Scholarlink

In the domain of natural language processing, adversarial attacks often exploit the importance of specific words to deceive models. Identifying these critical words is essential for developing robust defense mechanisms. The SHAP (SHapley Additive exPlanations) method provides a comprehensive approach to quantify feature importance in machine learning models. This paper discusses the advantages of integrating the SHAP method with the SelfDenoise technique, emphasizing how SHAP aids in pinpointing salient words that are typically targeted by attack strategies. Through this integration, we aim to enhance the model’s resilience against adversarial perturbations by reinforcing the denoising process on words deemed significant by SHAP.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Leveraging SHAP to Advance the Robustness of Large Language Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mohamed Yacine Djema,
  • Hacene Fouchal,
  • Olivier Flauzac

摘要

In the domain of natural language processing, adversarial attacks often exploit the importance of specific words to deceive models. Identifying these critical words is essential for developing robust defense mechanisms. The SHAP (SHapley Additive exPlanations) method provides a comprehensive approach to quantify feature importance in machine learning models. This paper discusses the advantages of integrating the SHAP method with the SelfDenoise technique, emphasizing how SHAP aids in pinpointing salient words that are typically targeted by attack strategies. Through this integration, we aim to enhance the model’s resilience against adversarial perturbations by reinforcing the denoising process on words deemed significant by SHAP.