An Optimized Black-Box Adversarial Simulator Attack Based on Meta-Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Deep neural networks (DNNs) exhibit security vulnerabilities, making them susceptible to adversarial attacks.
• Black-box adversarial attacks are particularly realistic due to the inherent opacity of DNNs.
• Existing black-box attack methods often fail to fully leverage available query information.

Purpose of the Study:

• To validate the utility of feature layer information from meta-learned simulator models in adversarial attacks.
• To introduce an optimized black-box attack method, Simulator Attack+, to improve query efficiency.
• To enhance the performance of adversarial example generation in DNNs.

Main Methods:

• The study validates the use of feature layer information from a meta-learned simulator model.
• Proposed Simulator Attack+ incorporates a feature attentional boosting module.
• Implemented a linear self-adaptive simulator-predict interval mechanism and an unsupervised clustering module for targeted attacks.

Main Results:

• Simulator Attack+ significantly reduces the number of queries required for successful adversarial attacks.
• The proposed optimizations enhance the efficiency of generating adversarial examples.
• Experimental results on CIFAR-10 and CIFAR-100 datasets demonstrate improved query efficiency while maintaining attack success.

Conclusions:

• Feature layer information from simulator models is crucial for effective black-box adversarial attacks.
• Simulator Attack+ represents a significant advancement in optimizing query efficiency for black-box attacks.
• The developed methods offer a more practical and efficient approach to assessing DNN security vulnerabilities.