EIFDAA: Evaluation of an IDS with function-discarding adversarial attacks in the IIoT.

Area of Science:

• Cybersecurity
• Machine Learning
• Industrial Internet of Things (IIoT)

Background:

• The increasing complexity of the Industrial Internet of Things (IIoT) necessitates advanced intrusion detection systems (IDSs).
• Machine learning-based IDSs are vulnerable to adversarial attacks, posing significant security risks.
• Existing defenses may not adequately address the sophisticated deception tactics employed by adversarial attackers in IIoT environments.

Purpose of the Study:

• To propose and evaluate a framework, EIFDAA (Evaluation of an IDS with Function-discarding Adversarial Attacks in the IIoT), for assessing the resilience of machine learning-based IDSs against adversarial attacks.
• To enhance the robustness of IDSs by identifying weaknesses through adversarial evaluation and applying adversarial training to mitigate vulnerabilities.
• To analyze the effectiveness of various adversarial attack algorithms in deceiving IDSs within the IIoT context.

Main Methods:

• Implementation of the EIFDAA framework, comprising adversarial evaluation and adversarial training processes.
• Simulation of adversarial environments using five distinct adversarial attack algorithms: FGSM, BIM, PGD, DeepFool, and WGAN-GP.
• Evaluation of mainstream machine learning techniques as intrusion detection models and their retraining using adversarial samples.

Main Results:

• Experimental results on the X-IIoTID dataset demonstrated that adversarial attacks can achieve near-zero detection rates, indicating black-box attack capabilities against current IDSs.
• Adversarial training significantly improved the robustness of the evaluated IDSs against adversarial attacks.
• Retrained IDSs effectively defended against adversarial attackers while preserving their original detection rates for legitimate attack samples.

Conclusions:

• The EIFDAA framework provides a viable method for evaluating and enhancing the security of IDSs in IIoT environments.
• Adversarial training is a crucial technique for bolstering the resilience of machine learning-based IDSs against sophisticated cyber threats.
• The proposed approach offers a promising solution for improving the overall robustness and security of IIoT systems.