IEEE Consumer Electronics Magazine - January/February 2023 - 121

Figure 4. Deep learning-based intrusion detection system.
single channel, while the VGG-16 architecture
accepts a fixed image with three channels input,
i.e., red, green, blue (RGB). Thus, the grayscale
images are converted into RGB with a fixed resolution
of 224 224 pixels. The binary data are represented
as grayscale images; every byte resembles
one pixel color of the image interpreted as a grayscale.
For black and white colors, 0 and 255 are
used,whereas other values are utilized for intermediate
colors of gray. These shades manifest visual
patterns utilized to interpret binary data displayed
as grayscale graphical depictions. It assists in identifying
different data regions and, therefore, helps
us to detect attacks or analytical tasks. Furthermore,
as suggested by Krizhevsky et al.,13 we subtracted
the mean RGB measured on the ImageNet
dataset. These RGB pixel values are utilized as
input features. We trained the VGG-16 architecture4
by transferring the pretrained the VGG-16
architecture's convolutional layers to the newly
trained layers. The parameters of assigned convolutional
layers are implemented to obtain the bottleneck
features. Finally, the fully connected layers
are used for the detection of different attack messages.
To overcome the spatial dimension, the
max-pooling layers are added at the end of convolutional
layers. The features for input are given as
VðMÞ
i
¼ BðMÞ
i þ
nX
ðM1Þ
1
'M
k¼1
In the abovementioned equation, ViðMÞ represents
an output layer; the base value is represented
with BiðMÞ, 'M
i;k indicates the filter mapping of the
kth feature value, while hk determines the M 1
output layer. We then applied the fully connected
January/February 2023
i;k hM1
k
:
(1)
layers to classify different types of sample messages.
Finally, the loss function is defined as a
type of content loss4 given as
Lfi;j
ð^y; yÞ¼
1
Wi;j;Hi;j
XWi;j
x¼1
fjð^yÞx;yÞ2:
XHi;j
y¼1
ðfiðyÞx;y
(2)
In the abovementioned equation, fi;j shows the
feature map received after activation from the
jth convolution and before the ith max-pooling
layer. The loss function of the VGG architecture
is the Euclidean distance between the feature
maps of an output image fiðyÞ and the actual
image fjð^yÞ. The corresponding dimensions of
feature maps' are represented with Wi;j and Hi;j.
EXPERIMENTAL RESULTS
This section elaborates on the experiments
and performance evaluation of the abovementioned
presented intrusion detection system.
During this analysis, CAN-intrusion-dataset is
used. We used three different kinds of CAN message
samples, i.e., DoS, fuzzy, and attack free
states (normal), from the dataset. The details of
the dataset are provided in Table 1.
The dataset is randomly splitted at the ratio of
70% and 30% for training and testing. The system
achieves good results, with an overall detection
accuracy of 95% for DoS, fuzzy, and 96% for attack
free state samples, as shown in Figure 5. The precision,
recall, and F1-score of the system for different
kinds of attacks and normal samples are
also shown in Figure 5. The system's accuracy
shows that the developed intrusion detection
121
IEEE Consumer Electronics Magazine - January/February 2023

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - January/February 2023
