A Hybrid Intrusion Detection Model Combining SAE with Kernel Approximation in Internet of Things

Area of Science:

• Computer Science
• Cybersecurity
• Machine Learning

Background:

• Network intrusion detection systems (NIDSs) struggle with high-dimensional data due to time and space complexity.
• Support vector machines (SVMs) face the "curse of dimensionality" in large-scale feature spaces.

Purpose of the Study:

• To propose a joint training model combining a stacked autoencoder (SAE) with SVM and kernel approximation.
• To address the limitations of traditional NIDSs in handling high-dimensional data.

Main Methods:

• Utilizing SAE for feature dimension reduction.
• Employing random Fourier features for kernel approximation to enable linear SVM approximation of Gaussian kernel SVM.
• Jointly training SAE with an efficient linear SVM.

Main Results:

• The proposed model demonstrates superior classification performance compared to existing methods.
• Significant reduction in training time was observed.
• The model proved feasible and efficient for large-scale datasets.

Conclusions:

• The joint SAE-SVM model with kernel approximation effectively mitigates the curse of dimensionality in NIDSs.
• This approach offers improved accuracy and efficiency for large-scale network security applications.