Defending Person Detection Against Adversarial Patch Attack by Using Universal Defensive Frame.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Person detection is crucial for human-centric computer vision but vulnerable to adversarial patch attacks.
• These attacks pose significant risks in safety-critical applications such as autonomous driving and security systems.
• Existing defenses against adversarial patch attacks on person detection are limited.

Purpose of the Study:

• To propose a novel defense strategy against adversarial patch attacks targeting person detection networks.
• To develop a defensive frame that mitigates the impact of adversarial patches while preserving performance on clean data.

Main Methods:

• A novel defense strategy is proposed, involving the optimization of a defensive frame for person detection.
• The defensive frame is generated using a competitive learning algorithm, featuring an iterative competition between detection threatening and shielding modules.
• This approach aims to alleviate the effects of adversarial patches without compromising detection accuracy on unperturbed images.

Main Results:

• The proposed defensive frame effectively mitigates adversarial patch attacks on person detection.
• Experimental results demonstrate the method's capability to defend person detection networks against such attacks.
• The defense strategy maintains person detection performance on clean images.

Conclusions:

• The developed defensive frame offers a robust solution for enhancing the security of person detection systems against adversarial patch attacks.
• This research addresses a critical gap in defending person detection against sophisticated adversarial manipulations.
• The proposed method shows promise for improving the reliability of computer vision systems in real-world applications.