AdaptDel: Adaptable Deletion Rate Randomized Smoothing for Certified Robustness
This addresses the challenge of certified robustness for sequence classification in natural language processing, where varying input lengths lead to suboptimal performance with fixed-rate methods, representing a novel method for a known bottleneck.
The paper tackled the problem of certified robustness for sequence classification against edit distance perturbations by introducing AdaptDel methods with adaptable deletion rates that dynamically adjust based on input properties, achieving up to 30 orders of magnitude improvement in median cardinality of the certified region over state-of-the-art certifications.
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.