Intrusion Detection Systems (IDS) are essential for safeguarding modern network infrastructures against an ever-evolving spectrum of cyberattacks. However, the high dimensionality and imbalance of network traffic data often hinder the performance and scalability of conventional deep learning models. This paper proposes a novel IDS framework that integrates Recursive Feature Elimination (RFE), the TabNet deep neural network, and deep supervision to enhance both accuracy and interpretability. The RFE module effectively reduces feature redundancy, while TabNet utilizes sequential attention to dynamically prioritize the most relevant attributes during classification. Deep supervision further improves convergence and stability by applying auxiliary loss functions to intermediate layers. The model was evaluated on the CIC-IDS2017 dataset and achieved outstanding results, with 98.7% accuracy, 98.5% precision, 98.2% recall, 98.4% F1 score, and an AUC of 0.979. Comparative analysis with recent studies confirms the superiority of the proposed approach in terms of performance, efficiency, and feature transparency. This framework offers a robust and scalable solution for real-world network intrusion detection applications.