Task-Aware Adversarial Feature Perturbation for Cross-Domain Few-Shot Learning

Abstract

Currently, metric-based meta-learning methods have achieved great success in few-shot learning (FSL). However, most works assume a high similarity between base classes and novel classes, and their performance can be greatly reduced when comes to domain-shift problem. As a result, cross-domain few-shot learning (CD-FSL) methods are proposed to tackle the domain-shift problem, which places a higher demand on the robustness of the meta-knowledge. To this end, we propose a feature augmentation method called Task-Aware Adversarial Feature Perturbation (TAAFP) to improve the generalization of the existing FSL models. Compared to the traditional adversarial training, our adversarial perturbations are generated from the feature space and contain more sample relationship information, which is discovered by the Task Attention Module. Task Attention Module is designed based on a transformer to capture more discriminative features in a task. Therefore, our perturbations can easily attack the extraction process for discriminative features, forcing the model to extract more robust discriminative features. In addition, a regularization loss is introduced to ensure the predictions of the adversarial augmented task remain similar to the original task. We conduct extensive classification experiments in five datasets under the setting of cross-domain few-shot classification. The result shows that our method can significantly improve the classification accuracy in both seen and unseen domains.