Empirical Risk Minimization with $f$-Divergence Regularization