Robust and Secure Deep Neural Network Watermarking Technique Shows Promise
Researchers from IMT Atlantique have presented a new deep neural network watermarking technique, RoSe-Mix, designed for black-box settings. According to the study, RoSe-Mix addresses limitations in existing DNN watermarking approaches by integrating key features from two established methods: RoSe, which uses cryptographic hashing to ensure security, and Mixer, which employs image Mixup to enhance robustness. The experimental results demonstrate that RoSe-Mix achieves security across various architectures and datasets with a robustness to removal attacks exceeding 99%.
Key Takeaways:
- The new technique, RoSe-Mix, is designed to provide robust and secure deep neural network watermarking in black-box settings.
- RoSe-Mix combines key features from two established methods: RoSe and Mixer, to enhance security and robustness.
- The experimental results show that RoSe-Mix achieves security across various architectures and datasets with a robustness to removal attacks exceeding 99%.
- The study concluded that RoSe-Mix addresses limitations in existing DNN watermarking approaches.
- The researchers include Tamara El Hajjar, Mohammed Lansari, Reda Bellafqira, Gouenou Coatrieux, Katarzyna Kapusta, and Kassem Kallas.
- The research was funded by the European Union, Cybaile Industrial Chair, and Agence Nationale De La Recherche.
Statistics:
- Robustness to removal attacks: 99%
- Security across various architectures and datasets: Achieved
- Number of researchers involved: 7
- Funding sources: European Union, Cybaile Industrial Chair, Agence Nationale De La Recherche
Sources:
- RoSe-Mix: Robust and Secure Deep Neural Network Watermarking in Black-Box Settings via Image Mixup. Machine Learning and Knowledge Extraction, 2023,7(2):32.
- https://doi.org/10.3390/make7020032
- Information Technology Newsweekly, July 8, 2025, p 1048.
- NewsRx. Researchers from IMT Atlantique Report on Findings in Knowledge Extraction (RoSe-Mix: Robust and Secure Deep Neural Network Watermarking in Black-Box Settings via Image Mixup).