Comparative Performance Analysis of Neural Network–Based Techniques for Botnet Detection in Smart Grid Cyber-Physical Systems

Abstract

The increasing integration of information and communication technologies into smart power grids has significantly improved operational efficiency but has also introduced critical cybersecurity vulnerabilities. Among emerging threats, botnet attacks pose a serious risk to smart grid cyber-physical systems due to their distributed, adaptive, and coordinated nature. This study presents a comparative performance analysis of three neural network–based techniques, Artificial Neural Networks (ANN), Deep Neural Networks (DNN), and Long Short-Term Memory (LSTM) networks, for detecting botnet-induced anomalies in smart grid environments. To guarantee fairness and repeatability, the models were trained and assessed using a publicly accessible smart-grid cyberattack dataset under similar preprocessing and experimental settings. Standard assessment measures, such as accuracy, precision, recall, F1-score, convergence behavior, and confusion matrix analysis, were used to gauge performance. The findings show that DNN produces better accuracy and overall classification stability through richer feature representations, whereas ANN offers a robust baseline with good recall. Because it can incorporate temporal relationships in smart-grid data, the LSTM model performs superior to both ANN and DNN on all measures, exhibiting greater accuracy and balanced detection capabilities. These results show that temporally aware deep learning models provide notable benefits for detecting coordinated botnet attacks and emphasize the significance of model architecture choices in cyber-physical intrusion detection. The study provides valuable insights for researchers and practitioners seeking effective data-driven security solutions for protecting critical smart grid infrastructure.