DIFF2: Differential Private Optimization via Gradient Differences for Nonconvex Distributed Learning