A Kernel-Based Method for Resolving Performance Inefficiencies in Mining Frequent-Patterns in Encrypted Data

Abstract

Big-data analytics is increasingly important in today's data-centric world. In this context, data encryption is a powerful tool for storing and analyzing private data. In particular, fully homomorphic encryption (FHE) is a promising encryption method that allows the analysis of encrypted data without need for decryption. FHE therefore enables users to outsource data storage and processing to a public cloud system without disclosing their data. However, FHE significantly increases data size and processing time, thus making it essential to improve performance in both I/O and processing. In most cases, the behavior of CPU resource consumption can be monitored and understood from code structure and logic. On the contrary, I/O resource consumption, which is controlled by the operating system kernel, is much harder to observe and understand. This paper addresses this issue in the context of a widely used data-analytics technique for secure frequent-pattern mining, called FHE Apriori. First, we propose a method for observing the I/O requests of FHE Apriori by modifying the operating system kernel. Second, we use the proposed method to characterize the I/O behavior of FHE Apriori and identify inefficiencies of storage access (that can be addressed to improve performance). Third, application-level changes based on this identification are described, enabling prefetching of data at runtime before they are needed. Fourth, the benefit of the described changes is quantitatively evaluated, showing that application performance improves by 23%.