Differentially Private Extreme Learning Machine

This paper presents a novel algorithm, the Differentially Private Extreme Learning Machine (DPELM), which guarantees pure differential privacy. The differential privacy budget is determined solely by one parameter of the ELM model: the number of hidden nodes. We demonstrate the effectiveness of DPELM by showcasing its reasonable utility, including its strong generalization performance, across benchmark datasets. Furthermore, DPELM offers a user-friendly experience, as it eliminates the need to consider data dimensionality, gradient norms, or objective function sensitivities.