Image-Level Adaptive Adversarial Ranking for Person Re-Identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Cybersecurity

Background:

• Deep neural networks are vulnerable, limiting person re-identification (ReID) in smart security.
• Existing attacks on ReID systems often focus on adversarial samples or metric distance disruption.
• Evaluating ReID robustness requires considering both adversarial samples and metric distances comprehensively.

Purpose of the Study:

• To propose an image-level adaptive adversarial ranking method for robust ReID model evaluation.
• To develop a strategy that adapts to real-world pedestrian variations and adversarial environments.
• To comprehensively assess the security of ReID methods against sophisticated attacks.

Main Methods:

• An image representation enhancement module uses channel-wise information entropy and a generative adversarial network to create refined adversarial samples.
• An adaptive weight confusion ranking loss is introduced to perturb ranking by adjusting sample distances.
• The method is adaptive, requiring no extra hyperparameters or training data.

Main Results:

• The proposed method effectively generates refined adversarial samples with richer information content.
• Adaptive perturbation of ranking successfully interferes with system performance by manipulating sample distances.
• Experiments on Market1501, CUHK03, and DukeMTMC datasets demonstrate significant effectiveness in attacking ReID systems.

Conclusions:

• The image-level adaptive adversarial ranking method provides a comprehensive approach to evaluating ReID security.
• This adaptive attack strategy enhances the understanding of ReID model vulnerabilities in adversarial settings.
• The findings are crucial for developing more robust and secure person re-identification systems for smart security.