Differentially Private Non-convex Learning for Multi-layer Neural Networks