Statistical Optimization for High Dimensional Machine Learning with Privacy and Sparsity Constraints