Attention based deep learning model for detecting copy move forgery

Abstract The growing reliance on digital imagery across platforms like social media, journalism, and academic publishing has amplified concerns regarding image authenticity. As digital images are frequently used as evidence and for critical decision-making, their susceptibility to manipulation poses a serious challenge. With the rise of advanced editing tools, image forgery particularly copy-move forgery, where a section of an image is duplicated within the same image, has become increasingly difficult to detect. This study introduces a deep learning-based framework enhanced with attention mechanisms to effectively identify such forgeries. The proposed system incorporates MultiResUNet as the core architecture, along with specialized attention modules such as CBAM, PAM, and SPAM. The proposed model employs Conditional Random Fields (CRF) to improve the accuracy of boundary segmentation. Extensive testing on standard benchmark datasets reveals that the model performs reliably across diverse manipulation types, indicating its robustness and suitability for practical use in digital forensic investigations.