Deep Model Intellectual Property Protection via Deep Watermarking.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) face significant intellectual property (IP) infringement risks, including model theft via fine-tuning or surrogate model training.
• Existing IP protection methods for DNNs are under-researched, particularly for low-level computer vision tasks.

Purpose of the Study:

• To propose a novel model watermarking framework for protecting deep networks used in image processing and low-level computer vision.
• To develop a robust and invisible watermarking mechanism resistant to various IP infringement attacks.

Main Methods:

• A task-agnostic barrier is integrated post-target model to embed a unified, invisible watermark into model outputs.
• A deep invisible watermarking mechanism is designed to handle watermarks ranging from binary bits to high-resolution images.
• Joint training of the target model and watermark embedding allows the barrier to be absorbed into the original model.

Main Results:

• The proposed framework successfully embeds and extracts hidden watermarks, even when attackers train surrogate models.
• Extensive experiments validate the framework's robustness against attacks employing different network structures and objective functions.
• The watermarking mechanism is effective for both binary and high-resolution image watermarks.

Conclusions:

• The developed deep model watermarking framework offers a viable solution for protecting intellectual property in computer vision models.
• The proposed method demonstrates strong resilience against common model stealing techniques.
• This research addresses a critical gap in the under-researched area of deep network IP protection.