Deep learning-based video anonymization for security and privacy

Abstract

In recent years, the design of systems has been heavily influenced by concerns about privacy and security. Companies have been particularly affected by the introduction of GDPR, which requires them to consider the privacy of individuals when storing personal data or facing significant fines. To address this issue, we have proposed a system that can detect and censor faces in a way that prioritizes privacy. Our system can process videos and automatically identify and blur facial features. We have customized our implementation using Facebook's object detector, Detectron2, and further developed it to enable facial detection and censorship. To evaluate the effectiveness of our approach, we conducted statistical tests and a small user study using videos that our system had censored. The results suggest that our implementation can reliably blur facial features to the point where the censored individuals are unrecognizable.