Lightweight IoT Intrusion Detection via Curriculum Knowledge Distillation

Abstract

The rapid development of Internet of Things (IoT) systems brings severe security challenges, as edge devices are restricted by limited computing and memory resources. Existing intrusion detection methods either deliver high accuracy at the expense of heavy computation, or cannot maintain stable performance under resource constraints. To address these issues, this article proposes a lightweight intrusion detection framework named curriculum knowledge distillation (CKD)-ADSCLNN, which jointly optimizes feature representation, model compression, and convolutional efficiency. Specifically, a two-stage feature optimization strategy combining mutual information (MI)-based filtering and sparse autoencoder (SAE) compression is adopted to remove redundant traffic information while retaining discriminative features. Furthermore, asymmetric depthwise separable convolution (ADSC) is introduced to reduce computational complexity while preserving strong local feature extraction capability. In addition, a CKD strategy is developed to enhance the model’s learning ability for minority attack classes under imbalanced traffic distributions. Extensive experiments on NSL-KDD, CICIDS2017, and Bot-IoT datasets demonstrate that the proposed method achieves competitive detection accuracy while greatly reducing computational cost and parameter size. The results show that our method demonstrates promising real-time performance on a single-core CPU, highlighting its potential for resource-constrained IoT deployment.