Wet Paper Coding-Based Deep Neural Network Watermarking.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural network (DNN) models face intellectual property risks due to widespread application.
• Existing watermarking schemes lack tamper-proofing and quantifiable embedding rates.

Purpose of the Study:

• To design a high embedding rate and tamper-proof watermarking scheme for DNN models.
• To address the limitations of current methods in preventing overwriting and quantifying embedding rates.

Main Methods:

• Employed wet paper coding (WPC) by classifying model parameters into 'wet' (important) and 'dry' (unimportant) blocks.
• Proposed an optimized probabilistic selection strategy (OPSS) to identify important parameters without affecting model functionality.
• Modified only unimportant parameters to embed watermarks, preserving critical model components.

Main Results:

• The proposed scheme demonstrates high fidelity, ensuring minimal impact on model performance.
• Achieved strong robustness against potential tampering and unauthorized modifications.
• Exhibited a high embedding rate, effectively protecting intellectual property.

Conclusions:

• The developed watermarking scheme provides a robust and efficient solution for protecting DNN intellectual property.
• The combination of WPC and OPSS offers superior tamper-proofing and embedding capacity compared to existing methods.