Image de couverture de livre

Concept-Aware Privacy Mechanisms for Defending Embedding Inversion Attacks

Fuente: arXiv

Enregistré dans:

Détails bibliographiques
Auteurs principaux:	Tsai, Yu-Che, Hsiao, Hsiang, Chen, Kuan-Yu, Lin, Shou-De
Format:	Preprint
Publié:	2026
Sujets:	Cryptography and Security Artificial Intelligence
Accès en ligne:	Acceder al recurso
Tags:	Ajouter un tag Pas de tags, Soyez le premier à ajouter un tag!

_version_	1866914311769686016
author	Tsai, Yu-Che Hsiao, Hsiang Chen, Kuan-Yu Lin, Shou-De
author_facet	Tsai, Yu-Che Hsiao, Hsiang Chen, Kuan-Yu Lin, Shou-De
contents	Text embeddings enable numerous NLP applications but face severe privacy risks from embedding inversion attacks, which can expose sensitive attributes or reconstruct raw text. Existing differential privacy defenses assume uniform sensitivity across embedding dimensions, leading to excessive noise and degraded utility. We propose SPARSE, a user-centric framework for concept-specific privacy protection in text embeddings. SPARSE combines (1) differentiable mask learning to identify privacy-sensitive dimensions for user-defined concepts, and (2) the Mahalanobis mechanism that applies elliptical noise calibrated by dimension sensitivity. Unlike traditional spherical noise injection, SPARSE selectively perturbs privacy-sensitive dimensions while preserving non-sensitive semantics. Evaluated across six datasets with three embedding models and attack scenarios, SPARSE consistently reduces privacy leakage while achieving superior downstream performance compared to state-of-the-art DP methods.
format	Preprint
id	arxiv_https___arxiv_org_abs_2602_07090
institution	arXiv
publishDate	2026
record_format	arxiv
spellingShingle	Concept-Aware Privacy Mechanisms for Defending Embedding Inversion Attacks Tsai, Yu-Che Hsiao, Hsiang Chen, Kuan-Yu Lin, Shou-De Cryptography and Security Artificial Intelligence Text embeddings enable numerous NLP applications but face severe privacy risks from embedding inversion attacks, which can expose sensitive attributes or reconstruct raw text. Existing differential privacy defenses assume uniform sensitivity across embedding dimensions, leading to excessive noise and degraded utility. We propose SPARSE, a user-centric framework for concept-specific privacy protection in text embeddings. SPARSE combines (1) differentiable mask learning to identify privacy-sensitive dimensions for user-defined concepts, and (2) the Mahalanobis mechanism that applies elliptical noise calibrated by dimension sensitivity. Unlike traditional spherical noise injection, SPARSE selectively perturbs privacy-sensitive dimensions while preserving non-sensitive semantics. Evaluated across six datasets with three embedding models and attack scenarios, SPARSE consistently reduces privacy leakage while achieving superior downstream performance compared to state-of-the-art DP methods.
title	Concept-Aware Privacy Mechanisms for Defending Embedding Inversion Attacks
topic	Cryptography and Security Artificial Intelligence
url	https://arxiv.org/abs/2602.07090