Related Experiment Video
Updated: Dec 21, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
926
Watermarking Deep Neural Networks in Image Processing
Summary
This study introduces a novel deep neural network (DNN) watermarking framework to protect intellectual property in computer vision. The method ensures model performance is maintained while preventing unauthorized use of pretrained models.
Area of Science:
- Computer Vision
- Deep Learning
- Intellectual Property Protection
Background:
- Sharing pretrained deep neural network (DNN) models is prevalent in computer vision.
- Protecting intellectual property and preventing misuse of these models is a growing concern.
Purpose of the Study:
- To develop a robust framework for watermarking DNNs, specifically for low-level image processing tasks.
- To safeguard the intellectual property of DNN model creators.
Main Methods:
- A black-box watermarking method exploiting DNN overparameterization was developed.
- Image denoising and superresolution tasks were used as case studies.
- An auxiliary module for watermark visualization was created for verification.
Main Results:
- The proposed watermarking framework demonstrated no significant impact on model performance.
- The method proved robust against common attacks.
- Watermark information could be effectively visualized.
Conclusions:
- The developed framework offers an effective solution for watermarking DNNs in image processing.
- It provides a means to protect intellectual property without compromising model utility or robustness.
