Anonymization for Skeleton Action Recognition

Skeleton-based action recognition attracts practitioners and researchers due to the lightweight, compact nature of datasets. Compared with RGB-video-based recognition, skeleton-based is a safer way protect privacy subjects while having competitive performance. However, improvements in skeleton algorithms as well motion depth sensors, more details characteristics can be preserved dataset, leading potential leakage. We first train classifiers categorize private information from trajectories investigate leakage Our preliminary experiments show that gender classifier achieves 87% accuracy on average, re-identification 80% average three baseline models: Shift-GCN, MS-G3D, 2s-AGCN. propose an anonymization framework based adversarial learning dataset. Experimental results anonymized dataset reduce risk marginal effects performance even simple anonymizer architectures. The code used our available at https://github.com/ml-postech/Skeleton-anonymization/