Multi-source refined adversarial domain adaptation with transfer complementarity infusion for IoT intrusion detection under limited samples.

The arrival of 5G has facilitated further development of the Internet of Things (IoT) which is vulnerable to hacking because of its widespread use. Large networks oriented toward endpoints risk privacy leakage if attacked. Therefore, studying intrusion detection in the IoT is crucial. Current models are trained on large samples, whereas IoT devices can only intercept a limited attack samples in some scenarios, resulting in poor detection performance. A domain adaptation algorithm based on transfer learning can effectively transfer the source domain samples to the target domain, as a solution to the above problem. However, inconsistencies in categories and domains are not adequately addressed. In view of this, we propose a multi-source refined adversarial domain adaptation model with transfer complementarity infusion (MRADDA-TC) for IoT intrusion detection under limited samples. First, a lightweight scale-aware bilateral feature pyramid network is built for intrusion feature extraction, with three additional parallel modules for the MRADDA-TC model. In the multi-source refined adversarial domain adaptation component, knowledge transfer between the multi-source and target domains is achieved by optimizing the difference in category distribution between each pair of domains. Finally, the additive margin Softmax and multi-classifier complementarity infusion modules are designed to determine the similarity of the adaptation modules. The results are fed back into the multi-source refined adversarial adaptation module to guide the adaptation process. Extensive results on four IoT datasets validate the feasibility and superiority of the proposed model for detecting IoT intrusions with limited samples. Specifically, it has an average detection accuracy of more than 95 % under different limited sample tasks for the four datasets and has good robustness in the presence of noise. [ABSTRACT FROM AUTHOR]