As the field of cybersecurity continues to develop with ever-complex cyber-attacks, older-generation intrusion detection systems have become less and less viable in confronting the challenges posed by these attacks. Thus, the proposed the system for detecting network intrusions (NIDS) aims to enhance models based on deep learning techniques with a feature-wise attention mechanism. This objective expounds on the idea of increasing detection accuracy by allowing models to dynamically shift their focus on the application of the most relevant features in relation to network traffic data in network. We have investigated and implemented three different deep learning architectures: RNN-based attention, GRU-based attention, and LSTM-based attention, which capture time dependencies and put more emphasis on important input dimensions. The XGBoost classifier is also included and serves as the comparison baseline for a machine learning model. Training and evaluation is performed using the SE-CIC-IDS 2018 dataset, which captures a considerable cross-section of real-life attack types. Data preprocessing techniques, including feature selection, normalization, and class balancing, were exhaustively performed to improve data quality and thereby enhance model performance. The experimental findings indicated that the LSTM model incorporating attention mechanisms produced the best performance in classification result, producing fewer false positives than any other architecture. The study considerably contributes to differentiating and discerning NIDS perstormers with thanks to attention mechanism applications.