Reinforced Embodied Active Defense: Exploiting Adaptive Interaction for Robust Visual Perception in Adversarial 3D...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Adversarial attacks threaten 3D visual perception systems in critical applications like autonomous driving.
• Current defenses (e.g., adversarial training) are often passive and struggle with dynamic 3D environments.
• Deep neural networks (DNNs) are vulnerable to manipulation via adversarial patches and objects in complex scenes.

Purpose of the Study:

• To introduce Reinforced Embodied Active Defense (REIN-EAD), a proactive framework to enhance DNN robustness in 3D adversarial settings.
• To develop an adaptive defense mechanism that explores and interacts with the environment.
• To improve the reliability of visual perception systems against sophisticated attacks.

Main Methods:

• REIN-EAD employs a multi-step objective balancing prediction accuracy and entropy minimization.
• An uncertainty-oriented reward-shaping mechanism facilitates efficient policy updates.
• The framework integrates proactive policy learning with embodied scene interaction.

Main Results:

• REIN-EAD significantly reduces adversarial attack success rates across diverse tasks.
• Standard accuracy is preserved during defense.
• The framework demonstrates robust generalization to unseen and adaptive attacks.

Conclusions:

• REIN-EAD offers a scalable and adaptable solution for securing DNN-based perception systems in dynamic 3D environments.
• The proactive and embodied approach overcomes limitations of passive defense strategies.
• This framework is suitable for real-world applications including 3D object classification, face recognition, and autonomous driving.