Adversarial perturbation and defense for generalizable person re-identification

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Domain Generalizable Person Re-Identification (DG Re-ID) performance relies heavily on the quality of identity-relevant descriptors.
• Hard-matching samples pose a challenge in DG Re-ID due to the difficulty of separating identity-relevant from irrelevant features, hindering generalization.

Purpose of the Study:

• To enhance a model's capability to distinguish identity-relevant features from identity-irrelevant features in hard-matching samples for improved DG Re-ID.
• To propose a novel Adversarial Perturbation and Defense (APD) Re-identification Method to achieve high-performance domain generalization.

Main Methods:

• Introduced a Metric-Perturbation Generation Network (MPG-Net) using metric adversariality to synthesize hard-matching samples by perturbing latent space metric relationships while preserving visual details.
• Developed a Semantic Purification Network (SP-Net) trained on synthesized hard-matching samples to capture high-quality identity-relevant features.
• Incorporated a Semantic Self-perturbation and Defense (SSD) Scheme within SP-Net to disentangle and purify identity-relevant features.

Main Results:

• The proposed APD method, utilizing MPG-Net and SP-Net with SSD, demonstrates effectiveness in the DG Re-ID task.
• Extensive experiments validate the APD method's ability to improve feature separation for hard-matching samples.

Conclusions:

• The APD method successfully enhances feature disentanglement for hard-matching samples in DG Re-ID.
• The proposed approach offers a promising direction for achieving robust domain generalization in person re-identification systems.