Image Manipulation Detection Based on Ringed Residual Edge Artifact Enhancement and Multiple Attention Mechanisms

As image editing techniques continue to evolve, concerns over the security risks associated with forged images are growing. Previous studies have suggested that tampering traces hidden at the edges of images are crucial for detecting manipulated regions. To better cope with image tampering post-processing methods, we propose REM-U2-Net, a U-shaped network designed to detect and localize image tampering traces by enhancing edge artifacts. REM-U2-Net extracts the noise distribution of an image by combining it with RGB features as input to capture subtle manipulation traces that may not be visible in the RGB domain. Additionally, ringed residual edge artifact boosting and symmetric attention module designs enable the model to detect both manipulated edges and regions, making it more effective in coping with a wide range of manipulation attacks. We conducted extensive experiments on multiple datasets and demonstrated the effectiveness of our method. Furthermore, REM-U2-Net exhibits excellent robustness to various post-processing methods.