研究人员提出文本嵌入通用转换方法 同时揭示向量数据库安全隐患

研究人员提出了一种无需配对数据、编码器或预定义匹配集的文本嵌入跨向量空间转换方法1。这一无监督方法基于柏拉图表示假设，即存在通用语义结构，将任何嵌入翻译到统一潜在表示，在不同架构、参数规模和训练数据集的模型间实现高余弦相似度1。

该研究同时揭示了这种转换能力对向量数据库安全的严重威胁1。攻击者仅通过嵌入向量即可提取底层文档的敏感信息1。该论文由Rishi Jha等研究人员完成，最初于2025年5月18日提交，并在2026年1月26日进行了最后修订1。

Researchers led by Rishi Jha have proposed an unsupervised technique for transforming text embeddings across different vector spaces without requiring paired data, encoders, or predefined matching sets 1. The method translates any embedding into a unified latent representation based on the Platonic Representation Hypothesis, which posits the existence of a universal semantic structure underlying different models 1. By leveraging this approach, the researchers achieved high cosine similarity across embeddings from models with varying architectures, parameter scales, and training datasets 1.

The work, first submitted on May 18, 2025, and last revised on January 26, 2026, represents the first translation technique of its kind that requires no external supervision 1. However, the research also reveals a significant security vulnerability in vector databases: attackers can extract sensitive information about underlying documents using only the embedding vectors themselves, enabling document classification and attribute inference 1.