Copyright protection of deep neural network models using digital watermarking: a comparative study

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Deep Neural Networks (DNNs) achieve high accuracy in various fields like healthcare and computer vision.
• Training DNNs is computationally intensive, making pre-trained models valuable assets.
• Unauthorized redistribution of pre-trained DNNs poses a significant copyright threat.

Purpose of the Study:

• To review digital watermarking techniques for DNN copyright protection.
• To compare the latest DNN watermarking methods.
• To propose optimizers for enhancing DNN accuracy against fine-tuning attacks.

Main Methods:

• Literature review of digital watermarking for DNNs.
• Comparative analysis of existing watermarking techniques.
• Development and testing of novel optimizers against fine-tuning attacks.
• Black-box experiments comparing proposed optimizers with Stochastic Gradient Descent (SGD).

Main Results:

• Digital watermarking offers effective copyright protection for DNNs.
• Proposed optimizers demonstrate improved accuracy against fine-tuning attacks compared to SGD.
• Comparative study highlights the strengths and weaknesses of various watermarking techniques.

Conclusions:

• Digital watermarking is essential for securing intellectual property in pre-trained DNNs.
• The proposed optimizers contribute to more robust DNN models against unauthorized modifications.
• Further research in secure DNN sharing and copyright protection is warranted.