Training differentially private machine learning models requires constraining
an individual's contribution to the optimization process. This is achieved by
clipping the $2$-norm of their gradient at a predetermined threshold prior to
averaging and batch sanitization. This selection adv