Abstract
Wireless Sensor Networks (WSNs) play a vital role in domains such as military surveillance, industrial automation, and environmental monitoring. However, their deployment in resource-constrained and adversarial environments exposes them to evolving cyber threats. Traditional intrusion detection systems often fail to handle high-dimensional data, temporal dependencies, and dynamic attack patterns typical in WSNs. To address these challenges, this paper proposes MultiNet-IDS, a framework combining DRL, LSTM networks, and a two-stage EFS mechanism for robust intrusion detection. The EFS process filters features using Pearson correlation and mutual information, then performs an exhaustive combinatorial search with K-fold cross-validation to select an optimal subset enhancing model efficiency without sacrificing accuracy. LSTM captures temporal dependencies in traffic flows, while DRL enables adaptive learning from changing attack behaviors. Together, these models form a hybrid ensemble tailored for the non-stationary and heterogeneous nature of WSN data. MultiNet-IDS was evaluated on three benchmark datasets, UNSW-NB15, CIC-DDoS2019, and Kitsune Network Attack, achieving up to 99.1% accuracy and significant reductions in false positives. Comparative results demonstrate that MultiNet-IDS consistently outperforms state-of-the-art intrusion detection models in adaptability and precision. These outcomes highlight its potential as an effective, lightweight, and generalizable IDS solution for real-world WSN deployments.