AdaptDel: Adaptable Deletion Rate Randomized Smoothing for Certified Robustness

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Huang, Zhuoqun, Marchant, Neil G., Ohrimenko, Olga, Rubinstein, Benjamin I. P.
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866909899463589888
author Huang, Zhuoqun
Marchant, Neil G.
Ohrimenko, Olga
Rubinstein, Benjamin I. P.
author_facet Huang, Zhuoqun
Marchant, Neil G.
Ohrimenko, Olga
Rubinstein, Benjamin I. P.
contents We consider the problem of certified robustness for sequence classification against edit distance perturbations. Naturally occurring inputs of varying lengths (e.g., sentences in natural language processing tasks) present a challenge to current methods that employ fixed-rate deletion mechanisms and lead to suboptimal performance. To this end, we introduce AdaptDel methods with adaptable deletion rates that dynamically adjust based on input properties. We extend the theoretical framework of randomized smoothing to variable-rate deletion, ensuring sound certification with respect to edit distance. We achieve strong empirical results in natural language tasks, observing up to 30 orders of magnitude improvement to median cardinality of the certified region, over state-of-the-art certifications.
format Preprint
id arxiv_https___arxiv_org_abs_2511_09316
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle AdaptDel: Adaptable Deletion Rate Randomized Smoothing for Certified Robustness
Huang, Zhuoqun
Marchant, Neil G.
Ohrimenko, Olga
Rubinstein, Benjamin I. P.
Computation and Language
Cryptography and Security
Machine Learning
We consider the problem of certified robustness for sequence classification against edit distance perturbations. Naturally occurring inputs of varying lengths (e.g., sentences in natural language processing tasks) present a challenge to current methods that employ fixed-rate deletion mechanisms and lead to suboptimal performance. To this end, we introduce AdaptDel methods with adaptable deletion rates that dynamically adjust based on input properties. We extend the theoretical framework of randomized smoothing to variable-rate deletion, ensuring sound certification with respect to edit distance. We achieve strong empirical results in natural language tasks, observing up to 30 orders of magnitude improvement to median cardinality of the certified region, over state-of-the-art certifications.
title AdaptDel: Adaptable Deletion Rate Randomized Smoothing for Certified Robustness
topic Computation and Language
Cryptography and Security
Machine Learning
url https://arxiv.org/abs/2511.09316