Investigating vulnerabilities of gait recognition model using latent-based perturbations

Area of Science:

• Computer Science
• Artificial Intelligence
• Security Systems

Background:

• Video surveillance is crucial for security, with gait recognition offering unique identification capabilities.
• Deep learning models for gait recognition are vulnerable to adversarial attacks, posing a significant security challenge.
• Existing attacks often require extensive model access or lack real-world applicability.

Purpose of the Study:

• To propose a novel, practical, and transferable black-box attack against gait recognition systems.
• To develop an attack method that is effective and perceptually realistic in limited-access scenarios.
• To evaluate the vulnerability of gait recognition models to sophisticated adversarial attacks.

Main Methods:

• Introduced the Black-box-Latent-GEI (BLG) attack, a novel black-box adversarial technique.
• Developed AdvHelper, a surrogate model to simulate the target gait recognition system.
• Implemented PerturbGen using an encoder-decoder framework with reconstruction and perceptual losses for realistic perturbations.

Main Results:

• The BLG attack achieved a high success rate of 94.33% on the CASIA-gait dataset.
• Adversarial samples generated were both effective and perceptually realistic.
• Demonstrated the feasibility of black-box attacks in realistic surveillance contexts.

Conclusions:

• The proposed BLG attack presents a significant advancement in understanding adversarial vulnerabilities in gait recognition.
• The method offers a practical and transferable approach for evaluating model robustness.
• Highlights the need for developing more resilient gait recognition systems against sophisticated adversarial threats.