Do Auto-Regressive Models Protect Privacy? Inferring Fine-Grained Energy Consumption From Aggregated Model Parameters

We investigate the extent to which statistical predictive models leak information about their training data. More specifically, based on use case of household (electrical) energy consumption, we evaluate whether white-box access auto-regressive (AR) trained such data together with background information, as aggregates (e.g., monthly billing information) and publicly-available weather data, can lead inferring fine-grained any particular household. construct two adversarial aiming infer consumption patterns. Both threat target households. The second adversary has AR model for a cluster households containing Using real-world datasets, demonstrate that this apply maximum a posteriori estimation reconstruct daily significantly lower error than first adversary, serves baseline. Such essentially expose private occupancy levels. Finally, differential privacy (DP) alleviate concerns in dis-aggregating Our evaluations show differentially parameters offer strong protection against moderate utility, captured terms fitness cluster.