FedDiSC: A computation-efficient federated learning framework for power systems disturbance and cyber attack discrimination

Detection of False Data Injection Attacks in Distribution Networks: A Vertical Federated Learning Approach

“…utilities, electricity providers, and system operators). This emerging scenario poses data-sharing challenges, as entities might be unwilling to share their data due to its sensitive, confidential and private nature [15]- [18]. Concerns about data-sharing further complicate the development of effective and collaborative FDIA detection models for smart grids.…”

“…Horizontal federated learning [19] has been utilized to address the aforementioned concerns by few studies [15]- [18] for FDIA detection. Federated learning can be classified into two main categories: horizontal federated learning (HFL) and vertical federated learning (VFL), based on the data partition scheme between participants.…”

“…The assumption is that the system is divided into sub-networks of equal size to have uniformity in measurements as required by their HFL-based framework, and each sub-network is equipped with an edge processor dedicated to decentralized FDIA detection based on HFL. In [18], a cross-silo HFL framework utilizing an autoencoder model is proposed. In [17], a cross-silo HFL framework with an incentive mechanism is proposed.…”

“…• First of all, the HFL-based studies [15]- [18] consider FDIA targeting state estimation at the transmission network where abundant measurements ensure the fulfilment of feature uniformity requirements. • These studies employing an HFL framework assume uniformity of measurement types at each bus [15] and additionally the equity in the number of buses within each sub-network [16]- [18]. Uniformity in measurements across all buses in the network cannot be guaranteed.…”

“…Furthermore, the assumptions made in these studies are not applicable to distribution networks due to the observability challenges (measurements are not abundant) and their inherent unbalanced multi-phase nature. • Lastly, it is important to note that these studies [15]- [18] solely process measurements at a single time step, overlooking the crucial temporal relationships between consecutive measurement data points. Given that distribution networks exhibit higher fluctuations compared to transmission systems, exploiting this temporal relationship significantly enhances the detection rate of attack samples within distribution networks.…”