Abstarct: With the increasing expansion of the Internet of Things (IoT) and the connection of billions of devices to the network, the issue of security and intrusion detection has become one of the fundamental challenges in this field. The data generated in IoT environments have high volume, complex dimensions, and diverse features, which makes the intrusion detection process face many difficulties. In this research, a new method for improving intrusion detection in IoT networks is presented, which is baxsed on two main axes: optimizing feature selection and increasing the accuracy of the deep learning model. In the first step, to reduce dimensions and eliminate unnecessary features, a combination of independent component analysis (ICA) and random forest-baxsed feature selection with sequential search (RF-SFS) is used to extract a set of effective and non-repeating features. Then, the selected features are fed into the optimized LSMT model with attention mechanism to more accurately identify complex behavioral patterns and time sequences of attacks. Simulation results show that the proposed method has significantly improved the detection rate, overall accuracy, and reduced false positive error rate compared to traditional and existing methods, and has achieved an average accuracy of 99.91% in detecting several types of cyber attacks.