Security Threat Analysis in 5G Cognitive Radio Networks: A Deep Learning Ensemble Approach

In reference [4], GRU -SVM-based threat detection system for CRNs. This approach combines the strength of GRU (Gated recurrent units) and SVM (support vector machine) to effectively notice the threats in CRNs. The GRU-SVM-based system has demonstrated improved detection accuracy and reduced false alarms compared to traditional threat detection approaches. The system effectively detected various types of security threats in CRNs, including primary user emulation attacks as well as spectrum sensing and data falsification attacks. In CRN, the GRU-SVM model has an 82.45 percent testing accuracy and an 80.99 percent training accuracy in identifying hostile users.

Benazzouza et al. [5] propose a novel prediction model that combines stacking and DL techniques for malicious user detection and spectrum sensing. The authors use a dataset of CRN simulations to evaluate the proposed model. The data set includes features like signal-to-noise ratio, signal strength, and modulation type. The outcomes determine that the suggested model achieves a higher detection accuracy of 95.6 percent and a low false alarm rate (2.1) for malicious user detection. This proposed model provides an effective solution for malicious user detection and spectrum sensing in CRNs. The combination of stacking and DL techniques improves the detection accuracy and reduces false alarms.

CRN [6] are vulnerable to SSDF (spectrum sensing and data falsification) attacks, which can compromise the security and reliability of the network. For SSDF attack detection in CRNs, the authors suggest a unique hierarchical ensemble extreme learning machine (HCME-ELM) based on cats and mice. Three steps make up the suggested approach: extreme learning machine, hierarchical ensemble, and data pre-processing. The authors evaluate the proposed HCME-ELM method using a simulated CRN network. The result shows that the proposed method achieves high detection accuracy (97.5) and low false alarm rate (2.1) for SSDF attack detection. The HCME-ELM method provides an effective solution for SSDF attack detection in CRNs. The detection accuracy is increased and the false alarm rate is decreased by using the hierarchical ensemble technique using ELM as the basic classifier.

The IoT (Internet of things) and 5^th generation networks are vulnerable [7] to various types of intrusions including malware, ransomware, and denial of service (Dos) attacks. These assaults are not detectable by conventional security measures like intrusion detection systems and firewalls. To identify breaches in IOT and 5G networks, the authors suggest a cognitive security framework that makes use of the DL approach. The authors evaluate the proposed framework using a dataset of network traffic data from IoT and 5G networks. The findings demonstrate that the suggested framework detects intrusions with a low false positive rate of 2.1% and a high accuracy of 95.6%.

Reference [8] presents a novel framework, Optimal Deep Learning Empowered Malicious User Detection for Spectrum Sensing (ODL-MUDSS), designed to automatically identify and classify malicious users (MUs) in Cognitive Radio Networks (CRNs). The ODL-MUDSS model leverages deep belief networks (DBNs) to detect MUs with high accuracy. To further enhance the recognition performance of DBNs, the model incorporates the sand cat swarm optimization (SCSO) algorithm, leading to improved detection results. The performance of the ODL-MUDSS technique is extensively validated through various experiments. The comprehensive evaluation results demonstrate the superiority of the ODL-MUDSS model over existing approaches, achieving exceptional performance metrics, including: Accuracy: 97.75%, Precision: 97.75%, Recall: 97.75%, F-score: 97.75%.

The proposed ensemble model harnesses the synergistic strengths of LSTMs and GRUs to enhance the precision and resilience of sequential data modeling. By integrating the long-term memory capabilities of LSTMs with the efficient gating mechanism of GRUs, the ensemble model can more effectively capture intricate patterns and relationships in sequential data. This section examines the LSTM&GRU ensemble approach's MU identification results. We employed our unique datasets, which were specifically designed and compiled for the experiments. The information about the dataset is outlined in Table 1. We prepared datasets of 10,000 samples. We conducted several thorough simulations to assess how well the suggested approach performed [14]. Thirty percent of the dataset is put aside for testing throughout the training phase, while the rest of the 70 percent is utilized for training. The model is trained utilizing 7000 data samples, and 3000 data samples are chosen for model testing. Table 1 demonstrates the sample dataset of characteristics, including modulation scheme, power level, frequency band, transmission time, entropy, SNR, and energy of the signal. The data set is labeled as malicious and primary users according to the set parameters. The frequency band used for this study is 2.4GHz.

The proposed model's performance in detecting MUs is evaluated using confusion matrices, as illustrated in Figures 4 and 5. These matrices facilitate a comprehensive evaluation of the model's accuracy, pinpointing its strengths and weaknesses. In the context of binary classification, the confusion matrix results comprise four essential variables, with true positives and true negatives being vital indicators of performance [15].

4.png

Figure 4. Confusion matrix of testing

5.png

Figure 5. Confusion matrix of training

The following measures are used to assess the model's performance:

True Positive (TP): Actual positive occurrences are appropriately predicted by the model as positive. Malicious users are effectively identified as such in this suggested strategy.

True Negative (TN): The actual negative instance is properly predicted by the model to be negative. Primary users are successfully identified as primary users in this model.

