On the Performance of Empirical Risk Minimization with Smoothed Data