RoPaSS: Robust Watermarking for Partial Screen-Shooting Scenarios
Authors: Zehua Ma, Han Fang, Xi Yang, Kejiang Chen, Weiming Zhang
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments have demonstrated the excellent performance of Ro Pa SS in partial screen-shooting traceability, with extraction accuracy of above 93% in frontal shooting and above 86% in 30 shooting even if only 50% of the image content is captured. |
| Researcher Affiliation | Collaboration | Zehua Ma1, Han Fang2*, Xi Yang3, Kejiang Chen1*, Weiming Zhang1, 4 1Anhui Province Key Laboratory of Digital Security, University of Science and Technology of China 2National University of Singapore 3Jinan University 4Institute of Hefei High Dimensional Data Ltd. |
| Pseudocode | No | The paper describes the framework and operations in text and diagrams (e.g., Fig. 2 and Fig. 3) but does not include any explicitly labeled pseudocode or algorithm blocks with structured steps. |
| Open Source Code | No | The paper mentions "Based on the source code and pre-trained models released, we adjusted the embedding strength of part comparison schemes..." but this refers to source code from *comparison* schemes, not the code for the method proposed in this paper. There is no explicit statement or link provided for the authors' own implementation. |
| Open Datasets | Yes | In the training stage, we randomly select 10000 images from the COCO dataset (Lin et al. 2014) as the training set. ... In experiments, we use images from the USC-SIPI dataset (USC-SIPI 2022) as host images. |
| Dataset Splits | No | The paper states, "In the training stage, we randomly select 10000 images from the COCO dataset (Lin et al. 2014) as the training set." and "In experiments, we use images from the USC-SIPI dataset (USC-SIPI 2022) as host images." While specific datasets are mentioned, there are no explicit details about how these datasets were split into training, validation, or test sets with percentages or sample counts for the authors' own model evaluation, beyond selecting 10000 images for training from COCO. |
| Hardware Specification | Yes | The proposed Ro Pa SS is implemented by Py Torch (Collobert, Kavukcuoglu, and Farabet 2011) and executed on NVIDIA RTX A6000. |
| Software Dependencies | No | The paper mentions "implemented by Py Torch" and "Homography transformation and restoration are implemented using Kornia" but does not provide specific version numbers for these software components. A Python version is also not specified. |
| Experiment Setup | Yes | Training images are randomly cropped and resized to 128 128 3 pixels, and the watermark message is randomly generated with length L = 30. The batch size is 64. ...L = λ1LI + λ2LM, where λ1 and λ2 are set as 400 and 1 by default. |