A Hybrid Intrusion Detection System for Smart Home Security Using LSTM and Autoencoder Neural Networks

Smart homes continuously grow in popularity and effectiveness, however, the interconnection of devices raises concerns over intrusion detection. The amount of smart devices continues to expand and the complexity of the intrusion detection systems needs to equally match these wide intruder capable environments. This article focused on this need by proposing a new solution, which is a hybrid intrusion detection system consisting of autoencoder based feature extraction and Long Short-Term Memory (LSTM) neural network in its classification phase. The entire process included a total data preprocessing stage that had various processes such as the normalization, data shuffling, data partitioning, data cleaning and label splitting and others. These processes were effective in achieving the objectives of the study. The proposed methodology was demonstrated through the application of the famous ‘NSL-KDD’ dataset where training achieved a significant accuracy of 99.467% while the test dataset produced an accuracy of 99.391%. This research is important not only because of its high accuracy but also because of the robust added value it provides to smart home protection. Enabling an autoencoder on a data set allowed us to emphasize specific features and therefore lower the dimensionality space remaining useful information. In addition, the further use of trained LSTM networks allowed the system to understand even more complex sequences of normal and hostile actions done by users improving the prediction. Furthermore, the strictly followed data preprocessing steps made sure that the model was exposed to good and reasonable data, which is an important factor for the efficiency and applicability of any IDS.