PulseAugur
EN
LIVE 11:05:03

New Shard method enhances privacy in dense retrieval systems

By PulseAugur Editorial · [3 sources] ·

Researchers have developed a new method called Shard to enhance privacy in dense retrieval systems, which are commonly used for semantic search and retrieval-augmented generation (RAG). Shard addresses the vulnerability of vector stores to attacks that can reveal underlying text by introducing a retrieval-preserving embedding transform. This transform splits embeddings into a public prefix for initial retrieval and a private residual sharded under secret keys, which are then reranked using CKKS to cancel keys and maintain exact inner products. The system aims to prevent alignment attacks and de-anonymization, offering a geometric defense rather than a cryptographic guarantee. AI

IMPACT Enhances privacy for semantic search and RAG systems, potentially enabling more secure data retrieval.

RANK_REASON The cluster contains a research paper detailing a new method for private dense retrieval.

Read on arXiv cs.AI →

AI-generated summary · Google Gemini · from 3 sources. How we write summaries →

New Shard method enhances privacy in dense retrieval systems

COVERAGE [3]

  1. arXiv cs.AI TIER_1 Italiano(IT) · Sergey Kurilenko ·

    SHARD: cell-keyed residual splitting for alignment-resistant private dense retrieval

    arXiv:2606.27976v1 Announce Type: cross Abstract: Dense embeddings underpin semantic search and RAG, yet a leaked vector store hands much of the underlying text back to whoever holds it. The attacks that make this possible (few-shot alignment, zero-shot inversion, unsupervised cr…

  2. arXiv cs.IR (Information Retrieval) TIER_1 Italiano(IT) · Sergey Kurilenko ·

    SHARD: cell-keyed residual splitting for alignment-resistant private dense retrieval

    Dense embeddings underpin semantic search and RAG, yet a leaked vector store hands much of the underlying text back to whoever holds it. The attacks that make this possible (few-shot alignment, zero-shot inversion, unsupervised cross-space translation) share one weakness: the pro…

  3. arXiv cs.IR (Information Retrieval) TIER_1 Italiano(IT) · Sergey Kurilenko ·

    SHARD: cell-keyed residual splitting for alignment-resistant private dense retrieval

    Dense embeddings underpin semantic search and retrieval-augmented generation, yet a leaked vector store hands much of the underlying text back. Modern inversion and alignment attacks share one weakness: the protected store is a single global geometry, and any single geometry can …

RELATED ENTITIES

RELATED TOPICS