Auditing Privacy in Power Consumption Forecasting: Evaluating Gradient Leakage and Differential Privacy Defenses

ABSTRACT As machine learning models continue to advance power consumption forecasting and smart grid analytics, concerns about privacy risks in electricity time‐series data have gained significant attention. Highlighting the unique characteristics of electricity data, such as its high periodicity and direct correlation with sensitive real‐world activities, this paper introduces a novel privacy auditing framework that systematically assesses the vulnerability of electricity forecasting models to gradient‐based data reconstruction attacks, where adversaries exploit model gradients to infer sensitive electricity usage data, posing a serious privacy risk to households, industries, and power operators. This paper introduces a novel privacy auditing framework that systematically assesses the vulnerability of electricity forecasting models to gradient‐based reconstruction attacks. Unlike traditional methods that primarily focus on privacy mitigation through model regularization or noise injection, our framework provides a quantitative evaluation of gradient leakage, allowing us to measure and analyze the extent to which adversaries can reconstruct private energy consumption data. Through experiments on benchmark electricity datasets (ETTh1, ETTh2, and ECL), we demonstrate that, without privacy‐preserving techniques, attackers can accurately reconstruct electricity consumption patterns, as indicated by the rapid decrease in mean squared error (MSE) during reconstruction iterations. However, the application of a standard defense mechanism, differential privacy (DP), significantly disrupts this process, increasing MSE and coefficient variance, thereby limiting an attacker's ability to recover meaningful energy consumption information. By formalizing the connection between gradient similarity and time‐series similarity, our study quantifies privacy risks in electricity forecasting models, and evaluates the effectiveness of DP as a representative defense mechanism. The main contribution is not an improvement to privacy‐preserving algorithms themselves, but rather the auditing framework used to measure their efficacy. Our analysis provides a quantitative framework for navigating the critical tradeoff between privacy preservation and model utility, while also considering the computational overhead of the auditing process. This work provides a systematic approach to privacy auditing, equipping researchers, policymakers, and energy providers with insights into the trade‐offs between model accuracy and data security. Future research should explore additional privacy‐preserving techniques such as homomorphic encryption, federated learning, and adversarial training to further enhance privacy protections in smart grid and energy forecasting applications.