Privacy-preserving distributed clustering: A fully homomorphic encrypted approach for time series

Abstract

In time series analysis, particularly in domains like smart metering, the drive for accurate predictions often depends on access to fine-grained, sensitive data. This need raises significant privacy concerns, especially in distributed data environments. To address these challenges, we apply the LINDDUN privacy threat modeling framework to identify and formalize privacy risks, and establish privacy requirements specific to distributed clustering of time series data. We extend the framework by integrating system design assumptions early on, and derive new attack trees that align with current threat patterns. We propose a distributed clustering protocol based on fully homomorphic encryption, and further enhance privacy guarantees by integrating differential privacy mechanisms and a software-based local caching strategy to bound computational costs. In the context of smart metering, assuming a semi-honest model where agents adhere to the protocol without collusion, our simulation results indicate a favorable trade-off between privacy and performance at $ϵ\simeq 3.0$. Our approach offers a blueprint for designing privacy-first systems that enable accurate predictions while safeguarding individual privacy.