As computer networks expand, robust Intrusion Detection Systems (IDS) are crucial. Traditional methods struggle with zero-day attacks and lack interpretability. This study uses the NSL-KDD dataset and three machine learning models: Random Forest, attention-based Multilayer Perceptron (MLP), and hybrid CNN–BiLSTM. Models were evaluated on accuracy, error rate, ROC–AUC, and confusion matrix. The hybrid CNN–BiLSTM achieved 77.84% accuracy; attention-based MLP reached 0.9076 AUC. SHAP (Shapley Additive Explanations) analyzed key features in the Random Forest model, enhancing transparency. Results demonstrate the framework’s improved real-time intrusion detection performance with increased explainability and practical network security applicability.