Sustainable AI for Cyber Defense: A Multi-Objective Benchmark

Intrusion Detection Systems (IDS) are extensively used in modern networking, to ensure safety over the internet. These systems are aggressively switching towards machine learning approaches to improve their performance, which in turn increases computational demands for deploying such systems. Machine learning approaches also require robustness evaluation so that minor changes in data do not threaten the IDS. Most evaluations today focus only on improving performance, neglecting the robustness and efficiency prospects of these systems. This study aims to benchmark supervised learning models for these multi-objective tradeoffs. It evaluates models based on three axes: classification performance using standard performance metrics, robustness under Gaussian noise injection, and efficiency using temporal and memory metrics. The results highlight tree-based models as strong performers with reasonable efficiency, while neural nets exhibit more stable performance under variance at higher computational costs. These findings put into light the importance of multi-criteria evaluation for practical and sustainable AI-driven cyber defense.