Privacy Beyond the Face: Assessing Gait Privacy Through Realistic Anonymization in Industrial Monitoring.

Area of Science:

• Computer Vision
• Biometrics
• Cybersecurity

Background:

• Camera-based monitoring is vital for industrial processes but raises privacy issues.
• Full-body anonymization can obscure identity while retaining data utility.
• The impact of anonymization on gait-based biometric recognition is not well understood.

Purpose of the Study:

• To evaluate if full-body anonymization methods disrupt gait characteristics crucial for identity recognition.
• To assess the effectiveness of pose-preserving anonymization in concealing gait-related biometric traits.
• To determine if anonymization impacts the linkability between original and anonymized video data.

Main Methods:

• Utilized DeepPrivacy2 for realistic full-body anonymization.
• Employed a custom CCTV-like industrial dataset with original and anonymized video sequences.
• Quantified keypoint shifts, analyzed gait identification accuracy, and measured domain transfer between data types.

Main Results:

• Anonymization effectively removed visual appearance but largely preserved gait identity.
• Pose-driven anonymization did not significantly disrupt gait-based person identification.
• Cross-domain linkability between original and anonymized data remained high, indicating privacy leakage.

Conclusions:

• Pose-preserving full-body anonymization is insufficient for protecting gait-related biometric privacy in industrial surveillance.
• Current anonymization techniques do not adequately address gait characteristics for robust identity concealment.
• Future privacy solutions must explicitly target gait features or employ domain-adaptation techniques to enhance protection.