Cyber-Security Audit for Smart Grid Networks: An Optimized
Detection Technique Based on Bayesian Deep Learning Alexander N. Ndife1, Yodthong Mensin1+, Wattanapong
Rakwichian1, and Paisarn Muneesawang2
1School of Renewable Energy and Smart Grid Technology 2Department of Electrical and Computer Engineering Abstract Security of computers, networks and
their communication protocols are vital in smart grid technology operation
and its management. This paper discusses a proposed Bayesian Neural Networks
for time-series TCP/IP packets intrusion detection and threats classification
in a grid network. This architecture termed SGtechNet detects invariants with
maximized detection accuracy by applying a robust method that approximates
the variation in posterior weights of neural networks with variational
inference to minimize the divergence between prior and true network posterior
distributions. Spatiotemporal feature engineering and uncertainty estimation
in Bayesian modeling, were leveraged to learn novel attack features and classify
attacks accordingly. This architecture reduced the size of the proposed model
to 25 % of the size of a pioneer model (AlexNet), hence, facilitating the
inference time compared to the baseline. SGtechNet was tested on NSL-KDD
datasets using two deep learning algorithms: CNN-LSTM and GRU, on two
classification categories (binary and multiclass) with Accuracy, Precision,
Recall, and F1-Score as the performance metrics. GRU algorithm comparatively
performed moderately well on both classification categories, unlike CNN-LSTM
that performed convincingly only on one test category. Comparing the result
of the SGtechNet model against a comparator model showed outstanding
performance in both model size, computational speed, and marginal improvement
in terms of accuracy. Chi-Square Test analysis determined that the degree at
which the training accuracy differed with validation accuracy was
statistically insignificant. Keywords: Bayesian, Cyber Threat,
Classification, Deep Learning, Intrusion detection, Neural Networks. +: Corresponding author: Yodthong Mensin Journal of Internet Services and
Information Security (JISIS), 12(2): 95-114, May 2022 |