Source
NeurIPS
DATE OF PUBLICATION
12/09/2022
Authors
Ivan Oseledets
Mikhail Pautov
Olesya Kuznetsova
Nurislam Tursynbek
Share
Smoothed Embeddings for Certified Few-Shot Learning
Abstract
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 ℓ2-bounded perturbations that may be useful in few-shot learning scenarios. Our theoretical results are confirmed by experiments on different datasets.
Similar publications
You can ask us a question or suggest a joint project in the field of AI
partner@airi.net
For scientific cooperation and
partnership
partnership
pr@airi.net
For journalists and media