SecureNet: Proactive intellectual property protection and model security defense for DNNs based on backdoor learning

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning Security

Background:

• Deep neural networks (DNNs) are increasingly vulnerable to privacy breaches and intellectual property (IP) theft.
• Existing watermarking techniques offer passive protection, failing to prevent unauthorized model use or embezzlement.
• Proactive protection methods are crucial for safeguarding DNN models against infringement.

Purpose of the Study:

• To propose SecureNet, a key-based access license framework for proactive intellectual property protection of DNN models.
• To develop mechanisms for license key injection, replacement, and fine-grained user authorization.
• To design defenses against adversarial and backdoor attacks within the SecureNet framework.

Main Methods:

• Implemented a key-based access license framework (SecureNet) using backdoor learning to embed license keys.
• Developed a license key replacement algorithm to ensure DNN model reusability.
• Designed defense mechanisms against adversarial and backdoor attacks, incorporating fine-grained authorization.

Main Results:

• SecureNet demonstrated superior computational efficiency (over 95% improvement) compared to state-of-the-art model parameter encryption.
• The framework effectively defended against adversarial and backdoor attacks without degrading overall model performance.
• Evaluated on diverse datasets (MNIST, Cifar10/100, FaceScrub, CelebA) and models (LeNet-5, VGG16, ResNet18/101, NFNet-F5, MobileNetV3).

Conclusions:

• SecureNet provides a robust and efficient solution for proactive DNN intellectual property protection.
• The key-based framework enhances model security by controlling access and defending against malicious attacks.
• SecureNet offers flexible authorization and maintains model performance, making it suitable for various applications.