Model compression using knowledge distillation with integrated gradients