# Glass-Box Deep Learning for FDIA Detection in Nonlinear Automatic Generation Control: A Kolmogorov-Arnold Network Approach

arXiv: https://arxiv.org/abs/2509.05259 · HTML: https://mlsanity.com/paper/2509.05259/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 5 of its 22 references that are in the map.
- First posted on arXiv 5 Sep 2025.

## Builds on (most cited first, 5 of 5)

- [KAN: Kolmogorov-Arnold Networks](https://mlsanity.com/paper/2404.19756/) (2024)
- [An Online Detection Framework for Cyber Attacks on Automatic Generation Control](https://mlsanity.com/paper/1712.06417/) (2017)
- [A Glass-Box Deep-Learning Method for Electrical Energy System Modeling Based on Kolmogorov-Arnold Network](https://mlsanity.com/paper/2409.08044/) (2024)
- [Machine Learning-Based Cyberattack Detection and Identification for Automatic Generation Control Systems Considering Nonlinearities](https://mlsanity.com/paper/2504.09363/) (2025)
- [Adversarial Attacks on Deep Learning-Based False Data Injection Detection in Differential Relays](https://mlsanity.com/paper/2506.19302/) (2025)

## Abstract (arXiv)

Automatic Generation Control (AGC) plays a critical role in maintaining power balance across multi-area power systems. However, its complete reliance on remotely communicated measurements makes it susceptible to cyber-induced False Data Injection Attacks (FDIAs), which can alter measurement values and destabilize system operation. Unlike prior studies that employ black-box Deep Learning (DL) models for FDIA detection, this paper proposes an interpretable and accurate Kolmogorov-Arnold Network (KAN)-based framework for detecting FDIAs in AGC systems, explicitly accounting for nonlinearities often overlooked in previous work. The proposed KAN model effectively identifies FDIAs in nonlinear AGC systems using only AGC measurements. Moreover, KANs inherently facilitate the extraction of symbolic equations, a capability absent in conventional DL models. After training, these equations can be directly utilized for FDIA detection, enhancing interpretability without compromising accuracy. The framework is trained offline to learn the nonlinear relationships among AGC measurements under diverse normal operating conditions and under FDIA scenarios that manipulate those measurements. Following training, pruning, and fine-tuning, symbolic expressions describing the model's decision logic are extracted. Both the trained KAN model and the extracted symbolic expressions are subsequently employed and evaluated for FDIA detection. Our results using a benchmark power system demonstrate that the proposed KAN-based framework and its associated symbolic representations accurately detect FDIAs targeting nonlinear AGC systems while preserving AGC reliability. The approach outperforms existing methods and provides a robust, interpretable mechanism for verifying AGC measurement authenticity against potential FDIAs.