S-GBT: Smooth Growth Bound Tensor for Certified Robustness Against Word Substitution Attacks in NLP
Researchers have developed a new method called the Smooth Growth Bound Tensor (S-GBT) to enhance the robustness of Natural Language Processing (NLP) models against word substitution attacks. Unlike previous methods that focus on first-order sensitivity, S-GBT utilizes second-order information by bounding the Hessian element-wise. This approach, integrated into the training objective, aims to control both the gradient and its variation, leading to tighter certified robustness bounds. Evaluations on benchmark datasets show that S-GBT can improve certified robust accuracy by up to 23.4% compared to existing techniques, while maintaining competitive clean accuracy. AI
IMPACT Enhances NLP model resilience to adversarial attacks, potentially improving reliability in sensitive applications.