Detecting and Mitigating Adversarial Attacks: Model Comparison Between Tree-Based and Deep Learning-Based Intrusion Detection Systems

Abstract

The study investigates the detection and mitigation of adversarial attacks targeting tree-based and deep learning-based intrusion detection systems (IDS). There has been a rise in advanced cyber threats; traditional IDS techniques often fall short, rendering them ineffective in some situations, prompting increased adoption of deep learning-based models as alternatives. These models are also prone to adversarial attacks that are creatively designed to deceive the system into classifying malicious traffic as benign. The Canadian Institute for Cybersecurity Intrusion Detection System 2017 (CICIDS2017) dataset was used for the study. Data were processed to handle missing values, normalisation features, and address the class imbalance using the ADASYN (adaptive synthetic) sampling approach. Evaluation was carried out on two models, namely the random forest classifier and the deep neural network system. An almost perfect accuracy of 99.80% was achieved with the random forest model, which shows its effectiveness against adversarial attacks like the fast gradient sign method (FGSM) and projected gradient descent (PGD). However, as promising as the deep neural network (DNN) system is, it only achieved an accuracy rate of 81%, which shows that it is open to adversarial attacks. Feature importance analysis highlighted critical network attributes, such as packet length and flow size, for intrusion detection. The finding shows the superior adversarial effectiveness of tree-based models relative to deep learning, which motivates for hybrid AI approaches and advanced future research. The study emphasises the integration of adversarial training, interoperability of models, and detection of anomalies. The study contributes to the ever-evolving development of robust and secure IDS, while not overlooking the recurring issues in cybersecurity.