Invisible DNN Watermarking Against Model Extraction Attack.

Area of Science:

• Computer Science
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) are valuable commercial assets, necessitating protection against unauthorized use.
• Current DNN watermarking methods are vulnerable to model extraction attacks due to independent watermark and model tasks.

Purpose of the Study:

• To propose a novel collaborative DNN watermarking framework to defend against model extraction attacks.
• To enhance the security and robustness of DNN models through improved watermarking techniques.

Main Methods:

• Developed a collaborative framework where watermark generation and embedding are interconnected.
• Utilized imperceptible trigger samples infused with target-label information for stealth and guidance.
• Implemented feature coupling to align trigger sample features with task distribution samples.

Main Results:

• The proposed framework successfully defends against model extraction attacks.
• Trigger samples are recognized as task distribution samples in stolen models, enabling ownership verification.
• Experiments on CIFAR10, CIFAR100, and ImageNet demonstrated superior performance.

Conclusions:

• The collaborative DNN watermarking framework offers effective protection against model extraction.
• The feature coupling mechanism ensures robust watermark embedding and ownership verification.
• This approach significantly improves the security of valuable DNN models.