Figure 14.
Normalized Confusion Matrix.
6. Discussion
We performed a comprehensive examination of the CNN-LSTM-GRU ensemble model
within the diverse and challenging domain of IoT security for EVCS. A comparative
analysis, as detailed in Table
7
, situates the ensemble model within the context of recent
advancements, delineating its standing against contemporary architectures in the field.
Table 7.
Comparison of Model Accuracies.
Model
Year
Accuracy (%)
2 Class
6 Class
15 Class
DNN [
33
]
2022
99.99
96.01
94.67
Inception Time [
34
]
2022
-
-
94.94
CNN-LSTM [
35
]
2022
100
98.69
-
VGG-16 [
36
]
2023
100
-
94.86
DeepAK-IoT [
37
]
2023
-
-
94.96
LNKDSEA [
38
]
2023
99.99
84.97
80.12
RNN [
39
]
2023
100
92.53
90.22
MAGRU [
40
]
2023
99.99
-
-
CNN-LSTM-GRU
2023
100
97.44
96.90
Mathematics 2024, 12, 571
23 of 26
In the binary classification domain, CNN-LSTM-GRU achieved parity with the unas-
sailable accuracy of its peers, where models such as CNN-LSTM [
35
], VGG-16 [
36
], and
RNN [
39
] also have perfect scores. This uniform excellence across models underscores a
maturing understanding and effective handling of binary classification tasks in the IoT
security domain.
The model’s success in achieving 100% accuracy in binary classification can be at-
tributed to the complementary strengths of its constituent architectures. The convolutional
layers effectively capture spatial hierarchies in the data, which is particularly useful in
identifying patterns indicative of intrusion within the IoT EVCS context. LSTM components
contribute to this high performance by capturing long-term dependencies, allowing for an
effective understanding of sequence progression in temporal data, a feature common in
network traffic. GRUs further refine the model’s capability by addressing the vanishing
gradient problem often encountered in recurrent networks, thereby enhancing the learning
process for long sequences without the need for extensive computational resources.
In the six-class classification scenario, CNN-LSTM-GRU displayed a notable accuracy
of 97.44%, surpassing most of its contemporaries and falling slightly behind CNN-LSTM’s
leading edge. This performance indicates CNN-LSTM-GRU robust feature extraction
and sequence learning capabilities, which are critical for distinguishing between a broad
spectrum of intrusion behaviors.
The fifteen-class classification, characterized by its intricacy and the granular distinc-
tion of intrusion types, demonstrated that CNN-LSTM-GRU maintained a high accuracy of
96.90%. This is a commendable achievement, especially when juxtaposed with DeepAK-
IoT [
37
] and Inception Time [
34
], which represent the upper echelon of performances in this
category. Notably, CNN-LSTM-GRU showed marked superiority over LNKDSEA [
38
] and
RNN [
39
], underscoring the efficacy of the ensemble approach in managing the increased
complexity of fine-grained classifications.
The CNN component of our ensemble model is primarily responsible for spatial
feature extraction. Unlike traditional models such as the DNN [
33
] and RNN [
39
], which
may lack depth in feature extraction, the CNN layers in our model provide a comprehensive
analysis of the input data’s spatial characteristics. This is evident in the binary classification
results, where our model matches the perfect accuracy of the CNN-LSTM [
35
] and the
VGG-16 [
36
], which are known for their strong feature extraction capabilities.
For temporal analysis, the LSTM and GRU components of our model are critical.
The LSTM layers capture long-term dependencies, while the GRU layers focus on shorter-
term data sequences. This dual approach allows our model to outperform traditional
architectures like the DeepAK-IoT [
37
] and LNKDSEA [
38
], particularly in the multi-class
classification tasks. It can recognize complex attack patterns that unfold over time, which
might be overlooked by models without this temporal depth.
These comparative outcomes not only validate the ensemble model’s capability but
also propel the discourse on the potential of hybrid models. The integration of multi-
ple neural network architectures may well be the harbinger of a new paradigm in IoT
security, where the complexity of threat detection is met with an equally sophisticated
analytical arsenal.
Moreover, the results present an impetus for the continued exploration of ensemble
methods in DL, pushing the envelope in terms of accuracy, adaptability, and computational
efficiency. As the digital infrastructure of IoT expands, the ensemble model’s adaptability
and learning depth will be pivotal in safeguarding the integrity and robustness of the
interconnected systems.
Considering these findings, the CNN-LSTM-GRU ensemble architecture emerges as a
potent architecture, heralding a promising direction for future research to further refine and
optimize DL strategies for intrusion detection, ensuring that they remain at the vanguard
of the ever-evolving cybersecurity landscape.
Mathematics 2024, 12, 571
24 of 26
|