STD-Net: A novel source-target discrimination network for copy-move forgery detection using consistency strategy

Most current Copy-Move Forgery Detection (CMFD) methods concentrate on the copy-move areas’ localization. However, accurate discrimination between the source and target regions in copy-move areas still a critical bottleneck. Although existing disambiguation methods have advanced the development of CMFD, they commonly treat disambiguation as a simple classification task or fail to trace effective and stable feature cues. To address these limitations, we convert this challenging task into a consistency identification problem and propose STD-Net, a disambiguation method for source-target discrimination in copy-move regions. Specifically, we take the genuine area as the third-party reference object, aiming to leverage a Consistency Detection between it and the suspicious (copy-move) areas. STD-Net consists of Duplicate Patches Extraction, Genuine Patches Extraction, and Discrimination. Duplicate Patches Extraction extracts duplicated copy-move patches, while the Genuine Patches Extraction matches each copy-move patch with a genuine patch of strictly identical size, which eliminates errors caused by asynchronous scaling and provides a unified benchmark for consistency analysis. The Discrimination phase employs a dual-stream structure, RGB-based Discrimination and Noise-based Discrimination. RGB-based stream detects multi-level texture consistency, while Noise-based stream focuses on high-frequency noise consistency. Both streams are learned for consistency measurement rather than general forged trace detection. Then, the final discrimination output is obtained by fusing the consistency results from both streams using learnable weighting. STD-Net is trained on two synthesized datasets and evaluated on CASIA V2.0, CoMoFoD, and Coverage datasets. The experimental results demonstrate that our method achieves state-of-the-art performance in source-target discrimination for copy-move forgery detection.