Eco-Friendly Intrusion Detection: Evaluating Energy Costs of Learning

The proliferation of Internet of Things (IoT) devices has led to an increased risk of security threats, making intrusion detection a major concern. Machine learning (ML) algorithms have shown promise in detecting and mitigating intrusions in IoT networks. However, in isolated or remote places IoT devices are often battery-powered or have limited energy sources, making energy efficiency a paramount concern. This paper aims to explore the impact of dataset size on the accuracy and energy consumption of ML models in IoT applications for intrusion detection. To investigate this connection, we conducted an empirical analysis using multiple datasets and commonly employed ML models for intrusion detection in IoT networks. We analyzed the impact of both the number of features and instances on the trade-off between accuracy and energy consumption during the learning phase of ML algorithms. Experimental results suggest that achieving 98.5 % of the maximum achieved accuracy is possible with only 10% of the dataset's samples while simultaneously reducing energy consumption. On the other hand, the number of features tends to decrease the energy consumption while negatively affecting the accuracy.