Protecting Cross-Camera Person Re-Identification Data with Image Differential Privacy
Lucas Maris, Yuki Matsuda, Keiichi Yasumoto · 2024
To achieve smart cities, leveraging data from cameras, which are often readily-installed IoT devices, can offer precious insights on the behavior of pedestrians and play a crucial role in designing and maintaining efficient transportation, appropriate infrastructure, or attractive tourism facilities. Such pedestrian flow data collection is often achieved through cross-camera person re-identification. This task is heavily privacy-invasive task by design, benefiting from rich visual data, which thus carries highly sensitive personal details about individuals. We here study how image data can be protected upfront, and introduce a novel image differential privacy mechanism leveraging both pixelization and color quantization for this purpose. Our extensive experiments show that through its random noise additions, our mechanism can obfuscate data more effectively than standard image obfuscation methods while retaining high utility for cross-camera re-identification, preserving reasonable re-identification metrics and demographic information even under low privacy budgets.