Private Model Compression via Knowledge Distillation