Personalized Privacy Control in LLMs via Attention Head Intervention

TL;DR - This paper introduces personalized privacy controls for LLMs, along with P3Bench for evaluating adherence to user-specific disclosure preferences. It shows that prompt-only policies are frequently ignored and proposes an inference-time attention-head intervention to improve compliance.

• P3Bench extends contextual privacy evaluation with policies reflecting individual users’ disclosure boundaries.
• Qwen2.5-7B and Gemma3-4B exhibit average policy-ignorance ratios of 51.25% and 74.28%, respectively.
• The proposed Repair method intervenes in attention heads at inference time to steer disclosure behavior.
• Repair reduces policy violations without relying solely on prompt-based instructions.