Scholarlink

Conventional neuron editing methods for language model privacy protection suffer from a critical limitation: uniform intervention strategies inadvertently exacerbate cross-category leakage via the privacy seesaw effect. Current methodologies overlook the functional divergence between category-specific neurons (activated solely by single-class private data) and cross-category neurons (responsive to multiple privacy classes). This study introduces a taxonomy-driven editing framework that identifies privacy neurons through integrated gradient attribution and categorizes them via multi-task activation patterns. Our method combines adversarial patching for category-specific neurons with class-conditional regularization for cross-category neurons. Evaluations on BERT and GPT-2 demonstrate that our method can effectively reduce privacy data exposure and leakage while maintaining model utility (perplexity increase <8%). The framework establishes an operational taxonomy for privacy neuron functionality, enabling architecture-specific solutions for regulatory-aligned AI systems.

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

Categorized Privacy Neuron Editing: Addressing Heterogeneity for Privacy Leakage Mitigation in Language Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Xiaole Yang,
  • Peng Wu

摘要

Conventional neuron editing methods for language model privacy protection suffer from a critical limitation: uniform intervention strategies inadvertently exacerbate cross-category leakage via the privacy seesaw effect. Current methodologies overlook the functional divergence between category-specific neurons (activated solely by single-class private data) and cross-category neurons (responsive to multiple privacy classes). This study introduces a taxonomy-driven editing framework that identifies privacy neurons through integrated gradient attribution and categorizes them via multi-task activation patterns. Our method combines adversarial patching for category-specific neurons with class-conditional regularization for cross-category neurons. Evaluations on BERT and GPT-2 demonstrate that our method can effectively reduce privacy data exposure and leakage while maintaining model utility (perplexity increase <8%). The framework establishes an operational taxonomy for privacy neuron functionality, enabling architecture-specific solutions for regulatory-aligned AI systems.