Hybrid Neural Network-Based Intrusion Detection System: Leveraging LightGBM and MobileNetV2 for IoT Security

The rapid expansion of the Internet of Things (IoT) has uncovered a significant asymmetry in cybersecurity, where low-power edge devices must face sophisticated threats from adversaries backed by ample resources. In our study, we employ a symmetry-based approach to rebalance these uneven scenarios. We propose a Hybrid Neural Network Intrusion Detection System (Hybrid NNIDS) that uses LightGBM to filter anomalies at the traffic level and MobileNetV2 for further detection at the packet level, creating a viable compromise between detection accuracy and computational cost. Additionally, the proposed Hybrid NNIDS model, on the ACI-IoT-2023 dataset, outperformed other intrusion detection models with an accuracy of 94%, an F1-score of 91%, and a precision rate of 93% in attack detection. The results indicate the developed asymmetry algorithm can greatly reduce processing overhead while still being able to be implemented in IoT environments. The focus of future work will be on the real-world deployment of these security infrastructures in the IoT and their adaptation to newer types of attack vectors that may be developed by malware.