Segment and Recover: Defending Object Detectors Against Adversarial Patch Attacks.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Cybersecurity

Background:

• Object detection models are vital for applications like autonomous driving and security.
• Adversarial patch attacks pose a significant threat to the reliability of these models, especially in dynamic environments.
• Current defenses struggle with variations in patch scale, position, and orientation.

Purpose of the Study:

• To develop a patch-agnostic defense scheme for object detection models against adversarial patch attacks.
• To address the limitations of existing defenses in dynamic and complex environments.
• To improve the robustness of object detection systems in safety-critical applications.

Main Methods:

• Introduced Segment and Recover (SAR), a defense scheme based on image preprocessing.
• Integrated a patch-agnostic detection frontend with a broken pixel restoration backend.
• Developed SAR to handle large-mask-covered object-hiding attacks and break limitations of patch scale, shape, and location.

Main Results:

• SAR accurately localizes adversarial patches and restores affected pixels.
• Evaluations show SAR is compatible with various pre-trained object detectors.
• Demonstrated notable resilience improvements over state-of-the-art methods against diverse patch types.

Conclusions:

• SAR offers a robust and adaptable defense against adversarial patch attacks in object detection.
• The patch-agnostic nature and preprocessing-based approach eliminate the need for additional model training.
• SAR enhances the reliability and safety of object detection systems in real-world, dynamic scenarios.