5. Experimental Results The essence of empirical validation lies in the rigor of experimental analysis, wherein
theoretical models confront the test of practical performance. This section presents a
detailed exposition of the experimental results derived from the evaluation of the ensemble
model, tailored for intrusion detection within the intricate framework of IoT-based EVCS.
Using a methodological approach, the model was subjected to various tests, ranging from
binary to multifaceted multiclass classifications. Each test was meticulously designed
to probe the model’s predictive prowess across a spectrum of scenarios that mirror the
heterogeneity of potential security breaches in IoT environments.
Binary classification trials were aimed at discerning the presence or absence of intru-
sion attempts, thus laying the groundwork for the model’s capability to distinguish between
normal operations and anomalies. Progressing to more granular levels, six-class and fifteen-
class classification tests were orchestrated to evaluate the model’s ability to identify specific
types of intrusions, each with its unique signature and implications (Table
3
).
Table 3. Model Performance Metrics.
Performance Metric 2 Class 6 Class 15 Class Test Loss
0.0000
0.0532
0.0632
Test Accuracy (%)
100
97.44
96.90
Epoch
6
50
50
Training time (s)
1885.46
14803.63
14719.47
Testing time (s)
42.53
42.20
40.65
5.1. Binary Classification Results
In an era where precision is paramount, the ensemble model demonstrates remarkable
proficiency in binary classification within the specialized sphere of IoT security for EVCS.
The model’s acumen, distilled through only six epochs, yielded a test loss imperceptible
to statistical significance and a test accuracy that epitomizes perfection (Figure
6
). Such
exemplary performance, encapsulated within 1885.46 s of training and only 42.53 s of
inferential judgment, heralds not only the model’s computational efficiency but also its
potential deployment in scenarios where the immediacy of threat detection is paramount.
The ensemble’s binary classification endeavor, delineating “No Intrusion” from “In-
trusion”, culminated in an exemplary synchronization with the ground truth, as evinced by
the congruence of precision, recall, and F1-score, each reaching the pinnacle of 1.00 for both
categories (Table
4
). This zenith of classification metric unanimity, seldom achieved in the
Mathematics 2024, 12, 571
16 of 26
intricate domain of cybersecurity analytics, highlights the model’s sophisticated capacity
to discern with meticulous exactitude.
