Behavioral Fingerprints in Smart Meter Data: Interpretable Deep Learning for Socio-Demographic Inference
ACM SIGEnergy Energy Informatics Review · 01.09.2026 · 6 (3) · 15-31
DOI: 10.1145/3861130.3861132 ↗Kurz gefasst
Fachartikel in der Zeitschrift ACM SIGEnergy Energy Informatics Review, online erschienen am 11.11.2026. Verfasst von Dejan Radovanovic und 4 weiteren Autor:innen. Die Arbeit nennt 64 Literaturangaben.
Abstract (Originalsprache)
Smart meter data expose behavioral fingerprints that can reveal sensitive socio-demographic characteristics of households. Prior studies establish the feasibility of such inference, but typically remain limited to single datasets, fixed model settings, and prediction-oriented evaluations. This restricts their ability to assess whether inference risks persist across regions, time periods, and data collection contexts. We address this gap with a cross-dataset study that infers socio-demographic characteristics from one-week energy-consumption profiles of multiple households across three datasets from different regions and time periods. We benchmark three deep-learning time-series models: a CNN architecture from prior work, InceptionTime, and PatchTST. We evaluate both predictive performance and privacy leakage. The results show that inference performance depends strongly on the dataset and target characteristic. Attributes such as Family Home, Single, Employment Status, Age Person, Income Group , and Social Class exhibit relevant leakage significantly across parts of the cross-dataset evaluation. Beyond classification scores, we analyze how models exploit load-profile structure. Using Integrated Gradients, we examine the inference of Single status across all three datasets. The resulting attribution maps indicate that models use temporally localized consumption patterns, including weekday evening activity, weekend midday peaks, and low weekday night demand. These findings show that socio-demographic leakage is not a dataset-specific artifact. Weekly load profiles can expose privacy-relevant household information across heterogeneous settings. The results provide empirical evidence for privacy assessments of smart meter data and inform future analyses of privacy-utility trade-offs in load-profile applications.
Zitieren
APA
Radovanovic, D., Schirl, M., Kirchgasser, S., Eibl, G., & Unterweger, A. (2026). Behavioral Fingerprints in Smart Meter Data: Interpretable Deep Learning for Socio-Demographic Inference. ACM SIGEnergy Energy Informatics Review, 6(3), 15-31. https://doi.org/10.1145/3861130.3861132
Vancouver
Radovanovic D, Schirl M, Kirchgasser S, Eibl G, Unterweger A. Behavioral Fingerprints in Smart Meter Data: Interpretable Deep Learning for Socio-Demographic Inference. ACM SIGEnergy Energy Informatics Review. 2026;6(3):15-31. doi:10.1145/3861130.3861132
BibTeX
@article{Radovanovic2026,
author = {Radovanovic, Dejan and Schirl, Maximilian and Kirchgasser, Simon and Eibl, Günther and Unterweger, Andreas},
title = {Behavioral Fingerprints in Smart Meter Data: Interpretable Deep Learning for Socio-Demographic Inference},
journal = {ACM SIGEnergy Energy Informatics Review},
year = {2026},
volume = {6},
number = {3},
pages = {15--31},
doi = {10.1145/3861130.3861132}
}
EUVIMED