Brief
Brave proposes distributed private text synthesis without a trusted curator
The new method releases DP‑protected embedding summaries and uses a custom secure protocol to guarantee privacy.
By Felo News Desk · Published
Researchers from Inria, Mohamed bin Zayed University of Artificial Intelligence, Brave Software and the National University of Singapore introduced a differentially private (DP) text synthesis approach designed for distributed users, according to Brave's research release on 16 Oct 2026.
The method replaces a trusted curator with a lightweight protocol that first creates a one‑time DP summary in embedding space, identifying frequent semantic regions and publishing DP‑protected centroids. A semantic support protection layer excludes rare text neighborhoods to lower re‑identification risk. A custom secure protocol then enforces end‑to‑end DP guarantees across user devices.
Benchmarks on four datasets show utility comparable to leading centralized DP synthesis methods, while avoiding repeated user participation and heavy synchronization.
Key facts
- The approach releases a one‑time DP summary of embeddings rather than raw text (brave.com)
- Semantic support protection prevents exposure of infrequent text regions (brave.com)
- A custom secure protocol enforces DP guarantees without a trusted curator (brave.com)
- Tests on four benchmarks match utility of state‑of‑the‑art centralized DP synthesis (brave.com)
Sources
- [1] brave.com — originally reported as “Distributed and Private Textual Data Synthesis from Embeddings”
