Towards Understanding the Adaptation Space of AI-Assisted Data Protection for Video Analytics at the Edge

Edge computing facilitates the deployment of distributed AI applications, capable of processing video data in real time. AI-assisted video analytics can provide valuable information and benefits in various domains. Face recognition, object detection, or movement tracing are prominent examples enabled by this technology. However, such mechanisms also entail threats regarding privacy and security, for example if the video contains identifiable persons. Therefore, adequate data protection is an increasing concern in video analytics. AI-assisted data protection mechanisms, such as face blurring, can help, but are often computationally expensive. Additionally, the heterogeneous hardware of end devices and the time-varying load on edge services need to be considered. Therefore, such systems need to adapt to react to changes during their operation, ensuring that conflicting requirements on data protection, performance, and accuracy are addressed in the best possible way. Sound adaptation decisions require an understanding of the adaptation options and their impact on different quality attributes. In this paper, we identify factors that can be adapted in AI-assisted data protection for video analytics using the example of a face blurring pipeline. We measure the impact of these factors using a heterogeneous edge computing hardware testbed. The results show a large and complex adaptation space, with varied impacts on data protection, performance, and accuracy.