Novel K-neighbors and gradient boosting based transfer features for optimizing network attack detection

Network attacks represent unauthorized activities on digital networks that aim to disrupt, damage, or gain unauthorized access to computer systems. These attacks range from denial-of-service (DoS) disruptions to sophisticated intrusion attempts, posing significant threats to the integrity, confidentiality, and availability of information resources. The repercussions of network attacks are extensive, leading to substantial financial losses, compromise of sensitive data, erosion of user trust, and potential damage to an organization’s reputation. This research aims to recognize network attacks using advanced machine learning approaches and transfer feature engineering. We utilized the benchmark CICDDOS2019 database for experiments. We balanced the preprocessed dataset during the initial analysis using the synthetic minority oversampling technique (SMOTE). We introduced an innovative transfer feature generation method that generates salient class probabilistic fused features using the K nearest neighbors classifier (KNC) and gradient boosting classifier (GBC). The proposed KNGB (KNC-GBC) model extracts transfer features, which are subsequently fed into machine learning models to identify network attacks. The experimental results indicate that, using KNGB features, the random forest model attained a high accuracy of 98% for network attack recognition. We also validated the applied machine learning model’s performance results using k-fold cross-validations and improved performance using hyperparameter tuning. Furthermore, we determined the computational complexity of the methods used and conducted a statistical t-test analysis to validate the significance of the results. This research aims to provide a robust, machine learning-based framework that can adapt to the dynamic nature of cyber threats, ensuring the integrity and security of information networks.