Physical Adversarial Attack Meets Computer Vision: A Decade Survey.

Area of Science:

• Computer Vision
• Artificial Intelligence Security

Background:

• Deep Neural Networks (DNNs) show vulnerability to adversarial attacks, degrading performance.
• These attacks extend beyond digital environments into the physical world, posing security risks.

Purpose of the Study:

• To provide a comprehensive overview of physical adversarial attacks.
• To introduce a new term, adversarial medium, for artifacts carrying perturbations.
• To propose a systematic evaluation metric for physical adversarial attacks.

Main Methods:

• Distilled four general steps for launching physical adversarial attacks.
• Introduced the concept of 'adversarial medium' to describe physical artifacts.
• Developed hiPAA, a novel evaluation metric for physical adversarial attacks.

Main Results:

• Physical adversarial attacks are influenced by artifacts, termed adversarial mediums.
• The hiPAA metric evaluates attacks across six perspectives: Effectiveness, Stealthiness, Robustness, Practicability, Aesthetics, and Economics.
• Comparative results across task categories are presented, offering insights.

Conclusions:

• Physical adversarial attacks require systematic evaluation due to their real-world implications.
• The adversarial medium concept and hiPAA metric offer a framework for assessing DNN security in physical contexts.
• Further research is suggested based on comparative analyses and observations.