<p>Traditional Intrusion Detection Systems (IDSs) are largely unable to provide the level of protection required to safeguard modern network environments against ever-evolving cyber threats. Conventional machine-learning-based feature representations are static and cannot capture temporal behaviour in network traffic, limiting performance under heterogeneous conditions. Deep learning approaches have improved detection accuracy, but challenges remain in feature diversity, class imbalance, and cross-environment generalisation. To overcome these limitations, we propose DeepShieldIDS, a hybrid deep learning–based network intrusion detection framework built on established convolutional neural networks (CNNs), long short-term memory (LSTM) networks, and an attention mechanism, all integrated into a single evaluation pipeline. The HybridIDSNet utilises spatial feature extraction, temporal sequence modelling and feature-weighting attention in a single framework to improve traffic classification performance. We conduct experiments on two benchmark datasets, CIC-IDS2017 and UNSW-NB15, in independent and cross-dataset evaluation settings. The model outperforms classical machine learning models (Random Forest, SVM, and XGBoost) and Categorical CNN and LSTM baselines, achieving accuracies of 97.98% and 98.02% with ROC-AUC scores of 99.05% and 99.12%, respectively. Evaluation is confined to supervised benchmark settings, while cross-dataset validation suggests better generalisation with heterogeneous traffic distributions. The proposed framework focuses on a deployment-oriented architecture and a formalised pre-processing procedure, providing an overview of scalable hybrid deep learning solutions for network intrusion detection.</p>