Unmasking insider threats using a robust hybrid optimized generative pretrained neural network approach

Area of Science:

• Cybersecurity
• Artificial Intelligence
• Network Security

Background:

• Insider threats pose significant risks to network infrastructure.
• Traditional detection models struggle with class imbalance and high-dimensional data.
• Neural network-based models offer improved performance but require robust solutions for data challenges.

Purpose of the Study:

• To propose a novel Hybrid Optimized Generative Pretrained Neural Network based Insider Threat Detection (HOGPNN-ITD) model.
• To address the class imbalance problem in insider threat detection.
• To enhance the precision and performance of identifying security breaches caused by insiders.

Main Methods:

• Utilized Adabelief Wasserstein Generative Adversarial Network (ABWGAN) with Expected Hypervolume Improvement (EHI) for adversarial sample generation.
• Employed L2-Starting Point (L2-SP) regularization on a pretrained Attention Graph Convolutional Network (AGCN) for insider identification.
• Incorporated Chebyshev Graph Laplacian Eigenmaps solver (CGLE) for dimensionality reduction and Insider State clustering via Density-Based Spatial Clustering of Applications with Noise (IS-DBSCAN).

Main Results:

• The HOGPNN-ITD model demonstrated high detection rates for insider threats.
• The proposed approach achieved a minimal false alarm rate.
• Experimentation on a benchmark dataset validated the model's effectiveness in detecting skeptical user behavior.

Conclusions:

• The HOGPNN-ITD approach offers a robust solution for insider threat detection, particularly in challenging high-dimensional and imbalanced datasets.
• The integration of generative adversarial networks and attention-based graph convolutional networks significantly improves detection accuracy.
• This method provides a promising advancement for securing network infrastructure against insider threats.