Toward Real-World Holistic Privacy-Preserving Person Re-Identification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Real-world person re-identification (Re-ID) systems face privacy risks from data and model leakage due to malicious attacks.
• Existing privacy-preserving person re-identification (PPPR) methods struggle to offer comprehensive protection without hindering Re-ID system functionality.

Purpose of the Study:

• To develop a robust and comprehensive privacy-preserving framework for Re-ID systems.
• To address the limitations of current PPPR methods in simultaneously resisting data leakage, model leakage, and combined data & model leakage.

Main Methods:

• Proposed SHIELD, a two-stage privacy-preserving framework.
• Implemented a self-supervised method for Protected Dataset Generation to resist data leakage.
• Introduced Original Feature Deconstruction and Protected Feature Alignment to resist model leakage while preserving retrieval accuracy.

Main Results:

• SHIELD demonstrates superior performance compared to existing PPPR methods.
• The framework offers robust and holistic protection for Re-ID systems.
• SHIELD maintains acceptable retrieval accuracy for authorized users.

Conclusions:

• SHIELD provides a pioneering solution for privacy preservation in Re-ID systems.
• The proposed framework effectively balances privacy protection and system functionality.
• SHIELD represents a significant advancement in the field of privacy-preserving Re-ID.