Learning to Learn Transferable Generative Attack for Person Re-Identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning person re-identification (reid) models are crucial for surveillance but vulnerable to adversarial attacks.
• Existing attacks lack effectiveness in cross-domain, cross-model, and cross-test scenarios.

Purpose of the Study:

• To develop a robust method for evaluating the real-world robustness of person re-identification models.
• To enhance the transferability of adversarial examples in black-box attacks.

Main Methods:

• Proposed Meta Transferable Generative Attack (MTGA) using meta-learning optimization.
• Mimicked cross-model & dataset black-box attacks using diverse re-id models and datasets.
• Introduced Perturbation Random Erasing and Normalization Mix strategies to improve attack generalization.

Main Results:

• MTGA demonstrates superior performance in cross-model & dataset attacks, outperforming SOTA by 20.0% mean mAP drop rate.
• MTGA achieves 11.3% better performance in cross-model & dataset & test attacks compared to SOTA.
• The method effectively generates adversarial examples transferable across diverse re-id models and domains.

Conclusions:

• MTGA provides a powerful tool for assessing the robustness of person re-identification systems.
• The proposed strategies enhance adversarial attack transferability, crucial for real-world security evaluations.
• This research contributes to understanding and mitigating adversarial vulnerabilities in AI-powered surveillance.