Image tampering detection combining dynamic attention mechanism and multidimensional feature input

With the diversification of media technologies, image tampering has become increasingly easy to accomplish, leading to decreased information credibility, the spread of fake news, and growing social impact. To address issues such as insufficient feature extraction, scattered attention, and noise interference in image tampering detection, we propose a dynamic attention mechanism and multi-dimensional feature input-based detection network—DSRU-Net. This network initially extracts and integrates image features from multiple perspectives at the input stage, emphasizing subtle residual noise and edge information. Additionally, a dynamic attention mechanism is embedded in the encoder, which adjusts the compression ratio based on different inputs to enable more precise information flow management and feature representation. Experimental results demonstrate that, compared to MVSS-Net, the proposed method improves the F1 score by 2.8% on the CASIA dataset and by 3.3% on the COLUMB dataset.