A Dual-Branch Feature Fusion Framework based on GenAI for Robust Detection of Medical CT Image Tampering

ICLR 2026 Conference Submission17438 Authors

Keywords: GenAI, Feature fusion, CT image forensics, Robust detection

Abstract: CT images are a critical diagnostic tool in modern medicine, yet they face risks to image authenticity posed by diverse image tampering techniques, which could disrupt the normal medical order and the societal trust system. Although image tamper detection technology has made some progress, techniques specifically targeting CT image tampering detection are extremely scarce. This paper proposes a dual-branch feature fusion framework based on generative artificial intelligence (GenAI) for CT image tampering detection. This framework utilizes a ResNet-based generator to create tampered images that rich in edge and noise features, which are then fed into a dual-branch discriminator to separately learn noise and edge features. For the features learned by the discriminator, we designed a feature fusion module that captures complex relationships between features and selects different feature weights through a cross self-attention mechanism and dynamic feature selection. Additionally, we created a CT image tampering dataset and conducted comparative experiments with existing mainstream methods on public image tampering datasets and the self-made CT tampering image dataset. Experimental results demonstrate that the proposed method possesses good accuracy and robustness, providing an effective solution for CT image tamper detection.

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

Submission Number: 17438