Towards Securing IoT: A Deep Autoencoder-Based Anomaly Detection System

This research introduces three NIDS systems, named new Parallel Deep Auto-Encoder (NEW PDAE), as opposed to APAE, and DFE that has been proposed, with the goal of reducing processing load while enhancing or maintaining detection accuracy. Each of these models leverages different deep learning techniques to create an optimized structure. The proposed models were trained and evaluated on three datasets—CICIDS2017, UNSW-NB15, and KDDCup99—and then compared to the NDAE and MemAE algorithms. For multi-class classification, the NEW PDAE model achieved accuracies of 99.43%, 99.84%, and 99.92% on CICIDS2017, UNSW-NB15, and KDDCup99 datasets, respectively. The APAE model yielded accuracies of 99.50%, 99.89%, and 99.94%, while the DFE model achieved 99.31%, 99.96%, and 99.92% on these datasets. These results demonstrate that the proposed models provide sufficient accuracy and outstanding performance on these benchmark datasets for NIDS applications. Conclusions: In this study, we presented three architectures, NEW PDAE, APAE, and DFE, for use in NIDSs. These architectures have demonstrated higher performance compared to simple encoder methods. The NEW PDAE method employs a parallel auto encoder technique with the ability to extract representations in different views. Due to the parallel feature extraction operations, NEW PDAE incurs less computational overhead, time, and fewer parameters compared to traditional methods for feature extraction. Another method, the APAE model, is based on an asymmetric auto encoder utilizing convolutional layers. This approach excels in extracting the best features in the encoder section by virtue of utilizing its modules effectively. Lastly, we introduced a very lightweight and powerful method called DFE, capable of using minimal processing and memory due to feature structuring while maintaining a very high detection accuracy. For model implementation and testing, we utilized three datasets: CICIDS2017, UNSW-NB15, and KDDCup99, comparing our architectures with MemAE and NDAE models. As observed from the results presented in Section Four, our models NEW PDAE, APAE, and DFE exhibit higher accuracy compared to other models while offering fewer parameters. Therefore, it can be concluded that our proposed models are more suitable choices for devices like Internet of Things, where computational time and cost are critical, providing a much more viable option. Considering the hardware and computational constraints of Internet of Things devices, and bearing in mind the mentioned limitations, the aim is to substitute conventional layers in the presented methods with novel, extremely lightweight, and efficient convolutional layers in the future..