Table 2 Performance comparison of ML models (n \(\rightarrow\) number of training instances, k \(\rightarrow\) number of clusters, I \(\rightarrow\) number of iterations, d \(\rightarrow\) number of features)
From: A survey on cognitive radio network attack mitigation using machine learning and blockchain
Parameters | SVM | KNN | Logistic regression | K-means | Decision tree |
|---|---|---|---|---|---|
Training time complexity | O(n3) | O(1) | O(k*n*d) | O(k*n*d*I*t) | O(n*d*log(n)) |
Prediction time complexity | O(m*n) | O(n*d) | O(d) | O(k*d) | O(log(n)) |
Detection accuracy | Moderate to high | Require careful tuning and feature selection for good results | Limited in complex scenarios | Depends on K and the features quality | Vary |
False positive rate | High when imbalanced data | Good at majority class and high for the minority class | Occur if the data are not well separated | Occur when the behaviour of normal user is classified as anomalous | Low when high risk of overfitting and high on unseen data |
Precision | Low when imbalanced data | Low if noisy data | Low if data has overlapping patterns | Assessed based on the accuracy of identifying true positives | High when accurately separate classes and low if they overfit training data |
Receiver operating characteristic | Deviate from ideal shape when imbalanced data | Do not produce traditional ROC | Achieves good when classes are well separated | Not directly applicable | Achieves good when classes are well separated |