Federated Hybrid Deep Learning with Adaptive Nature-Inspired Optimization for Privacy-Preserving Intrusion Detection in IoT Networks

We “Internet of Things (IoT)” networks has made this need even more pressing, with an ever-growing need to have trustworthy and privacy-preserving IDS systems that detect advanced cyberattacks without revealing sensitive network data. To tackle this issue, this work introduces a “Federated Hybrid Deep Learning with Adaptive Nature-Inspired Optimization (FHDL-ANIO)” framework, which combines “Federated Learning (FL)” with “Adaptive Particle Swarm Optimization (APSO)”, “Adaptive Ant Colony Optimization (AACO)”, and Hybrid CNN–Attention–BiGRU architecture, to support distributed intrusion detection. To optimise features and hyperparameters, the system employs APSO and AACO, while federated learning trains a model independently of raw data. The UNSW-NB15 dataset is used to test the suggested model of 175341 training and 82332 testing samples with nine attack types. Results show that “the hybrid approach achieves the accuracy of 92.92%, precision of 93.67%, recall of 93.45%, F1 score of 93.56%, and ROC-AUC of 0.9827”. APSO then filters the feature space bringing this down from 194 to 114 features (41.24% reduction) reducing the computational complexity of the system, yet maintaining a high rate of detection. The outcomes demonstrate that the suggested architecture for implementing intelligent intrusion detection in dispersed IoT systems is accurate, scalable, effective, and privacy-preserving.