Enforcing Privacy Preservation on Edge Cameras using Lightweight Video Frame Scrambling
Alem Fitwi, Yu Chen, Sencun Zhu
Abstract
Privacy protecting is a very challenging task in a highly surveilled world with zillions of surveillance cameras deployed. The difficulty mainly lies in the facts: (i) there is not a distinctively defined boundary between usability and privacy, (ii) video frames indiscriminately created and collected by the edge cameras could be abused and intercepted, and (iii) it is difficult to enforce the commonly used compute-intensive standard techniques as-is on the edge cameras because of limited computational resources. In this paper, we propose a lightweight and secure scheme to Enforce Privacy-preservation on Edge Cameras (EnPEC) using deep learning and a sinusoidal chaotic-map. The proposed EnPEC architecture comprises a lightweight frame classifier designed to label frames as offensive and harmless depending on their content to ensure the practice of selective surveillance following a frame approximation process and a novel sinusoidal-map-based chaotic image scrambling technique that enciphers frames color-channel wise to ensure end-to-end privacy of frame contents. The extensive analysis of the functionality, performance and security of the EnPEC scheme, and comparison with related works verify that the EnPEC scheme is more feasible, robust and secure when it runs in real-time on edge cameras equipped with computational power equivalent to the Raspberry PI 4.