Towards Privacy-Preserving Anomaly-Based Intrusion Detection in Energy Communities
Zeeshan Afzal, Giovanni Gaggero, Mikael Asplund
TL;DR
The paper addresses the security of energy communities (ECs) by proposing an anomaly-based intrusion detection system using a deep autoencoder trained on simulated EC data. It formalizes time-series anomaly detection with reconstruction error $e(X)$ and a threshold $\tau$, and it demonstrates a privacy-preserving implementation via federated learning (FL) to train the model without exposing raw data. Key contributions include an open-source EC dataset, an LSTM-based autoencoder IDS, and a federated learning demonstration showing comparable performance to centralized training while preserving privacy. The work highlights that such a privacy-conscious anomaly detector can enhance EC resilience in smart grids, though it acknowledges false negatives for certain attack types and points to avenues for optimization and real-world validation.
Abstract
Energy communities consist of decentralized energy production, storage, consumption, and distribution and are gaining traction in modern power systems. However, these communities may increase the vulnerability of the grid to cyber threats. We propose an anomaly-based intrusion detection system to enhance the security of energy communities. The system leverages deep autoencoders to detect deviations from normal operational patterns in order to identify anomalies induced by malicious activities and attacks. Operational data for training and evaluation are derived from a Simulink model of an energy community. The results show that the autoencoder-based intrusion detection system achieves good detection performance across multiple attack scenarios. We also demonstrate potential for real-world application of the system by training a federated model that enables distributed intrusion detection while preserving data privacy.