Adversarial attacks against supervised machine learning based network intrusion detection systems.

Area of Science:

• Cybersecurity
• Machine Learning
• Network Security

Background:

• Adversarial machine learning investigates methods to attack and defend AI systems.
• Intrusion Detection Systems (IDS) are crucial for network security but vulnerable to sophisticated attacks.
• Generative Adversarial Networks (GANs) can create realistic synthetic data for testing security models.

Purpose of the Study:

• To evaluate the impact of adversarial attacks on machine learning-based IDS.
• To analyze the effectiveness of evasion and poisoning attacks on Decision Tree and Logistic Regression models.
• To assess the influence of synthetic intrusion traffic generated by GANs on IDS accuracy.

Main Methods:

• Implemented evasion and poisoning attack scenarios using a GAN to generate synthetic intrusion traffic.
• Tested attacks on Decision Tree and Logistic Regression models.
• Evaluated performance using the CICIDS2017 dataset by comparing IDS accuracy before and after attacks.

Main Results:

• Evasion attacks decreased the testing accuracy of both network intrusion detection systems (NIDS) models.
• The Decision Tree model was more susceptible to evasion attacks than the Logistic Regression model.
• Poisoning attacks disrupted the training process of NIDS, with the Logistic Regression model being more affected than the Decision Tree model.

Conclusions:

• Adversarial attacks pose a significant threat to the integrity and performance of machine learning-based IDS.
• Different machine learning models exhibit varying levels of resilience against specific adversarial attack types.
• Further research is needed to develop robust defenses against adversarial machine learning in network security.