generative adversarial learning with deep neural networks, effectively identifying both tampered and AI-generated content. We validated our framework using public image tampering datasets, a specialized AIGC image dataset, and datasets containing both tampered and AI-generated images. Experimental outcomes demonstrate that our framework significantly enhances accuracy by approximately 20\%, 30\%, and 20\% for tampered, AI-generated, and tampered AIGC image datasets, respectively. Code will be available on request.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 16038