Where Privacy Risk Lives in English-Source Multilingual RAG: A Stage-Decomposed Audit Across Five Query Languages

The study challenges the common assumption that non-English queries inherently weaken privacy protections in multilingual RAG. The authors built an English-source corpus with synthetic personal information and tested five query languages, using a Qwen2.5-7B-based pipeline with translator, judge, back-translator, and generator components. A two-stage defense (LLM input judge plus regex output filter) was evaluated stage-by-stage. Findings are explicitly pipeline-conditional, meaning results may shift with different models or defenses. The paper contributes a stage-decomposed audit methodology, which could help developers localize where privacy leaks occur in multilingual RAG deployments.