Anonymising Video Data Collection at the Edge Using DeepDish

Abstract

Increase in popularity of various deep learning methodologies has led to a huge increase in demand for training data. These training data often contain video featuring people but these videos often cannot be used due to privacy concerns as people are identifiable. We present an extension to the DeepDish object tracking system running on a Raspberry Pi. DeepDish performs object detection in the Raspberry Pi itself, avoiding the need to ship raw video data off the device. Our extension adds online face detection and blurring to preserve privacy. We implement and evaluate four different face detection algorithms, achieving 83% accuracy using YuNet, while increasing the latency by 75%. We then implement and evaluate three different object tracking algorithms to reduce the latency increase by running the simpler object tracking algorithm, reducing the number of calls to the expensive face detection algorithm. The result is a significant decrease in latency at the cost of a small decrease in accuracy. We also implemented and compared three different face obfuscation algorithms, one of which achieves differential privacy.