FP-ZOO: Fast Patch-Based Zeroth Order Optimization for Black-Box Adversarial Attacks on Vision Models.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) excel in vision tasks but are vulnerable to adversarial attacks.
• Evasion attacks, particularly in black-box settings, threaten DNN reliability in real-world applications.
• Existing methods like Zeroth Order Optimization (ZOO) struggle with high-resolution images due to inefficiency and memory complexity.

Purpose of the Study:

• To propose a novel patch-based fast zeroth order optimization (FP-ZOO) attack.
• To address the limitations of existing ZOO attacks on high-resolution images.
• To improve the efficiency and success rate of black-box evasion attacks against vision models.

Main Methods:

• FP-ZOO partitions images into patches for targeted perturbation.
• It utilizes probability-based sampling and an epsilon-greedy scheduling strategy for effective perturbation generation.
• Large-scale evaluations were conducted on CIFAR-10, CIFAR-100, and ImageNet datasets.

Main Results:

• FP-ZOO achieved 97-100% attack success rate on ImageNet for untargeted attacks.
• The attack demonstrated generation times up to 10 seconds faster than the standard ZOO attack.
• FP-ZOO showed relatively lower performance in targeted attacks, indicating an area for future research.

Conclusions:

• FP-ZOO offers a significant improvement in efficiency and effectiveness for untargeted evasion attacks against vision models.
• The proposed method enhances adversarial robustness by overcoming the limitations of previous ZOO attacks.
• Further research is needed to optimize FP-ZOO for targeted attack scenarios.