With the rapid increase of the Internet of Things (IoT), ensuring the safety of IoT devices has become very important since IoT devices generate and transmit large and sensitive data. Due to this current traditional security measures are not enough to protect this sanative data. To protect the IoT network from traditional and potential attacks or hackers a network Intrusion detection system (NIDS) is placed. In this proposed study, we introduce a two-level feature selection procedure for increasing network security in the Internet of Things (IoT) environment. In this proposed feature selection, we aim to address the challenges caused by high-dimensional datasets by identifying and selecting the most relevant feature set from the dataset to improve detection accuracy and computation overhead. Our proposed methodology combines correlation analysis and Recursive Feature Elimination (RFE) operations to identify the highly important features set from the dataset. The extracted features are the next step in training the deep learning models. The result of this work holds promise for creating a robust and better IoT security mechanism that can tackle traditional and new attacks. We use the IoT-Botnet dataset for 2020 and by applying the proposed feature selection method we achieved the feature reduction from 80 to 16. These models are adopted for the test purpose in the framework of reduced feature sets so that their capabilities can be compared and evaluated. The DL models include LSTM, RNN, GRU, CNN, and DNNs. The experiments reveal that better detection accuracy and lower FAR are achieved with our proposed method which uses the DNN model. Through the comparative study, we found that other deep-learning approaches have a decrease in model complexity and more features.