An Energy-Efficient Intrusion Detection Offloading Based on DNN for Edge Computing

To improve the accuracy of deep neural networks (DNNs) applied to network intrusion detection systems (NIDS) researchers often increase the complexity of their designed model. Given the processing limitations of resource-constrained devices, researchers have proposed offloading the NIDS task to the cloud. However, simultaneously ensuring energy efficiency and detection accuracy remains a challenge. This article proposes a DNN-based NIDS through early exits that operate following an energy-efficient edge-computing architecture implemented twofold. First, we propose a DNN-based NIDS employing multiobjective optimization for efficient inference and computation offloading. It is designed to perform the classification task at the edge device and configured to proactively offload events to the cloud when additional processing capabilities are required. Our insight is to utilize multiobjective optimization to identify the optimal balance between accuracy and energy efficiency in task offloading. Second, the final DNN branch performs classification with a reject option to ensure reliability when analyzing new network traffic patterns, while calibration adjusts the model’s confidence values to enhance generalization. The rejection mechanism allows the model to accept only its most confident classifications enhancing the model’s generalization capabilities. Experiments conducted with our proposal’s prototype through a new intrusion dataset encompassing one-year-long network traffic with over 7TB of data attested to our proposal’s feasibility. It can reduce the edge device’s energy consumption and processing costs to only 1%, while maintaining accuracy. This is achieved while demanding the offloading of only 10% of network events to the cloud, optimizing resource utilization across both the edge and cloud infrastructures.