Sliding principal component and dynamic reward reinforcement learning based IIoT attack detection

Area of Science:

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

Background:

• The Internet of Things (IoT) and Industrial IoT (IIoU) are increasingly vulnerable to cyber threats, necessitating robust attack detection systems.
• Existing methods struggle with high error rates and detecting diverse attacks in large IIoT datasets.

Purpose of the Study:

• To introduce a novel reinforcement learning-based attack detection method, Sliding Principal Component and Dynamic Reward Reinforcement Learning (SPC-DRRL), for enhanced IIoT network security.
• To improve the accuracy and reduce the error rate in detecting various IIoT network attacks.

Main Methods:

• Data preprocessing using min-max normalization on the TON_IoT dataset.
• Feature extraction via a sliding principal component algorithm with a sliding window.
• Development of a dynamic reward reinforcement learning model with an incident repository for adaptive training.

Main Results:

• The SPC-DRRL system demonstrated significant reductions in IIoT attack detection time, overhead, and error rates.
• Achieved higher accuracy in attack detection compared to traditional reinforcement learning methods.
• Validated effectiveness on the ToN_IoT dataset.

Conclusions:

• The proposed SPC-DRRL method offers a more effective and efficient solution for securing IIoT networks against cyber threats.
• This approach significantly improves upon existing methods by minimizing detection errors and enhancing overall system performance.