Differential Privacy Preservation in Interpretable Feedforward-Designed Convolutional Neural Networks