A federated deep learning approach for SDN security with quantum optimized feature selection and hybrid MSDC net architecture

Software-Defined Networking (SDN) is increasingly exposed to complex cyberattacks, requiring advanced, adaptive, and efficient intrusion detection mechanisms. This study presents LightIDS-SDN, a federated and explainable intrusion detection framework tailored for SDN environments. At its core, the system employs Dual Fitness Enhanced Quantum-Inspired Particle Swarm Optimization (DFE-GQPSO) for feature selection, which identifies the most informative network attributes while eliminating redundant or irrelevant features. This quantum-optimized feature selection significantly improves detection performance by reducing overfitting and enhancing generalization. The framework incorporates a hybrid deep learning architecture, MSDC-Net, combining Transformer layers, Capsule Networks, and BiLSTM units to capture contextual, spatial, and sequential dependencies in network traffic. Federated learning using FedAvg enables collaborative model training across multiple SDN controllers while preserving data privacy. Explainable AI modules, based on SHapley Additive exPlanations (SHAP) and Gradient-weighted Class Activation Mapping (Grad-CAM), provide both global and local interpretability, ensuring transparent and accountable decision-making. Experiments on the InSDN dataset demonstrate the effectiveness of the proposed system, achieving 98.73% accuracy, 98.80% precision, 98.65% recall, and 98.72% F1-score. Comparative analysis confirms that DFE-GQPSO outperforms traditional feature selection methods, enhancing model robustness and training efficiency. Overall, LightIDS-SDN effectively detects a wide range of SDN attacks while addressing limitations of conventional IDS approaches, including limited scalability, lack of interpretability, and computational inefficiency. This work lays the foundation for deploying quantum-optimized, explainable, and federated intrusion detection systems in SDN networks.