Skip to content
InfoResearchPreprintLLM-specific

Ask in the Crowd: Differentially Private LLM Inference via Dummy-Augmented Shuffling

Published
Record updated
View JSON

Summary

The paper proposes DP-SPD, a differentially private inference framework for LLM-as-a-Service that hides a user's real query among client-generated dummy queries and randomly shuffles all queries before they reach the server. It reuses transformer key-value caches across the shared prefixes of the queries to improve decoding efficiency. The authors report strong inference-time privacy and near-plaintext utility on classification and generation tasks, evaluated against QIA, EIA and GPT-based attacks.