Differentially Private Stochastic Gradient Descent with Low-Noise