MOVE: Effective and Harmless Ownership Verification via Embedded External Features.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning Security

Background:

• Deep neural networks (DNNs) are valuable intellectual property due to extensive training resources.
• Model stealing attacks pose a significant threat, allowing adversaries to create functional copies of DNNs.
• Existing defenses may introduce new security risks or fail against diverse attack types.

Purpose of the Study:

• To propose an effective and harmless model ownership verification (MOVE) method.
• To defend against various model stealing techniques simultaneously without introducing new vulnerabilities.
• To protect the intellectual property of well-trained DNN models.

Main Methods:

• Embedding defender-specified external features into training samples using style transfer.
• Training a meta-classifier to identify stolen models based on the presence of embedded knowledge.
• Developing and analyzing the MOVE method for both glass-box and closed-box scenarios.

Main Results:

• The proposed MOVE method demonstrates effectiveness in verifying model ownership.
• Experiments confirm the method's robustness against different model stealing attacks.
• MOVE shows resistance to potential adaptive attacks, ensuring comprehensive protection.

Conclusions:

• MOVE provides a secure and effective solution for DNN model ownership verification.
• The technique successfully detects stolen models by verifying embedded external features.
• MOVE offers a promising approach to safeguard intellectual property in machine learning.