Adversarial Metric Attack and Defense for Person Re-Identification

Area of Science:

• Computer Vision
• Machine Learning
• Cybersecurity

Background:

• Person re-identification (re-ID) is crucial for video surveillance.
• Existing re-ID distance metrics are vulnerable to adversarial examples with imperceptible perturbations.
• This vulnerability poses significant security risks for deployed re-ID systems.

Purpose of the Study:

• To investigate the vulnerability of person re-identification (re-ID) distance metrics to adversarial attacks.
• To propose a novel attack methodology, Adversarial Metric Attack, for metric-based re-ID.
• To develop and evaluate defense mechanisms against such adversarial attacks.

Main Methods:

• Proposed Adversarial Metric Attack, a novel methodology for metric analysis in re-ID.
• Conducted comprehensive experiments to demonstrate adversarial effects on re-ID systems.
• Developed a metric-preserving network as a defense strategy against adversarial attacks.

Main Results:

• Demonstrated the extreme vulnerability of current distance metrics in person re-ID to adversarial examples.
• Showcased the effectiveness of the proposed Adversarial Metric Attack.
• Validated the potential of metric-preserving networks in defending against adversarial attacks.

Conclusions:

• Adversarial attacks pose a significant threat to the security of person re-identification systems.
• The proposed Adversarial Metric Attack provides a new perspective on re-ID security.
• Further research into adversarial attack and defense is crucial for secure metric-based applications.