PDSProxy++: Proactive Proxy Deployment for Confidential Ad-hoc Personalization of AI Services

Abstract

Personal data stores (PDS) typically provide extensions for external confidential processing, allowing ad-hoc personalization of AI services on nearby (third-party) Internet of Things (IoT) devices. However, these extensions entail a high initialization overhead due to the underlying cryptographic mechanisms. While some approaches provide first optimizations by pre-initializing this confidential environment, it is unclear which devices need to be pre-initialized – too many unnecessary devices are inefficient, and ad-hoc initialization still takes too long, especially when the user is moving. In this paper, we tackle this initialization issue by proposing PDSProxy++—a PDS extension for proactive multi-hop deployment of AI services. Inspired by the human eye, PDSProxy++ is based on a central cone (foveal vision) and a surrounding smaller circle (peripheral vision), which determine the nearby IoT devices to be initialized. Using a city-wide, real-world smart street lamp dataset and emulations, we show the feasibility of PDSProxy++ and its efficiency: it outperforms the currently-practiced ad-hoc mode and other deployment baselines in different smart city scenarios.