The authors address a gap in privacy-preserving LLM inference: prior DP methods were designed for short or moderate contexts and fail at million-token scales because of limited context capacity, severe utility loss under DP noise, and high inference overhead.
Memory-guided framework brings differential privacy to million-token LLM inference
DP-MIND chunks and locally perturbs long inputs, propagates a task-relevant memory between chunks, and denoises answer spans without exposing clean context to the remote model.
Chinese Tech
Zhihao Liu · Yitian Cheng · Shuang Liu · Rui Wen · Yuke Hu · Zhan Qin
Zhejiang University · Ant Group · Institute of Science Tokyo · King Abdullah University of Science and Technology
Research Digest··3 min read
Zhihao Liu and colleagues at Zhejiang University, Ant Group, the Institute of Science Tokyo and KAUST propose DP-MIND, a differential privacy (DP) inference framework that scales to contexts of up to a million tokens.
Why this paper
From Ant Group and 3 others
In one line
DP-MIND enables private inference on million-token contexts by perturbing chunks, propagating memory, and local denoising.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§