DynamicPAE: Generating Scene-Aware Physical Adversarial Examples in Real-Time.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Physical adversarial examples (PAEs) highlight real-world risks in deep learning.
• Current PAE generation lacks adaptability to diverse, dynamic scenes.
• There is a need for real-time, observation-conditioned dynamic PAEs.

Purpose of the Study:

• To develop the first generative framework for scene-aware, real-time physical adversarial attacks (DynamicPAE).
• To address challenges in learning sparse relations under noisy feedback during attack training.
• To align generated PAEs with real-world scenarios for effective physical attacks.

Main Methods:

• Introduced residual-guided adversarial pattern exploration to overcome noisy feedback.
• Modeled training degeneracy with limited feedback information restriction.
• Proposed distribution-matched attack scenario alignment, including conditional-uncertainty-aligned data and skewness-aligned objective re-weighting.

Main Results:

• DynamicPAE demonstrates superior attack performance in digital and physical evaluations.
• Achieved a 2.07x boost and 58.8% average AP drop against object detectors.
• Outperformed state-of-the-art static PAE generation methods.

Conclusions:

• DynamicPAE is the first framework enabling end-to-end modeling of dynamic PAEs.
• The proposed methods effectively address noisy feedback and scenario alignment challenges.
• DynamicPAE significantly advances the capability of real-time physical adversarial attacks.