Personalized Privacy Control in LLMs via Attention Head Intervention
Researchers propose a new approach to controlling personalized privacy in large language models (LLMs). They introduce 'personalized privacy' that considers individual users' disclosure preferences. A benchmark called P3Bench is developed to test this concept. However, existing methods fail to enforce these policies reliably. To address this issue, the authors suggest a method called Repair, which adjusts the model's attention heads at inference time to ensure policy-consiste
Researchers propose a new approach to controlling personalized privacy in large language models (LLMs). They introduce 'personalized privacy' that considers individual users' disclosure preferences. A benchmark called P3Bench is developed to test this concept. However, existing methods fail to enforce these policies reliably. To address this issue, the authors suggest a method called Repair, which adjusts the model's attention heads at inference time to ensure policy-consistent responses.
---
Why it matters: This matters because LLMs often access sensitive user data, raising concerns about privacy. The ability to personalize privacy control can help mitigate these issues and improve trust in AI systems.
Source: https://arxiv.org/abs/2608.21209
This article was originally published at: https://arxiv.org/abs/2608.21209