Internet Of Things Intrusion Detection Based on Deep Learning

Abstract

The fast-growing Internet of Things (IoT) has exposed more attack vectors of the connected devices and heightened the necessity of effective intrusion detection systems (IDS). Nevertheless, the three main challenges that still limit the practical use of IoT intrusion detection include the high complexity of models that are only runnable on resource-constrained devices, low generalization in the presence of small and unbalanced labelling, and the lack of privacy-concerning collaborative detection systems among heterogeneous settings. Here, it is possible to identify a multi-stage research program based on three complementary frameworks that consolidate these challenges in this paper. The main shortcoming of the paper is that the authors first propose lightweight intrusion detection model, LIIDXC, to describe binary XNOR-based convolution of a two-layer long short-term memory (LSTM) and entropy-guided feature selection with a focus on computational overhead reduction without losing detection power. Second, a superior framework of transfer learning, EMTD-SSC, combines a residual convolutional autoencoder, multilayer multi-kernel maximum mean discrepancy (MLMK-MMD) along with fine-tuning techniques to enhance cross-domain transfer when the sample size is small. Third, a federated contrastive learning framework, ID-CFL, enables collaborative intrusion detection without raw data sharing and improves robustness under non-IID client distributions by adaptive node-correlation aggregation. Experiments on N-BaIoT, CIC-DDoS2019, IoT-23, ToN-IoT, and BoT-IoT demonstrate that LIIDXC achieves 87.4% accuracy with approximately fivefold training acceleration, EMTD-SSC reaches up to 94.8% accuracy and maintains 82.8% accuracy with only 100 samples per class, and ID-CFL attains 84.79% average personalized local-client accuracy while improving non-IID collaborative detection performance by up to 38% relative to conventional federated baselines. Taken together, the results show that lightweight computation, cross-domain knowledge transfer, and privacy-preserving distributed learning can be integrated into a coherent IoT IDS research agenda that is efficient, robust, and scalable.