False Negative (FN): Primary users are mistakenly thought to be malicious.

False Positive (FP): The Malicious users are incorrectly identified as the primary users.

Table 1. Data set model

The goal of this DL models is to identify complex patterns as well as relationships within training data, enabling accurate predictions and classification decisions. Evaluating precision, accuracy, recall, as well as F-score is essential to ensure optimal model performance, as these metrics offer a comprehensive understanding of model strengths, weaknesses, and opportunities for refinement.

Accuracy: The percentage of correctly categorized nodes relative to the total number of nodes evaluated is known as accuracy. Its definition is as follows:

Accuracy $=\frac{T P+T N}{T P+T N+F P+F N}$      (5)

Precision is a quality metric when assessing a machine-learning model. When making positive predictions, it gauges how accurate the model is. It may be explained as:

Precision $=\frac{T P}{T P+F P}$              (6)

An ML model can be assessed as a quantity metric. It computes the overall number of successful positive predictions made out of all possible positive predictions. It is characterized as Recall.

Recall $=\frac{T P}{T P+F N}$        (7)

The balance between precision as well as recall is evaluated by the F score, that is the harmonic mean of both precision and recall. The ideal value of the F_score is 1.

$\mathrm{F}_{\text {score }}=\frac{2 *(\text { Precision } * \text { Recall })}{(\text { Precision }+ \text { Recall })}$      (8)

MCC measures the correlation between predicted and actual classes.

$M C C=\frac{(T P * T N-F P * F N)}{\sqrt{(T P+F P)(T P+F N)(T N+F P)(T N+F N)}}$      (9)

Table 2 demonstrates the performance matrices of the recommended model. Comparing the models using a curve-based method, such as the PR (Precision-Recall) curve, is another way to explore the various models' classifier performance. Figures 6 and 7 illustrate the suggested model's P-R curve. The precision-recall curves focus on the trade-off between precision & recall. The PR curve uses the x-axis to show the recall value and the y-axis to show the precision value. A higher recall indicates fewer false negatives, whereas a higher accuracy value indicates fewer false positives.

Table 2. Performance calculations of training data

6.png

Figure 6. P-R Curve of testing model

7.png

Figure 7. PR Curve of training model

A steep precision-recall curve indicates a high-performance classifier. The average precision (AP) is the summary metric that measures the classifier's performance across all recall values. A higher AP indicates a better classifier. In the proposed model, the AP value of the primary user is 0.93, and for malicious users, AP=0.99 A high-quality classifier will always have an AP value greater than 0.5 or equal to 1. Therefore, the suggested model provides a superior classification.

The performance of a classifier is visually represented by the ROC curve, that contrasts the true positive rate and false positive rate across distinct thresholds [16]. This plot provides another way to visualise the performance of a classifier [17]. The percentage of real positive cases that the classifier accurately recognizes is evaluated by the true positive rate. A good classifier, requires a high TPR preferably greater than 0.9. The proposed model classifier shows a good TPR for testing and training models.

The classifier's efficacy is evaluated using the AUC metric, demonstrating a strong link between scores (0.5-1) and accuracy. Figures 8 and 9 illustrate the proposed LSTM-GRU model's exceptional performance, achieving AUC values of 0.992 (training) and 0.9907 (testing) with an 80:20 train-test split, outperforming other models. Moreover, training accuracy and validation accuracy are crucial parameters, indicating the model's reliability. The training and validation accuracy plots in Figures 10 and 11 provide insight into the model's performance. A well-performing model should achieve high accuracy in both training and validation phases, indicating good generalization and robustness [18]. By dividing the total number of training samples by the number of correct predictions, the training accuracy has been determined. To calculate validation accuracy, divide the number of correct predictions by the total number of validation samples [19]. Separate datasets are used for validation accuracy, which are not used during training. Validation accuracy is a better indicator of the model performance on real-world data. The accuracy of the training and validation sets is summarized in Table 3. As the model undergoes more training epochs, its validation accuracy consistently improves, outperforming other models and showcasing its enhanced capabilities.

8.png

Figure 8. ROC curve of the testing model

9.png

Figure 9. ROC curve of training model

10.png

Figure 10. Training & validation accuracy

11.png

Figure 11. Training & validation loss

Table 3. Training & validation accuracy and loss

Overfitting is the reason why training accuracy is usually higher than validation accuracy. For a model to perform well, both validation and training accuracy should be high, and the margin between the two should be minimal. A low training loss suggests that the model is fitting in the training data effectively. Loss of training is a measure of the proposed model's error. Usually, validation loss exceeds training loss [20]. The average loss over the validation samples on a separate validation data set is used to calculate validation loss. Table 3 demonstrates the accuracy as well as training loss along with validation data for different epochs. A comparison of the LSTM-GRU model with existing methods is presented in Figure 12. The simulation results reveal that the GRU-SVM model underperforms, whereas the GRU-LSTM model achieves enhanced performance in detecting malicious users in cognitive radio networks (CRN).

12.png

Figure 12. Comparison of different models