Comparison of classical ML models and neural networks for network intrusion detection

We compare SVM, random forest, light GBM, logistic regression, KNN, and neural networks for classifying whether a network connection is a cyberattack or normal IoT activity based on packet capture data. We analyze model accuracy, model memory usage, and inference times. Our classical models are trained using Sklearn while our neural nets are implemented in PyTorch. We apply post training quantization and quantization aware training (8-bit) in PyTorch and TensorRT to significantly reduce the inference times of our neural nets.