Black-box dnn watermarking copyright protection method based on scene perception and optical modeling
By generating starburst triggers based on scene perception and optical modeling, the problems of insufficient semantic rationality and trigger stability in black-box watermarking technology are solved, achieving effective copyright protection of deep neural networks and improving the concealment and robustness of watermarks.
Patent Information
- Application Number
- CN202610830153.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-10
- Publication Date
- 2026-07-17
AI Technical Summary
Existing black-box watermarking technology struggles to balance semantic rationality and trigger stability, resulting in insufficient watermark concealment and robustness, making it difficult to effectively protect the intellectual property rights of deep neural networks in practical applications.
A scene-aware and optical modeling approach is adopted to locate the light source position in the image through a large visual language model, generate a starburst trigger based on Fraunhofer diffraction theory, and construct a starburst watermark image through pixel-level optimization and neighborhood brightness statistics. The model is trained by combining the basic loss and center loss functions to achieve implicit embedding and verification of the watermark.
It achieves deep integration of triggers and scene semantics, generates glare phenomena in natural optical imaging, ensures the physical rationality and stability of trigger positions, improves the concealment and robustness of watermarks, and avoids destructive changes to images.
Smart Images
Figure CN122415309A_ABST