EVENT DOSSIER
In response to the threat of embedding reversal attacks in vector databases, researchers proposed the SHAQ (Shadow Query Generation) defense mechanism. This method utilizes generative language models to create diverse shadow queries to replace directly stored document embeddings. Through semantic decomposition and embedding decoupling, it breaks the strong coupling between embeddings and original texts. Experimental results show that SHAQ improves privacy protection while maintaining retrieval efficiency: the recovery rate is reduced to 0.2104, protecting 19.50% more tokens compared to baseline methods, and achieving a MAP@10 score of 0.7967, resulting in a 5.53% improvement in efficiency.
提出 SHAQ（阴影查询生成）防御向量数据库中的嵌入反转攻击。SHAQ 利用生成式语言模型创建多样化阴影查询以替代直接存储文档嵌入，通过语义分解和嵌入解耦切断嵌入与原文的强耦合关系。实验表明，该方法在提升隐私保护的同时保持检索效用，将恢复率降至 0.2104，防御效果比基线方法多保护 19.50% 的 token，并在 MAP@10 指标上达到 0.7967，效用提升达 5.53%。