Evaluating Sustainability and Social Costs of Adversarial Training in Machine Learning

The growing prevalence of adversarial attacks on machine learning models in consumer electronics necessitates enhancing adversarial robustness. Although adversarial training improves the robustness of a model against adversarial attacks, its sustainability remains a critical concern due to carbon emissions and the environmental impact of the extensive computational demands. To address this, we use the Robust Carbon Trade-Off Index metric, which establishes a relationship between robustness and carbon emissions, and introduce the Cost Per Unit of Robustness Change metric to quantify the economic impact of increasing robustness in terms of carbon emission costs measured by an economic metric quantifying the costs associated with carbon emissions. By examining the theoretical foundations, practical quantification techniques, and interdisciplinary research areas, we shed light on the multifaceted aspects of building sustainable and scalable models with robust adversarial defenses.