Query-Efficient Black-Box Adversarial Attack With Customized Iteration and Sampling.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) are vulnerable to adversarial attacks in black-box settings.
• Existing black-box attacks, like transfer-based and decision-based methods, suffer from overfitting and low query efficiency, respectively.

Purpose of the Study:

• To develop a new framework for query-efficient black-box adversarial attacks.
• To propose a novel attack method, Customized Iteration and Sampling Attack (CISA), that bridges transfer-based and decision-based approaches.

Main Methods:

• Investigated the relationship between noise variance, sampling, noise compression monotonicity, and decision-based attack transition functions.
• Developed CISA, which estimates decision boundary distance for stepsize, uses dual-direction iteration for adversarial examples, and employs customized pixel-wise sampling for noise compression.
• Relaxed the state transition function for enhanced query efficiency.

Main Results:

• CISA demonstrates significant advantages in query efficiency compared to existing black-box adversarial attack methods.
• The framework provides insights into optimizing decision-based attacks.

Conclusions:

• CISA offers a more efficient approach to black-box adversarial attacks on DNNs.
• The proposed framework enhances understanding and performance of decision-based attacks by addressing limitations of current methods.