AdaptDel: Adaptable Deletion Rate Randomized Smoothing for Certified Robustness

2025-11-12

AI Summary

Existing randomized smoothing methods for certified robustness of sequence classification against edit-distance perturbations employ fixed deletion rates, rendering them ill-suited for variable-length natural language inputs and limiting both certified accuracy and scalability. This work pioneers the extension of randomized smoothing theory to adaptive deletion rates, introducing a length- and structure-aware deletion strategy that enables dynamic, fine-grained certification against edit-distance perturbations. Key contributions include: (1) a deletion probability that dynamically adjusts according to sequence-specific attributes (e.g., length and syntactic structure); and (2) theoretically grounded certification radii that are significantly enlarged compared to prior approaches. Empirically, on standard NLP benchmark tasks, the median size of the certified region improves by up to 30 orders of magnitude over state-of-the-art methods, achieving both rigorous theoretical guarantees and superior empirical performance.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Adversarial Learning & Robustness
• Computer Vision: Adversarial Attacks & Robustness

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Security and Privacy: Large-scale security measurements

Abstract

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.

Problem

• Certified robustness for sequence classification against edit distance perturbations
• Addressing suboptimal performance from fixed-rate deletion mechanisms
• Developing adaptable deletion rates for varying input lengths

Innovation

• Adaptable deletion rates adjust dynamically to input properties
• Extends randomized smoothing framework for variable-rate deletion
• Achieves certified robustness against edit distance perturbations