Smoothed Embeddings for Certified Few-Shot Learning
This work addresses adversarial robustness for few-shot learning, an incremental extension of existing randomized smoothing methods.
The paper tackles the problem of extending certified adversarial robustness to few-shot learning models by applying randomized smoothing to normalized embeddings, deriving a robustness certificate against ℓ₂-bounded perturbations and confirming results with experiments on datasets.
Randomized smoothing is considered to be the state-of-the-art provable defense against adversarial perturbations. However, it heavily exploits the fact that classifiers map input objects to class probabilities and do not focus on the ones that learn a metric space in which classification is performed by computing distances to embeddings of classes prototypes. In this work, we extend randomized smoothing to few-shot learning models that map inputs to normalized embeddings. We provide analysis of Lipschitz continuity of such models and derive robustness certificate against $\ell_2$-bounded perturbations that may be useful in few-shot learning scenarios. Our theoretical results are confirmed by experiments on different datasets.