Lightweight Deep Learning Models for Intrusion Detection in Resource-Constrained Cyber-Physical Devices

Abstract

Important infrastructures like water treatment and smart grids rely heavily on cyber-physical systems (CPS), which are increasingly vulnerable to new and developing threats.  The advent of Intrusion Detection Systems (IDS) tailored to CPS design has become one of the essential tactics for safeguarding them, despite the fact that traditional security measures like firewalls and encryption do not function well enough with CPS architecture. The study article presents a deep learning (DL) approach to an IDS to identify security attacks in a network of resource-constrained CPS using the CICIDS2017 database. Some preprocessing operations, such as cleaning, transformation, one-hot encoding, outliers removal using the Z-score method, and class balancing using SMOTE, were performed on the dataset. The tasks of dimensionality reduction and feature selection were performed with the help of Principal Component Analysis (PCA). The binary and multi-class intrusion detection was done with a built and tested six-layer Deep Neural Network (DNN) and a Convolutional Neural Network (CNN). The DNN could obtain 99.7% and the CNN 99.5%. The two models were good generalizers, with low overfitting, and high precision (PRE), recall (REC), and F1-Score (F1). Comparative performance analysis of the existing methods and the suggested models proved that the suggested models possess a higher detection accuracy and can be utilized much faster, hence, they may be an appropriate option when it comes to the deployment of a strong and scalable IDS in a cyber-physical environment.