IEEE Consumer Electronics Magazine - January/February 2024 - 33

and cross entropy as loss function for 4000 epochs
andabatchsize of 128.
The objective function is minimized by precisely
estimating the coefficients through SGD
algorithm during the neural-net's training phase.
This was possible because of the backpropagation
procedure, the workhorse of SGD that cognitively
computes the objective function's gradient with
respect to the neural-net's key parameters, i.e., the
biases and weights. As the name suggests, the
backpropagation procedure iteratively computes
the loss-function's derivative for the coefficients
propagated fromprevious layers.
Figure 2 illustrates the history of neuralnetwork's
learning convergence of discriminating
attacks vectors from normal samples during the
training phase. It was observed that as the training
epoch prolonged, accuracy increased gradually
with the decline in themodel's overall loss, depicting
the model's efficiency in learning the pattern of
the input data with optimal convergence.
Hence, the increase in neurons per layer,
network's overall depth and sample volume are
directly proportional to the increase in computational
cost, which is observed as an essential
critic was elevated in this study by deploying
the learning model on IBM-Power9 Server cluster
with NVIDIA's K80-GPUs.
Evaluation Metrics
Research findings of intelligent intrusion detection
proposed in the literature claim their detection
supremacy based on the confusion-matrix scores.
True positive ðTPÞ that marks an actual attack as
an attack, true negative ðTNÞ that marks a normal
incident as normal, FP that marks a normal incident
as an attack, and false negative ðFNÞ that
marks an actual attack as normal incident are the
four main metrics of the confusion matrix. The
study carried out and presented in this research
article adopted the performance metrics presented
in Table 2 for evaluating the attack-detection proficiency
of the proposed network-stack model.
STACKED-MODEL'S PERFORMANCE
OVER HETEROGENEOUS
INDUSTRIAL DATASETS
The performance of the proposed neural-network
stack classifier was first evaluated using the
January/February 2024
TABLE 1. Structure followed by each deep-learning model.
Operation
Data
dimensions
Input
Dense
ReLu
#####
#####
#####
122
XXXXX ------ 9840
80
BatchNormalization m j s ------ 320
80
Dropout
Dense
ReLu
jj j ------ 0
80
#####
#####
60
BatchNormalization m j s ------ 240
60
Dropout
Dense
ReLu
jj j ------ 0
60
#####
#####
60
BatchNormalization m j s ------ 240
60
Dropout
Dense
sigmoid
jj j ------ 0
60
1
#####
XXXXX ------ 61
#####
data recorded at Mississippi State University
(MSU's)Mississippi State University Power-System
SCADA laboratory,10 which is equipped with
industry-scale components, such as power generators
ðG1G2Þ, intelligent electronic devices
ðR1R4Þ for switching breakers, and breakers
ðBR1BR4Þ. The entire setup has both a distance-based
protection scheme to trip breakers
and command-issuing modules for operators to
control IEDs manually. Figure 3 summarizes the
overall testbed deployed at the MSU-laboratory
where SCADA-traffic was logged under 37-scenarios:
8 events while the testbed operates normally,
1 event where the testbed is at idle state, 28 events
where attacks were simulated under short circuit
(fault), line-maintenance, remote-trip (command
33
0.3%
1.2%
0.0%
#####
XXXXX ------ 3660
19.0%
1.2%
0.0%
#####
XXXXX ------ 4860
25.3%
1.7%
0.0%
51.2%
Weights
(N)
Weights
(%)
IEEE Consumer Electronics Magazine - January/February 2024

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - January/February 2024
