Person identity shift for privacy-preserving person re-identification

Shuguang Dou, Xinyang Jiang, Qingsong Zhao, Yansen Wang, Dongsheng Li, Cairong Zhao

Abstract: Recently, privacy concerns of person re-identification (ReID) have raised more and more attention, and protecting personal information in the privacy-sensitive images used by ReID methods has become essential. In order to utilize data from video surveillance without leaking pedestrians’ private information, person de-identification (DeID) is a simple and effective method of alleviating privacy issues by removing identity-related information from the data. Most of the existing DeID methods focus on identity-irrelevant tasks such as pose and action recognition and tend to remove all identity-related information. However, this compromises the usability of de-identified data in the ReID task. In this paper, we aim to develop a technique to achieve a good trade-off between privacy protection and data usability for person ReID. To achieve this, we propose a novel de-identification method designed explicitly for person ReID, named person identity shift (PIS). PIS removes the absolute identity in a pedestrian image while preserving the identity relationship between image pairs by exploiting the interpolation property of the variational auto-encoder. Experimental results show that our method has a better trade-off between privacy-preserving and model performance compared to existing de-identification methods and can defend against human and model attacks for data privacy. The codebase of PIS is available at https://github.com/Vill-Lab/2025-SCIS-PIS.