Impartial Adversarial Distillation: Addressing Biased Data-Free Knowledge Distillation via Adaptive Constrained Optimization

Abstract

Data-Free Knowledge Distillation (DFKD) enables knowledge transfer from a pretrained teacher to a light-weighted student without original training data. Existing works are limited by a strong assumption that samples used to pretrain the teacher model are balanced, which is, however, unrealistic for many real-world tasks. In this work, we investigated a pragmatic yet under-explored problem: how to perform DFKD from a teacher model pretrained from imbalanced data. We observe a seemingly counter-intuitive phenomenon, i.e., adversarial DFKD algorithms favour minority classes, while causing a disastrous impact on majority classes. We theoretically prove that a biased teacher could cause severe disparity on different groups of synthetic data in adversarial distillation, which further exacerbates the mode collapse of a generator and consequently degenerates the overall accuracy of a distilled student model. To tackle this problem, we propose a class-adaptive regularization method, aiming to encourage impartial representation learning of a generator among different classes under a constrained learning formulation. We devise a primal-dual algorithm to solve the target optimization problem. Through extensive experiments, we show that our method mitigates the biased learning of majority classes in DFKD and improves the overall performance compared with baselines. Code will be available at https://github.com/ldpbuaa/ipad.