MT-KD: Multi-Teacher Knowledge Distillation for Privacy-Aware Unlearning in IIoT Intrusion Detection Systems

The growth of Industrial Internet of Things (IIoT) systems creates a direct conflict for intrusion detection systems: while essential for security, their inability to selectively forget specific attack patterns hinders adaptability and impedes compliance with privacy regulations like General Data Protection Regulation. To resolve this conflict, we introduce the Multi-Teacher Knowledge Distillation (MT-KD) unlearning framework, designed to erase knowledge of specified attack classes without the prohibitive cost of full model retraining. Our approach employs deep convolutional neural networks, multi-teacher guidance, and gradient-based noise injection to train a student model that forgets targeted attacks while retaining high accuracy on remaining classes. We validate this by converting network traffic from the CIC-IDS2017 and USTC-TFC2016 benchmark datasets into grayscale images for extensive experiments. The results demonstrate that our MT-KD framework achieves complete forgetting of targeted classes, verified by Membership Inference Attack resistance, while maintaining over 95% accuracy on retained classes. By outperforming existing unlearning baselines in both effectiveness and efficiency, our work delivers a practical and scalable solution for privacy-preserving intrusion detection in real-world IIoT scenarios. The source code is publicly available at https://github.com/NJS-HZ/MT-KD.