Differentially Private Empirical Risk Minimization: Efficient Algorithms and Tight Error Bounds