有限查詢存取下的文本嵌入逆推攻擊

Abstract

本文提出了一個針對文本嵌入逆推攻擊的解決方法，適用於當查詢數量有限的情境。為了應對這個挑戰，我們提出了一個名為"輔助文檔嵌入攻擊與查詢選擇（SADE）"的模型，旨在適應實際應用的場景。具體而言，我們引入了一種利用外部文檔的方法來克服當私有文檔非常有限的情況，這在實際場景中是一個常見的問題。此外，我們提出了一種新穎的查詢策略，解決了當私有編碼器的查詢數量非常有限的情況。與之前的方法不同的是，我們的方法利用少量的查詢即可實現高度的準確性，而不需要大量的私有文檔和無限制的查詢私有編碼器。整體來說，我們提出的模型和查詢策略證明了我們的方法在文本嵌入逆推攻擊中的有效性和實用性。

This paper proposes a solution for embedding inversion attack of textual embedding in scenarios where the number of queries is limited. To address this challenge, we propose a model called Surrogate-Assisted Document Embedding Attack with Query Selection (SADE) that is designed to fit practical scenarios. Specifically, we introduce a means to exploit external documents to overcome the challenge of limited private documents, which is a common issue in practice. Additionally, we propose a novel query strategy that addresses the issue of limited queries to the private encoder. Unlike previous works that require a large number of private documents and unlimited query access, our approach utilizes a small number of queries to achieve high retrieval accuracy. Overall, our proposed model and query strategy demonstrate the effectiveness and practicality of our approach for embedding inversion attacks of textual embedding.