Uncertainty-Quantified, Robust Deep Learning for Network Intrusion Detection

Cyber threats are moving beyond human comprehension and reaction capability in a rapidly evolving world. Deep learning models for network intrusion detection are becoming evermore crucial in processing network traffic to filter benign content from malicious activity. However, novel attacks such as zero-days are becoming more frequent, demonstrating the need for robust deep learning models to flag attacks while providing predictive certainty guarantees. Therefore, detecting out-of-distribution (OOD) inputs at inference time is crucial to address the rapidly changing environment while keeping up with evolving cyber threats. We develop multi-class deep learning models for network intrusion detection, comparing deterministic with Bayesian neural networks estimated using Hamiltonian Monte Carlo. We also propose new uncertainty quantification scoring measures for performance evaluation to evaluate certainty in predictions. During our experimentation, our best performing proposed Bayesian deep learning model detected 89.1% and 86.9% of the OOD packets at the 5% and 0.1% significance levels, respectively.