Targeted Adversarial Camouflage Texture for Fooling Object Detectors via Native Supervision Redirection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Cybersecurity

Background:

• Adversarial camouflage enables persistent, multi-view attacks in physical environments, surpassing single-view methods.
• Current methods often focus on non-targeted attacks, limiting their destructiveness and stealth for security evaluations.

Purpose of the Study:

• To develop a targeted adversarial camouflage approach (TACT) for more effective physical attacks.
• To address the limitations of non-targeted attacks in real-world vision system security evaluations.

Main Methods:

• TACT utilizes a full-coverage physical camouflage pipeline, replacing original category supervision with a predefined target class.
• It redirects optimization gradients to guide 3D texture towards target category features using existing object detector mechanisms.
• No network redesign, novel loss functions, or rendering pipeline modifications are required.

Main Results:

• TACT-person achieved a 51.91% average targeted attack success rate across seven object detectors, demonstrating transferability.
• In physical tests, TACT-bird reduced mAP50-95 by 59.87% on YOLOv8.
• A gap between physical and digital attack success rates indicates the physical pipeline acts as a low-pass filter for target class granularity.

Conclusions:

• Native supervision redirection is a viable strategy for targeted physical adversarial attacks.
• Coarse-grained target classes transfer more robustly through physical pipelines than fine-grained ones.
• Target class feature granularity significantly impacts physical-domain attack effectiveness.