Network technologies play a vital role in the rapid growth of the technological era and living standards. However, the weaknesses of traditional security mechanisms lead to the difficulty of stable network development. Therefore, this work suggests a novel graph transformation method for network traffic data and an Intrusion Detection System (IDS) using the Graph Attention Network (GAT). This IDS includes two distinct scenarios: the first for detecting popular attacks like DoS and Probe, and the second for identifying less common intrusions such as U2R and R2L. Both scenarios are constructed based on the analysis stage using Principal Component Analysis. Moreover, Particle Swarm Optimization is utilized for feature selection and GAT is combined with multiple Machine Learning and Deep Learning classifiers for classification stages. The optimal architectures of the chosen models and the reliable validation performances are achieved by applying 5-fold cross-validation procedures. As a result, the high-quality extracted feature sets in two scenarios possibly improve the effectiveness of the proposed IDS, with GAT identified as the most powerful classifier for efficient IDS construction. Additionally, both common and uncommon network attacks are detected with high performance in the suggested scenarios, demonstrating the potential to solve security problems in practical environments.