Deep Transfer Learning Based on Hybrid Swin Transformers With LSTM for Intrusion Detection Systems in IoT Environment

Extensive growth in the number of Internet Of Things (IoT) devices has significantly increased susceptibility to various cyber-attacks and hence emphasized the need for robust intrusion detection systems (IDS) for ensuring IoT network security. While deep learning (DL) methodologies have proven effective in the application of IDS, their success greatly depends on the availability of large datasets and significant computational resources during training. To overcome the limitations associated with this dependence on large datasets and significant computational capacity for training, the current work suggests employing the transfer learning (TL) mechanism by combining Swin Transformers with long short-term memory (LSTM) networks. Utilizing the beneficial properties of Swin Transformers in learning hierarchically structured data combined with the proficiency of LSTM in processing sequential dependencies, the hybrid model generates pre-trained weights in the first phase. These pre-trained weights are further transferred into another instance of the new model for subsequent fine-tuning. Experiments are carried out on several benchmarking datasets, namely NSL-KDD, ToN-IoT, BoTIoT, MQTTIoT, and CICIoT2023, which include both binary and multi-class classification scenarios. The proposed model outperforms state-of-the-art DL models, for example, the Autoencoders, ResNets, CNN, RNN, and LSTM models, and achieved an average of 98.97% in accuracy, of 98.97% in precision, of 99.02% in recall, of 98.97% in F1 score, across all datasets. Experimental results establish that the hybrid approach achieves better detection accuracy and better performance measures compared to the latest state-of-the-art methods, thus proving itself effective in increasing the scalability and adaptability of IDS in IoT.