A Supervised Machine Learning and Temporal Convolutional Network Based Framework for Denial-of-Service Attack Detection and Mitigation

This article proposes a supervised machine learning (ML) based strategy for detecting Denial-of-Service (DoS) attacks and a Temporal Convolutional Network (TCN) based mitigation approach for synchrophasor data transmitted from Phasor Measurement Units (PMUs) to a Phasor Data Concentrator (PDC). The proposed work addresses the urgent need for real-time detection and reconstruction of DoS-affected data loss in synchrophasor networks, which is essential for secure and reliable grid operation. It enables deployment-ready detection with low computational overhead. After decoding and storing synchrophasor data for training, feature extraction is performed, followed by dimensionality reduction using Principal Component Analysis (PCA). Subsequently, the dataset is classified using various ML models with their hyperparameters selected via Randomized Search Cross-Validation (RSCV). A meta-learning layer is introduced that adapts to any dataset based on its statistical features and selects the best-performing model, ensuring accurate DoS attack detection and minimal detection time. The total meta-training time is approximately 3.4<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">s</i>, and detection time is around 1<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ms</i>. Further, a Euclidean distance-based trust score is used to trigger retraining in the presence of concept drift. Once a DoS attack is detected, TCN reconstructs the lost data using historical data within approximately 2<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ms</i>, mitigating adverse effects. The methodology is validated using a cyber-physical testbed simulating synchrophasor data flow from multiple PMUs to a PDC under various DoS scenarios. The outcomes indicate that the meta-learning layer selects the optimal model, and TCN effectively reconstructs missing phasors. The overall processing time (1<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ms</i>+2<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ms</i>=3<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ms</i>) remains within the packet reporting time of 16.67<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ms</i> for a 60 <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Hz</i> system.