RESONANCE: robust synergistic consistent and covariant learning for adversarial few-shot classification

Few-shot learning (FSL) aims to address data scarcity by enabling models to adapt to new scenarios with limited samples. Adversarial samples, which expose the vulnerabilities of models trained on large datasets, pose an even greater challenge to FSL models. This is primarily due to two reasons. First, the task sampling mechanism—commonly adopted in traditional FSL paradigms—often leads to poor generalization. Second, existing feature learners for FSL are typically optimized without prior knowledge of adversarial threats, resulting in non-robust feature representations. To mitigate these issues, we propose a Robust synErgiStic cOnsisteNt ANd Covariant lEarning (RESONANCE) strategy to develop a few-shot learning framework without relying on meta-task sampling. Our method provides more generalized embeddings for unseen adversarial samples because it extracts consistent and covariant features that enhance robustness to perturbations. Specifically, to encourage the model to capture relationships among images, we construct guidance groups comprising natural and adversarial samples for training. Additionally, to better preserve inherent semantic information, we design two auxiliary branches that provide supervised signals to guide the embedding learning process. Furthermore, weighted distillation is introduced to selectively transfer knowledge, allowing the model to focus on discriminative regions. Extensive experiments demonstrate that our approach outperforms state-of-the-art methods under various adversarial attacks while minimizing performance degradation on natural inputs.