The rise in internet users and connected devices has heightened cybersecurity risks, demanding more effective intrusion detection systems (IDS). Traditional IDS models often rely on synthetic datasets and struggle to handle class imbalance and high-cardinality categorical features, limiting their effectiveness in real-world scenarios. The main objective is to design and evaluate an IDS capable of improving detection accuracy and reducing false positives in real network environments. The model designed for binary classification integrates hierarchical attention layers, EQLv2 a loss function for dynamic class balancing, and CatBoost encoding to manage high-cardinality categorical data. Trained on real network traffic from Litnet, the model overcomes the limitations of synthetic datasets, achieving 99.92% detection rate, 99.92% F1-score, and a low 0.075% false positive rate (FPR). These results demonstrate the model’s superiority over traditional IDS approaches, reinforcing its applicability for real-world cybersecurity.