Dynamic Routing and Knowledge Re-Learning for Data-Free Black-Box Attack.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep learning models are powerful but vulnerable to adversarial examples.
• Existing data-free black-box attacks have limitations, including static substitute models and reliance on target model data statistics.

Purpose of the Study:

• To propose a novel Dynamic Routing and Knowledge Re-Learning framework (DraKe) for effective data-free black-box adversarial attacks.
• To address limitations of previous methods by enabling dynamic substitute model learning and knowledge adaptation.

Main Methods:

• Developed a dynamic substitute structure learning strategy using a policy network to adapt substitute models to different targets.
• Implemented graph-based structure information learning to capture knowledge from target models.
• Introduced a dynamic knowledge re-learning strategy to refine the substitute model by re-learning hard samples.

Main Results:

• DraKe achieved significant improvements over state-of-the-art competitors on image classification and face recognition benchmarks.
• The framework demonstrated consistent attack superiority across various target models, including residual networks and vision transformers.

Conclusions:

• DraKe offers a more practical and effective approach to data-free black-box adversarial attacks.
• The dynamic and adaptive nature of DraKe shows great potential for real-world applications involving deep learning security.