The Model's Language Matters: A Comparative Privacy Analysis of LLMs

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Mishra, Abhishek K., Boutet, Antoine, Magnana, Lucas
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866918157911851008
author Mishra, Abhishek K.
Boutet, Antoine
Magnana, Lucas
author_facet Mishra, Abhishek K.
Boutet, Antoine
Magnana, Lucas
contents Large Language Models (LLMs) are increasingly deployed across multilingual applications that handle sensitive data, yet their scale and linguistic variability introduce major privacy risks. Mostly evaluated for English, this paper investigates how language structure affects privacy leakage in LLMs trained on English, Spanish, French, and Italian medical corpora. We quantify six linguistic indicators and evaluate three attack vectors: extraction, counterfactual memorization, and membership inference. Results show that privacy vulnerability scales with linguistic redundancy and tokenization granularity: Italian exhibits the strongest leakage, while English shows higher membership separability. In contrast, French and Spanish display greater resilience due to higher morphological complexity. Overall, our findings provide the first quantitative evidence that language matters in privacy leakage, underscoring the need for language-aware privacy-preserving mechanisms in LLM deployments.
format Preprint
id arxiv_https___arxiv_org_abs_2510_08813
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle The Model's Language Matters: A Comparative Privacy Analysis of LLMs
Mishra, Abhishek K.
Boutet, Antoine
Magnana, Lucas
Computation and Language
Cryptography and Security
Large Language Models (LLMs) are increasingly deployed across multilingual applications that handle sensitive data, yet their scale and linguistic variability introduce major privacy risks. Mostly evaluated for English, this paper investigates how language structure affects privacy leakage in LLMs trained on English, Spanish, French, and Italian medical corpora. We quantify six linguistic indicators and evaluate three attack vectors: extraction, counterfactual memorization, and membership inference. Results show that privacy vulnerability scales with linguistic redundancy and tokenization granularity: Italian exhibits the strongest leakage, while English shows higher membership separability. In contrast, French and Spanish display greater resilience due to higher morphological complexity. Overall, our findings provide the first quantitative evidence that language matters in privacy leakage, underscoring the need for language-aware privacy-preserving mechanisms in LLM deployments.
title The Model's Language Matters: A Comparative Privacy Analysis of LLMs
topic Computation and Language
Cryptography and Security
url https://arxiv.org/abs/2510.08813