EdgeBox: Confidential Ad-Hoc Personalization of Nearby IoT Applications

Steady progress in ubiquitous technologies and machine learning (ML) facilitates ever-new and better personalized IoT applications (e.g., digital assistants). However, current architectural concepts for personalization are either impractical or cannot support both sides without leaking sensitive data, namely the user's personal data or the provider's protected ML models. In this paper, we propose a novel approach termed EdgeBox for ad hoc personalizing nearby IoT applications while allowing both, (1) keeping the providers' ML models confidential and (2) privacy-preserving processing of personal user data, even on untrusted IoT devices. In short, confidential parts of the EdgeBox (e.g., the personalization task) are performed in an isolated trusted execution environment (TEE) within the IoT device's processor; the remote attested code in the TEE allows to confidentially load latest provider models and stream the required user data over wireless direct link ad-hoc protocol on demand. We evaluate our implemented proof-of-concept prototype (ad-hoc-initialized vs. pre-initialized) w.r.t. setup time and performance overhead. The results show real-time capabilities with an acceptable overhead, enabling a novel and confidential way to ad hoc personalize a user's dynamic environment in everyday life, supporting several assistant use cases.