Safe and Robust Watermark Injection with a Single OoD Image

Area of Science:

• Artificial Intelligence
• Computer Science
• Machine Learning Security

Background:

• Deep neural networks (DNNs) demand significant data and computational power.
• Protecting intellectual property (IP) for DNNs is a critical challenge.
• Existing watermarking methods often rely on training data poisoning, raising privacy concerns and lacking robustness against model modifications.

Purpose of the Study:

• To propose a safe and robust backdoor-based watermarking technique for DNNs.
• To develop a method for IP verification independent of training data.
• To enhance watermark resilience against common removal attacks.

Main Methods:

• A novel watermarking technique leveraging a single out-of-distribution (OoD) image as a secret key.
• Injection of backdoor triggers using the OoD image's diverse knowledge.
• Inducing robustness through random perturbation of model parameters during watermark injection.

Main Results:

• The proposed method is time- and sample-efficient, requiring no access to training data.
• Demonstrated robustness against common watermark removal attacks like fine-tuning, pruning, and model extraction.
• Effective IP verification through the unique secret key derived from the OoD image.

Conclusions:

• The developed watermarking approach provides a secure and resilient solution for DNN intellectual property protection.
• The method overcomes the limitations of traditional watermarking techniques by eliminating the need for training data and enhancing robustness.
• This technique offers a practical and efficient way to safeguard commercial ownership of deep learning models.