P2Skill: Privacy-Preserving Skill Distillation for Cloud-Local LLM Inference
A recent study published on arXiv (2608.14094) presents P2Skill, a novel method for skill distillation based on prompts, aimed at enhancing cloud-local LLM inference systems. These systems strive to merge the analytical capabilities of extensive cloud models with the confidentiality of on-device processing. To safeguard personally identifiable information (PII), it is crucial to sanitize requests sent to the cloud. Current techniques such as prompt perturbation, entity masking, or fine-tuning can alter semantics or necessitate additional training. In contrast, P2Skill employs a local small language model (SLM) that independently executes decomposition, PII-aware routing, paraphrasing, and reconstruction by adhering to skill prompts. Skills are progressively improved through feedback from a cloud LLM, enabling the SLM to extend its capabilities beyond memorized PII patterns without needing privacy-specific fine-tuning or additional detectors. The paper’s evaluation details are briefly noted but incomplete. It is classified as a cross-type announcement, suggesting potential presentations at various venues.
Key facts
- Paper arXiv:2608.14094 proposes P2Skill, a prompt-based skill distillation method.
- P2Skill targets cloud-local LLM inference systems to protect user privacy.
- It addresses PII leakage by having a local SLM perform decomposition, PII-aware routing, paraphrasing, and reconstruction.
- Skills are refined iteratively from execution failures by a cloud LLM.
- P2Skill requires no privacy-specific fine-tuning or learned auxiliary detectors.
- Existing methods include prompt perturbation, entity masking, and model fine-tuning.
- The paper is a cross-type announcement on arXiv.
- Evaluation details are incomplete in the source.
Entities
Institutions
- arXiv