Inferring the Hidden: Privacy Risks of Microaggregation in Smart Meter Data

Smart meter data, while essential for energy systems, pose significant privacy risks due to the behavioral information embedded in household electricity consumption patterns. Microaggregation has emerged as a promising anonymization technique to mitigate these risks. However, it remains unclear whether such aggregated profiles retain an identifiable structure that enables group membership inference while maintaining utility as it perturbs the data. In this paper, we present a replicable methodology to evaluate the trade-off between utility and privacy in micro-aggregated smart meter data. We assess utility through household-level day-ahead load forecasting and evaluate privacy by implementing an unsupervised group membership inference attack. The attack combines distance-based record linkage with a two-stage majority voting scheme and is applied across a range of anonymity levels ( k = 5 to 200) using both domain-specific features and deep neural representations. Our results reveal a utility-privacy trade-off: while forecasting accuracy degrades only moderately (maximum 14% loss), group membership inference remains highly effective at lower k values, with success rates up to 80 times higher than random guessing. These findings indicate that structural patterns persist through aggregation and can be exploited by adversaries, even without household-level identification, to enable targeted advertising, discriminatory profiling, or dynamic pricing. As such, microaggregation provides meaningful privacy protection at sufficiently higher k levels, underscoring the need for context-aware deployment in energy data sharing.