I See and I Forget: Aged Data Removal for Continuous Data Sensing Systems
Zhuan Chen, Kai Shen · UR Research (University of Rochester)
Continuous data collection and processing is important for emerging application areas like cognitive assistance, smart sensing, and intelligent transportation. These systems may collect and buffer sensitive data including human-identifying features and privacy-concerning (e.g., health) information. We perform a full-system (including the OS) dynamic tainting study which finds that sensitive data may stay long in the system and be repeatedly accessed. However, accesses to old data are not critical for the correctness and data processing semantics of many applications. This paper presents a new data protection approach for mobile and field-deployed data sensing devices that guarantees the removal of sensitive data and derived information over a certain age (i.e., time since arrival or creation). This is realized by periodic virtual machine (VM)-enabled clean system renewals, during which we switch the data input and production operation to a new clean VM and then securely delete the previous production VM. We identify characteristics in continuous data sensing applications that maintain correct execution through clean system renewals. We further propose a shadow warmup mechanism that lets the renewed system use limited recent input data to produce accurate processing outputs. Experiments with our QEMU/KVM-based prototype implementation and several human and transportation sensing applications demonstrate the applicability of aged data removal at acceptable overhead (no more than 6% slowdown compared to virtualized original execution, or below 19% slowdown compared to native execution). The energy cost is no more than 10% compared to native execution.