Evaluating Privacy Measures for Load Hiding
Vadim Arzamasov, Klemens Böhm
TL;DR
This work tackles privacy evaluation for load-hiding in smart grids, where household consumption data can reveal sensitive information. It develops a framework to test 25 proposed privacy measures against natural properties (monotonicity, symmetry, boundary cases) and uses both real (Smart* and CER) and synthetic data to study how well measures reflect stealth of an appliance-usage secret. The study finds that only five measures pass most tests, with a variant of mutual information, $MI^i$, standing out as most effective for detecting appliance-usage leakage; estimator choice (histogram vs nearest-neighbor) significantly affects results. The findings provide a practical blueprint for selecting privacy measures in load-hiding research, favoring $MI^i$ (with NN when possible) and offering guidance on when to use $ ext{Delta}R_2^2$ or $R_2^2$ depending on adversary knowledge. The work has implications for standardizing privacy assessment in smart-meter deployments and for designing load-hiding schemes with reliable privacy guarantees.
Abstract
In smart grids, the use of smart meters to measure electricity consumption at a household level raises privacy concerns. To address them, researchers have designed various load hiding algorithms that manipulate the electricity consumption measured. To compare how well these algorithms preserve privacy, various privacy measures have been proposed. However, there currently is no consensus on which privacy measure is most appropriate to use. In this study, we aim to identify the most effective privacy measure(s) for load hiding algorithms. We have crafted a series of experiments to assess the effectiveness of these measures. found 20 of the 25 measures studied to be ineffective. Next, focused on the well-known "appliance usage" secret, we have designed synthetic data to find the measure that best deals with this secret. We observe that such a measure, a variant of mutual information, actually exists.