A Local Adversarial Attack with a Maximum Aggregated Region Sparseness Strategy for 3D Objects.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Cybersecurity

Background:

• Deep neural networks for object detection are vulnerable to adversarial attacks.
• Existing 3D adversarial attacks (3D-AE) are often conspicuous and ineffective in real-world scenarios due to large, dispersed modifications.

Purpose of the Study:

• To develop a subtle and efficient 3D adversarial attack method for object detection models.
• To address the challenge of maximizing attack effectiveness with minimal and concentrated camouflage.

Main Methods:

• Proposed a local 3D attack method driven by a Maximum Aggregated Region Sparseness (MARS) strategy.
• Utilized aggregation and sparseness regularization for concentrated and stealthy camouflage.
• Employed neural rendering for multi-angle augmented data to identify universal critical decision regions.

Main Results:

• Achieved an average attack efficiency of 1.724 on YOLOv3 and v5 networks using the CARLA dataset, a 134% improvement over baseline methods.
• Demonstrated that the MARS attack method is both stealthy and aggressive across different viewpoints.
• Showcased strong attack transferability, decreasing average precision by 0.488 and 0.662 when combined with texture optimization methods.

Conclusions:

• The MARS strategy offers a significant advancement in 3D adversarial attacks, enhancing stealth and effectiveness.
• The method highlights potential security risks to real-world object detection systems.
• The research provides a foundation for developing more robust defense mechanisms against sophisticated adversarial attacks.