Towards Robust Person Re-Identification by Defending Against Universal Attackers

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models for person re-identification (re-ID) are susceptible to adversarial attacks.
• Enhancing the robustness of re-ID systems against these attacks is crucial for real-world applications.

Purpose of the Study:

• To develop a comprehensive framework for analyzing attack-defense dynamics in person re-ID.
• To introduce a novel, adaptive adversarial attack and a robust defense mechanism using meta-learning.

Main Methods:

• A holistic attack-defense framework was established to investigate the interplay between adversarial attacks and defenses.
• A combinatorial adversarial attack was designed, incorporating pixel and color space distortions adaptable to new domains and models.
• A virtual-guided meta-learning algorithm was proposed, utilizing a virtual dataset to explore cross-domain constraints for improved generalization and robustness.

Main Results:

• The combinatorial attack demonstrated effectiveness and universality across different models and datasets.
• The meta-learning algorithm consistently improved various attack and defense strategies.
• The resulting defense models exhibited robustness against state-of-the-art attacks not encountered during training.

Conclusions:

• The proposed attack-defense framework provides a valuable tool for understanding and improving person re-ID robustness.
• The combinatorial attack and virtual-guided meta-learning approach offer significant advancements in adversarial robustness for re-ID systems.