Trustworthy Generative Few-Shot Learning Based Intrusion Detection Method in Internet of Things

• Published in: IEEE transactions on consumer electronics 2024-10, p.1-1
• Main Authors: Zhang, Ze, Wang, Pan, Zhang, Tianqi, Liu, Minyao, Zhou, Xiaokang
• Format: Article
• Language: English
• Subjects: Artificial intelligence; Consumer electronics; Data models; Diffusion models; Feature extraction; Few shot learning; Generative Few-Shot Learning; Internet of Things; Model interpretability; Network intrusion detection; Telecommunication traffic; Training
• ISSN: 0098-3063; 1558-4127
• DOI: 10.1109/TCE.2024.3473304

The application of Artificial Intelligence (AI) and Internet of Things (IoT) technologies enhances device smart decision-making, yet it also makes them more susceptible to cyber-attacks. AI-enabled Network Intrusion Detection System (NIDS) has been regarded as an effective way to mitigate such attacks. However, large-scale attacks data collection and labeling is prohibitively costly and time-consuming. Therefore, effective Few-Shot Learning based NIDS for IoT are crucial. Variational Autoencoders and Generative Adversarial Networks are widely used in generative few-shot learning but struggle with limited sample diversity and unstable training. Meanwhile, Deep Learning-based NIDS often lacks trustworthiness due to the "black box" problem. In response to these challenges, we propose a network traffic sample generation method that leverages the Conditional Denoising Diffusion Probabilistic Model (CDDPM) to tackle the issues of samples. Then, we introduce an Intrusion Detection method named CNNBiGRU that combines Convolutional Neural Networks with Bidirectional Gated Recurrent Units based on the synthetic data generated by CDDPM. Furthermore, we apply the SHAP method to interpret the results to ensure model users' trust. Finally, the proposed method is evaluated on the two public datasets, which illustrate that our approach can successfully generates high-quality samples and improves the performance of NIDS.