INSPIRE: Instance-Level Privacy-Pre Serving Transformation for Vehicular Camera Videos

• DOI: 10.1109/icccn58024.2023.10230162
• Publication Date: 2023-09-01T17:22:54Z
• Authors: Zhouyu Li; Ruozhou Yu; Anupam Das; Shaohu Zhang; Huayue Gu; Xiaojian Wang; Fangtong Zhou; Aafaq Sabir; Dilawer Ahmed; Ahsan Zafar
• Keywords: Identification; Privacy Protection; Spoofing attack

The wide spread of vehicular cameras has raised broad privacy concerns. Ubiquitous capture bystanders like people or cars nearby without their awareness. To address concerns, most existing works either blur out direct identifiers such as vehicle license plates and human faces, obfuscate whole video frames. However, the former solution is vulnerable to re-identification attacks based on general features, latter severely impacts utility transformed videos. In this paper, we propose an INStance-level PrIvacy-pREserving (INSPIRE) transformation framework for camera INSPIRE leverages deep neural network models detect replace sensitive object instances in videos with non-existent counterparts. We design a modular enable flexible customization protected instance categories protection modules. An implementation focused protecting described, which tested six datasets three real-world evaluate its preservation capability. Results show that can thwart 97% while maintaining 0.75 detection mean average precision instances. also demonstrate experimentally robust against model inversion attacks. Compared solutions provide comparable protection, achieves relatively 1.76 times higher counting accuracy 31.61% precision.