Learning Without Adversarial Training: A Physics-Informed Neural Network for Secure Power System State Estimation under False Data Injection Attacks

A team of researchers from Cyprus, led by Solon Falas, has developed a novel Physics-Informed Neural Network (PINN) for secure power system state estimation (PSSE). The model is specifically designed to withstand stealth-constrained AC False Data Injection Attacks (FDIAs), a growing threat as power grids become more digitalized and communication-intensive. Unlike existing neural network-based approaches that require adversarial training, this PINN embeds power-flow consistency directly into its learning objective, using a dynamic loss-weighting formulation based on homoscedastic uncertainty. This technique automatically learns the relative scaling of supervised data-fit and physics-residual terms during training, reducing sensitivity to manual weight tuning and improving robustness.

The model was rigorously evaluated on the IEEE 118-bus system, a standard benchmark for power systems research, against four representative FDIA families: state distortion, load redistribution, line overloading, and residual-constrained stealth corruption. Performance was measured using Mean Absolute Error (MAE) on voltage magnitudes and phase angles. Results showed that the proposed PINN achieves higher accuracy and stability than existing fixed-weight PINN variants across all attack scenarios. This work, published on arXiv (2604.22784), highlights the potential of physics-informed machine learning for securing critical infrastructure without the computational overhead of adversarial training.