LLM‑as‑a‑Service has become the dominant way to access large language models without local deployment, yet user queries often contain sensitive information, raising serious privacy concerns. Differential privacy (DP) offers a rigorous protection mechanism, but current DP inference techniques are designed for short or moderate contexts and struggle with million‑token inputs because of limited context capacity and severe utility loss from DP noise.
DP‑MIND introduces a memory‑guided framework that scales DP inference to million‑token contexts. The method first perturbs the raw input locally with DP noise, then processes the noisy text chunk by chunk while maintaining a compact memory that preserves task‑relevant information across chunks. To make the memory robust to DP noise, DP‑MIND directly optimizes the memory construction policy on perturbed inputs via reinforcement learning, using a reward that balances final answer correctness and memory consistency.
After the model generates outputs, a local denoising module restores answer spans corrupted by noise, all without exposing the original context to external models. We provide a formal privacy analysis showing that DP‑MIND satisfies $(\epsilon,\delta)$‑DP. Extensive experiments on several long‑context benchmarks, with context lengths up to 1 M tokens, demonstrate that DP‑MIND consistently outperforms existing DP and long‑context inference baselines in both scalability and utility.
Review: By integrating a memory mechanism with reinforcement‑learning‑driven optimization, DP‑MIND overcomes the long‑context limitation of differential privacy, offering a practical solution for privacy‑preserving LLM‑as‑a‑Service.