SHIELD: A Statistical Machine Learning Framework for Diversity Enabled Ensemble Robustness

This proposal aims to develop an integrated statistical machine learning framework for quantifying the diversity and improving the robustness of machine learning ensembles against adversarial attacks. Our framework, named SHIELD (Sparse High-dimensional Inference for Ensemble Learner Diversification), lays down a solid theoretical foundation for understanding the behavior of diversified machine learning ensembles and paves the way for creating formal guarantees of ensemble-based classifier against various attacks. Such a framework will lead to innovative defense method for machine learning ensembles that have high impact in real-world applications, including mobile robots, autonomous vehicles, intelligent sensing systems, etc.