Figure 6.
Model accuracy and loss.
Table 4.
Classification Report.
Precision
Recall
F1-Score
Support
No Intrusion
1.00
1.00
1.00
279,968
Intrusion
1.00
1.00
1.00
109,122
accuracy
1.00
389090
macro avg
1.00
1.00
1.00
389,090
weighted avg
1.00
1.00
1.00
389,090
The classification report and ensuing confusion matrix–both in their raw and nor-
malized states–serve as a testament to the model’s impeccable discriminative abilities
(Figures
7
and
8
). They exhibit an unequivocal dichotomy between normalcy and intrusion,
a dichotomy that is stark and devoid of the ambiguity that often plagues classification en-
deavors. This absolute bifurcation in the model’s predictive capabilities marks a significant
milestone in the quest for robust, fail-safe security systems in the burgeoning field of IoT.
Figure 7.
Confusion Matrix.
Mathematics 2024, 12, 571
17 of 26
Figure 8.
Normalized Confusion Matrix.
5.2. Six-Class Classification Results
The model’s ability to distinguish between the six nuanced threat landscapes is re-
flected in an impressive accuracy of 97.44%, a metric that stands as a testament to its
robustness and the veracity of its training (Figure
9
). This level of accuracy, particularly in
the complex and often chaotic environment of IoT security, speaks to the model’s sophis-
ticated feature extraction and classification capabilities. Although the test loss of 0.0532
indicates room for refinement, it remains a commendable figure given the intricacy of the
task at hand (Figure
9
). Further, the extended training duration of 14,803.63 s indicates
the model’s intensive learning process and the rapid testing time of 42.20 s underscores its
practical efficiency. This juxtaposition of extended training with swift testing is emblematic
of a model that, once trained, can offer real-time, reliable threat detection, which is crucial
for the active defense of IoT systems.
Figure 9.
Model accuracy and loss.
The proposed model exhibits exceptional precision in the “Normal” category, achiev-
ing perfect scores across precision, recall, and F1-score metrics (Table
5
). The “DDoS” and
“Scanning” categories also showed high metrics, demonstrating the model’s adeptness at
identifying these particular types of intrusions. Challenges surfaced in the “Injection” and
Mathematics 2024, 12, 571
18 of 26
“Malware” categories, where the precision and recall metrics indicated a greater difficulty in
class distinction, suggesting potential avenues for future research and model enhancement.
|
