Optimizing ML models for cybercrime detection: balancing performance, energy consumption, and carbon footprint through multi-objective optimization

This study aims to enhance computational performance while minimizing environmental impact in AI (Artificial Intelligence) and ML (Machine Learning) applications, especially in cybersecurity, by developing energy-efficient models using a multi-objective optimization approach. The research utilizes the Non-Dominated Sorting Genetic Algorithm II (NSGA-II), a multi-objective evolutionary algorithm (MOEA), to reduce data dimensionality and identify key features for traffic analysis related to global human trafficking (HTr). The study addresses the increasing energy consumption and carbon emissions due to the rapid adoption of AI technologies, especially during training and deployment. The methodology involves using NSGA-II for feature selection and evaluating energy consumption (Econ) and carbon footprint (CFP) with tools like CodeCarbon and EmissionsTracker. The performance of models such as LSTM (Long short-term memory) and SVM (Support Vector Machine) is assessed in terms of F-measure, with computational environments (Google Colab vs. personal laptop) compared for sustainability. The study applies two datasets: Transnational Terrorist Hostage Event (TTHE) and Counter Trafficking Data Collaborative (CTDC), focusing on key attributes for optimization. Results show that the LSTM model achieved a top F1 score of 98.93%, with precision and recall at 98.96% and 98.81%, respectively. In Pareto front evaluations, the CTDC dataset showed a 94.89% F-measure with a reduction in energy consumption and carbon footprint by 8%, while the TTHE dataset reached a 97.87% F-measure with an 11% reduction. These findings underscore the potential for balancing high performance with reduced environmental impact, contributing to sustainable AI innovations in cybersecurity.