Scalable and Efficient Training of Large Convolutional Neural Networks with Differential Privacy