Do Text Embeddings Perfectly Encode Text?
Text embeddings, widely used in Retrieval Augmented Generation (RAG) systems and stored in vector databases, are thought to secure text data. However, new research shows that text embeddings can be inverted to recover the original text with high accuracy, raising serious privacy and security concerns.
The article discusses the rise of vector databases for RAG systems, which use embeddings to represent documents. Embeddings are vector representations from neural networks, and they appear as random numbers, prompting many to consider them secure. However, the paper "Text Embeddings Reveal As Much as Text" (EMNLP 2023) demonstrates that text can be recovered from embeddings. The researchers developed vec2text, an iterative optimization method that, after 50 steps, recovers 92% of 32-token sequences exactly and achieves a BLEU score of 97. This proves that embeddings retain nearly all information from the input text, contradicting assumptions of security. The article also notes that similar inversion is possible in computer vision and raises questions about defending against such attacks and the relationship between text length and embedding size.
- Сокращения
- RAG = Retrieval Augmented Generation — Генерация с дополнением по поиску
- BLEU = Bilingual Evaluation Understudy — Двуязычная оценка подмастерья
- CNN = Convolutional Neural Network — Свёрточная нейронная сеть
Source: The Gradient —
original