<p>A cyber-defense framework (CDF) monitors the operation of computer systems by integrating physical processes, computing resources, and communication networks. The increasing scale and complexity of IoT-enabled cyber-physical systems generate high-volume and high-velocity network traffic, introducing significant cybersecurity challenges that demand real-time and computationally efficient intrusion detection mechanisms. Detecting cyber-physical attacks in such environments remains difficult, and traditional intrusion detection systems (IDS) often struggle with accuracy and efficiency. To address these limitations, we propose a deep learning-based IDS optimized for high-performance computing (HPC) environments to enable parallel and real-time cyber threat detection. The model incorporates batch normalization and dropout layers to improve training stability and generalization. The NSL-KDD dataset is utilized to train and evaluate the system, with pre-processing steps including data cleansing and normalization; additionally, experiments are conducted on the UNSW-NB15 to assess robustness and generalization under more recent network traffic conditions. To address class imbalance, the extended synthetic sampling method (ExSSM) is applied. Kernel-assisted principal component analysis (KerPCA) is used for feature extraction, which involves computationally intensive nonlinear transformations that benefit from parallel execution. The chicken swarm genetic algorithm (CSGA) is employed for feature selection and hyperparameter tuning. The system’s effectiveness is validated using key performance metrics such as accuracy of 98.20%, precision of 97.10%, and F1 score 97.84%. Results indicate that the proposed method outperforms traditional IDS approaches, offering improved detection performance. This research contributes to enhancing cyber defense in IoT systems through an HPC-aware deep learning architecture, and outlines directions for future work in mitigating emerging threats like backdoor attacks.</p>