Query-Efficient Black-Box Adversarial Attacks Guided by a Transfer-Based Prior.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models are vulnerable to adversarial attacks, necessitating research into identifying and mitigating these weaknesses before deployment.
• Black-box adversarial attacks pose a significant challenge as they require crafting adversarial examples without direct access to the target model's gradients.

Purpose of the Study:

• To develop more effective and efficient black-box adversarial attack methods.
• To address the limitations of existing gradient approximation techniques in high-dimensional input spaces.

Main Methods:

• Proposed two prior-guided random gradient-free (PRGF) algorithms: one based on biased sampling and the other on gradient averaging.
• Integrated a transfer-based prior from a surrogate model's gradient with query feedback information.
• Utilized theoretical analysis to determine an optimal coefficient for integrating prior knowledge and query information.

Main Results:

• Demonstrated significantly higher attack success rates compared to state-of-the-art methods.
• Achieved these results with substantially fewer queries to the black-box models.
• Validated the effectiveness of the proposed PRGF algorithms through extensive experiments.

Conclusions:

• The proposed PRGF algorithms offer a superior approach to black-box adversarial attacks.
• These methods effectively leverage both surrogate model gradients and direct query feedback for improved performance.
• The findings contribute to a better understanding of adversarial robustness in deep learning.