GridEye: Secure UAV-Based Remote Monitoring Framework for Smart Grid's SCADA System

This paper presents a comparative analysis of deep learning models - 1D Convolution Neural Networks (1D-CNN), Long-Short-Term Memory (LSTM) and Gated Recurrent Units (GRU) - to classify UAV-based network attacks on smart grid supervisory control and data acquisition (SCADA) systems. The models were trained using Adam and RMSprop optimizer over 50 epochs. The results show that LSTM and GRU outperform 1D-CNN in terms of accuracy and generalization. Specifically, the GRU model with the Adam optimizer achieved a test accuracy of 97.6 %, precision of 0.976, recall of 0.976, and an F1 score of 0.976 with a test loss of 0.066. The LSTM model exhibited similar performance, with slightly more stable training and validation accuracy over time. The 1D-CNN model initially achieved decent performance with test loss of 0.198 and test accuracy of 92 % but demonstrated higher fluctuation in validation accuracy and loss. Overall, LSTM and GRU with the Adam optimizer have shown the best generalization capabilities.