Query-Efficient Hard-Label Attack: A Prior-Guided Adam Ray Search Optimization.

Area of Science:

• Artificial Intelligence
• Computer Vision

Background:

• Deep neural networks (DNNs) are susceptible to adversarial examples.
• Attacking DNNs in hard-label black-box settings is challenging due to limited query access and high dimensionality.

Purpose of the Study:

• To propose two novel, query-efficient attack methods for generating adversarial examples in hard-label black-box settings.
• To enhance the gradient estimation and convergence speed of adversarial attacks.

Main Methods:

• Adam-OPT: Integrates Adam-based adaptive optimization into a ray-search framework for stabilized zeroth-order gradient updates.
• Prior-Adam-OPT: Incorporates transfer-based priors from surrogate models to improve gradient estimation accuracy and query efficiency.

Main Results:

• Both methods demonstrate superior performance on CIFAR-10, ImageNet, and zero-shot CLIP models.
• Consistent reduction in perturbation magnitudes and improved attack efficiency compared to existing state-of-the-art hard-label attacks.

Conclusions:

• Combining adaptive optimization with transfer-based priors offers a scalable and robust framework for generating high-quality adversarial examples.
• Ablation studies confirm the significance of gradient estimation vectors and surrogate model quality for attack effectiveness.