A Vulnerability in Video Anonymization – Privacy Disclosure from Face-obfuscated video

This work studies a vulnerability in face obfuscation techniques intended to preserve the privacy of individuals. There have been several attempts to prevent unauthorized face recognition from being performed, aiming to guarantee anonymity in video data. Most of these attempts have focused on facial areas that are thought as sensitive to contribute most to facial recognition. However, obfuscation of such facial areas is insufficient to preserve privacy because gait information such as arm movements and step characteristics can be used to identify individuals and other personal information such as gender. In this paper, we claim that individual tracking and gender estimation are possible just from the gait information extracted from a video without using face-related data. We propose a set of biometric features and an algorithm to estimate gender from skeleton data. Our experiments with more than 100 subjects demonstrate that gender is estimated with a significant accuracy of 99.86%. The proposed identification algorithm, which is based on pattern-matching techniques, is robust against changes in the manner of walking and successfully identifies subjects with only small error of 0.036.