RobEns: Robust Ensemble Adversarial Machine Learning Framework for Securing IoT Traffic.

Area of Science:

• Cybersecurity
• Machine Learning
• Internet of Things (IoT)

Background:

• Machine Learning (ML)-based Intrusion Detection Systems (IDS) are crucial for Internet of Things (IoT) security.
• These ML-based IDSs are vulnerable to adversarial examples generated through Adversarial Machine Learning (AML) attacks.
• Attackers exploit these vulnerabilities to degrade IDS performance and evade detection, necessitating robust defense strategies.

Purpose of the Study:

• To introduce RobEns, a robust ensemble framework for enhancing ML-based IDSs in IoT networks.
• To investigate the impact of black-box AML evasion attacks on various ML-based IDSs.
• To evaluate the robustness of these IDSs after implementing defense mechanisms.

Main Methods:

• Developed RobEns, an ensemble framework integrating state-of-the-art ML models and ensemble techniques for IoT IDSs.
• Investigated four typical AML attacks against six ML-based IDSs using three benchmarking datasets in multi-class classification scenarios.
• Implemented two defense mechanisms: feature squeezing (data-based) and adversarial training (model-based).

Main Results:

• Experiments revealed a significant drop in detection accuracy for some ML-based IDSs under AML attacks.
• The implemented defense mechanisms (feature squeezing and adversarial training) significantly enhanced IDS robustness.
• RobEns achieved up to 100% accuracy in black-box attack scenarios while maintaining high performance without adversaries.

Conclusions:

• The RobEns framework effectively improves the robustness of ML-based IDSs against AML attacks in IoT environments.
• Defense strategies like feature squeezing and adversarial training are vital for hardening IDSs.
• The proposed approach ensures secure IoT networks by maintaining high detection accuracy even when facing sophisticated threats.