What is it about?
This study proposes a two-stage semi-supervised deep learning framework to detect and locate false data injection attacks (FDIAs) in smart grids. Such attacks can stealthily alter measurement data, misleading the power system’s state estimation and possibly triggering large-scale blackouts. The proposed method first performs unsupervised detection of abnormal data patterns using a ResNet-based Deep Support Vector Data Description (Deep SVDD) model. Once an anomaly is detected, a fine-grained localization stage identifies which sensors or buses were compromised, using shared-parameter 1D-ResNet networks to speed up training. Case studies on IEEE 14-bus and 118-bus systems show that the approach achieves an F1-score near 0.99, outperforming state-of-the-art deep learning methods like CNN, GGNN, and Transformer-based models.
Featured Image
Photo by Matthew Henry on Unsplash
Why is it important?
Cyberattacks on smart grids are rare but extremely destructive, and traditional detectors often fail when attackers manipulate data carefully enough to stay below alarm thresholds. Existing learning-based detectors assume frequent attacks and balanced datasets, which do not reflect reality. This paper addresses that gap by leveraging the natural data imbalance of real systems—using abundant benign samples to learn normal patterns and detect deviations. The result is a robust, efficient, and label-light framework that accurately detects and pinpoints stealthy FDIAs with minimal human supervision, enhancing the resilience and reliability of future intelligent power systems.
Perspectives
This work demonstrates a paradigm shift from purely supervised classification to semi-supervised anomaly learning for cyber-physical security. The two-stage design enables both global anomaly awareness and localized fault diagnosis, making it adaptable to large-scale grid deployments. Future extensions will explore physics-informed neural networks and federated defense architectures to handle parameter uncertainty, communication latency, and privacy constraints in real-world grid environments. Ultimately, this research contributes to creating secure, data-driven smart grids capable of defending against evolving cyber threats.
Professor/Clarivate Highly Cited Researcher/Associate Editor of IEEE TSG/TII/TSTE Yang Li
Northeast Electric Power University
Read the Original
This page is a summary of: Two-stage identification of false data injection attacks in power systems via semi-supervised deep learning, Applied Soft Computing, December 2025, Elsevier,
DOI: 10.1016/j.asoc.2025.113672.
You can read the full text:
Contributors
The following have contributed to this page
