Data-Free Knowledge Distillation Using Adversarially Perturbed OpenGL Shader Images

Logan Frank, Jim Davis

Knowledge distillation (KD) has been a popular and effective method for model compression. One important assumption of KD is that the original training dataset is always available.…