Toward Better Target Representation for Source-Free and Black-Box Domain Adaptation

Domain adaptation aims at aligning the labeled source domain and the unlabeled target domain, and most existing approaches assume the source data is accessible. Unfortunately, this paradigm raises concerns in data privacy and security. Recent studies try to dispel these concerns by the Source-Free setting, which adapts the source-trained model towards target domain without exposing the source data. However, the Source-Free paradigm is still at risk of data leakage due to adversarial attacks to the source model. Hence, the Black-Box setting is proposed, where only the outputs of source model can be utilized. In this paper, we address both the Source-Free adaptation and the Black-Box adaptation, proposing a novel method named better target representation from Frequency Mixup and Mutual Learning (FMML). Specifically, we introduce a new data augmentation technique as Frequency MixUp, which highlights task-relevant objects in the interpolations, thus enhancing class-consistency and linear behavior for target models. Moreover, we introduce a network regularization method called Mutual Learning to the domain adaptation problem. It transfers knowledge inside the target model via self-knowledge distillation and thus alleviates overfitting on the source domain by learning multi-scale target representations. Extensive experiments show that our method achieves state-of-the-art performance on several benchmark datasets under both settings.