Differentially Private Neural Network Training under Hidden State Assumption