EPC: a video analytics system with efficient edge-side privacy control

Edge-cloud video analytics systems capture video streams by edge cameras and send the video streams to the cloud for analytics to support applications like video surveillance, VR/AR, autonomous driving, etc. Video streams captured at the edge may contain sensitive objects, e.g., a human being. Existing studies propose adding noise to the intermediate video analytics results, encrypting video frames, etc. In this paper, we take an orthogonal approach where we remove, a.k.a. denaturing, the sensitive objects at the edge side before sending a video frame to the cloud. Edge devices are highly resource-constrained, and the denaturing operation has non-trivial computation costs. More specifically, before denaturing, one needs to locate the sensitive objects by object detection; such object detection computation is resource intensive. In this paper, we propose EPC, an edge-cloud video analytics system that leverages a trajectory prediction model to locate sensitive objects in video frames. We formally analyze EPC and show that EPC can guarantee privacy. We evaluate EPC with two applications, person counting and vehicle detection. Evaluation results show that EPC can prevent privacy leakage under visual data attack with 95% video analytics accuracy and a 4x speedup compared to existing privacy control mechanisms.