DP-SPD hides real queries among dummy prompts for private LLM inference
Researchers propose a differentially private inference framework that mixes a user's real query with client-generated dummy queries and shuffles them before they reach the server.
In plain words
When people send prompts to a hosted language model, the service can see the sensitive text. DP-SPD, a proposed research framework, hides the real prompt among fake ones generated on the user's side and shuffles them so the server cannot tell which is genuine. The authors report the approach keeps accuracy close to normal inference while adding privacy protection.
What was announced
The paper, posted on arXiv, addresses privacy risks in LLM-as-a-Service deployments, where queries sent for remote inference can expose sensitive user inputs. The authors argue that existing differentially private methods, which inject noise into input embeddings, struggle to balance privacy and utility because stronger protection often degrades accuracy. DP-SPD instead hides the real query among client-generated dummy queries, and all queries are randomly shuffled before reaching the server, preventing direct identification of the user's true intent.
To limit the cost of processing extra queries, the framework exploits shared-prefix structure among the queries and reuses transformer key-value (KV) caches during decoding. The authors report a privacy analysis, evaluations on classification and generation tasks, and tests against query-injection, embedding-inversion, and GPT-based attacks, describing strong inference-time protection, near-plaintext utility, and improved efficiency from the cache reuse. These results come from the paper's own experiments and have not been independently confirmed.
What it does not show
- The evidence is the paper's abstract only; full method details, experimental setup, and exact results are not available in the supplied passages.
- Reported utility and attack-resistance results are the authors' own evaluations and do not constitute independent confirmation.
For security and IT teams
- Scope
- DP-SPD is a research framework described in an arXiv abstract; the passages do not state any product release, deployment, or availability.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.