Noisy Labels And Overconfidence

A classifier already has good accuracy, but on borderline names it assigns 99% probability too often and the labels are believed to contain small noise. Which regularization change best targets that failure mode?