PINS: A Pivot-based Intersection Noise Scheme for Dense Embeddings Privacy in Information Retrieval
Abstract
Modern Information Retrieval (IR) systems, including small and medium-scale ones, often search across millions of documents. In many cases, data owners outsource retrieval to a third-party host to keep the data accessible. Furthermore, IR has also moved toward semantic matching with dense embeddings; although these embeddings are less interpretable than sparse representations, they remain vulnerable to model inversion attacks. Sending document embeddings to a third-party retrieval provider can, therefore, create privacy risks. An honest-but-curious host may try to reconstruct the original documents and exploit their content. One option is to send encrypted or obfuscated embeddings instead of the original vectors. Existing methods, however, have important limits: Asymmetric Scalar Product-preserving Encryption (ASPE) is vulnerable to inversion, while homomorphic encryption is often too costly. Building on the ASPE literature, we propose the Pivot-based Intersection Noise Scheme (PINS), a noise-injection method designed for dense IR. PINS aims to preserve retrieval quality while reducing leakage from document embeddings. It selects a set of pivots to define an obfuscation space, then chooses or samples an obfuscation point in that space for each document embedding. When measuring privacy-utility trade-off in terms of AUC, PINS reaches up to twice the effectiveness of state-of-the-art embedding obfuscation methods and also works with asymmetric bi-encoders, overcoming the current state-of-the-art. To favour reproducibility, we release PINS at https://github.com/guglielmof/2026-CIKM-PINS.