[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-seo-148675-105":3,"detail-sidebar-cat-0-en-105":81,"doc-detail-148675-en":131},{"code":4,"msg":5,"data":6},0,"ok",{"site_id":7,"language":8,"slug":9,"title":10,"keywords":11,"description":12,"schema_data":13,"social_meta":74,"head_meta":76,"extra_data":78,"updated_unix":80},105,"en","the-models-language-matters-a-comparative-privacy-analysis-of-llms","The Model’s Language Matters - A Comparative Privacy Analysis of LLMs","","Large Language Models (LLMs) are increasingly used in multilingual settings where sensitive data is processed, yet scale and linguistic variability can intensify privacy exposure. This work studies how language structure affects privacy leakage in LLMs trained on English, Spanish, French, and Italian medical corpora. Six corpus-level linguistic indicators are quantified and vulnerability is tested under extraction, counterfactual memorization, and membership inference attacks. Results show leakage follows structural properties: Italian has strongest exposure, English yields clearest membership separability, while French and Spanish are comparatively more resilient.",{"@graph":14,"@context":73},[15,34,56],{"@type":16,"itemListElement":17},"BreadcrumbList",[18,23,27,31],{"item":19,"name":20,"@type":21,"position":22},"https://docshare.wps.com","Home","ListItem",1,{"item":24,"name":25,"@type":21,"position":26},"https://docshare.wps.com/document/","Document",2,{"item":28,"name":29,"@type":21,"position":30},"https://docshare.wps.com/document/research-report/","Research & Report",3,{"item":32,"name":10,"@type":21,"position":33},"https://docshare.wps.com/document/the-models-language-matters-a-comparative-privacy-analysis-of-llms/148675/",4,{"url":32,"name":10,"@type":35,"image":36,"author":41,"headline":10,"publisher":44,"fileFormat":47,"inLanguage":8,"description":12,"dateModified":48,"datePublished":49,"encodingFormat":47,"isAccessibleForFree":50,"interactionStatistic":51},"DigitalDocument",{"url":37,"@type":38,"width":39,"height":40},"https://docshare.wps.com/thumbnails/the-models-language-matters-a-comparative-privacy-analysis-of-llms/148675.png","ImageObject",300,407,{"name":42,"@type":43},"Tawan","Person",{"url":19,"name":45,"@type":46},"DocShare","Organization","application/pdf","2026-10-16","2026-08-26",true,{"@type":52,"interactionType":53,"userInteractionCount":55},"InteractionCounter",{"@type":54},"ViewAction",17,{"@type":57,"mainEntity":58},"FAQPage",[59,65,69],{"name":60,"@type":61,"acceptedAnswer":62},"What question does the paper investigate about multilingual LLMs?","Question",{"text":63,"@type":64},"It examines how the linguistic structure of different languages influences privacy leakage in LLMs trained on language-specific medical corpora.","Answer",{"name":66,"@type":61,"acceptedAnswer":67},"Which languages and data domains are evaluated?",{"text":68,"@type":64},"The study evaluates English, Spanish, French, and Italian using medical corpora to reflect sensitive-domain deployment conditions.",{"name":70,"@type":61,"acceptedAnswer":71},"How is privacy vulnerability measured in the experiments?",{"text":72,"@type":64},"The paper quantifies six corpus-level linguistic indicators and evaluates attacks from three families: extraction, counterfactual memorization, and membership inference.","https://schema.org",{"og:url":32,"og:type":75,"og:title":10,"og:site_name":45,"og:description":12},"article",{"robots":77,"canonical":32},"index,follow",{"doc_id":79,"site_id":7},148675,1787782936,{"code":4,"msg":82,"data":83},"success",[84,88,92,96,101,106,111,115,120,123,127],{"id":22,"doc_module":4,"doc_module_name":25,"category_name":85,"show_sort_weight":86,"slug":87},"Story & Novel",90,"story-novel",{"id":26,"doc_module":4,"doc_module_name":25,"category_name":89,"show_sort_weight":90,"slug":91},"Literature",80,"literature",{"id":33,"doc_module":4,"doc_module_name":25,"category_name":93,"show_sort_weight":94,"slug":95},"Exam",70,"exam",{"id":97,"doc_module":4,"doc_module_name":25,"category_name":98,"show_sort_weight":99,"slug":100},5,"Comic",60,"comic",{"id":102,"doc_module":4,"doc_module_name":25,"category_name":103,"show_sort_weight":104,"slug":105},6,"Technology",50,"technology",{"id":107,"doc_module":4,"doc_module_name":25,"category_name":108,"show_sort_weight":109,"slug":110},7,"Healthcare",40,"healthcare",{"id":112,"doc_module":4,"doc_module_name":25,"category_name":29,"show_sort_weight":113,"slug":114},8,30,"research-report",{"id":116,"doc_module":4,"doc_module_name":25,"category_name":117,"show_sort_weight":118,"slug":119},9,"Religion & Spirituality",20,"religion-spirituality",{"id":118,"doc_module":4,"doc_module_name":25,"category_name":121,"show_sort_weight":118,"slug":122},"World Cup","world-cup",{"id":124,"doc_module":4,"doc_module_name":25,"category_name":125,"show_sort_weight":124,"slug":126},10,"Lifestyle","lifestyle",{"id":128,"doc_module":4,"doc_module_name":25,"category_name":129,"show_sort_weight":97,"slug":130},19,"General","general",{"code":4,"msg":82,"data":132},{"doc_id":79,"user_id":133,"nickname":42,"user_avatar":134,"doc_module":4,"category_id":112,"category_name":29,"doc_title":10,"doc_description":12,"doc_content":135,"file_id":136,"file_url":137,"file_type":138,"file_size":139,"view_count":55,"is_deleted":4,"is_public":22,"is_downloadable":22,"audit_status":22,"page_count":140,"language":141,"language_code":8,"site_id":7,"html_lang":8,"table_of_contents":142,"faqs":143,"seo_title":144,"seo_description":12,"update_tm":80,"read_time":145,"preview_page_count":4},2336475104042,"https://ap-avatar.wpscdn.com/avatar/22000c4c32af1715be0?x-image-process=image/resize,m_fixed,w_180,h_180&k=1786537525561427321","The Model’s Language Matters: A Comparative Privacy Analysis of LLMs  \nAbhishek Kumar Mishra  \nInria [abhishek.mishra@inria.fr](abhishek.mishra@inria.fr)  \nAntoine Boutet  \nINSA Lyon, Inria CITI, UR3720  \nLucas Magnana  \nInria, INSA Lyon CITI, UR3720  \n[antoine.boutet@insa-lyon.fr](antoine.boutet@insa-lyon.fr) [lucas.magnana@inria.fr](lucas.magnana@inria.fr)  \nAbstract  \nLarge Language Models (LLMs) are increasingly deployed in multilingual settings that process sensitive data, yet their scale and linguistic variability can amplify privacy risks. While prior privacy evaluations focus predominantly on English, we investigate how language structure shapes privacy leakage in LLMs trained on English, Spanish, French, and Italian medical corpora. We quantify six corpus-level linguistic indicators and evaluate vulnerability under three attack families: extraction, counterfactual memorization, and membership inference.  \nAcross languages, we find that leakage systematically tracks structural properties: Italian exhibits the strongest exposure, consistent with its highest redundancy and longer lexical units, whereas English shows the clearest membership separability, aligning with its higher syntactic entropy and stronger surface-identifiable cues. In contrast, French and Spanish remain comparatively more resilient overall, aided by higher morphological complexity. These results provide quantitative evidence that language matters for privacy leakage, motivating language-aware and structure-adaptive privacypreserving mechanisms for multilingual LLM deployments.  \n1 Introduction  \nRapid advances in natural language processing (NLP) have fueled its adoption in many industries worldwide. Large language models (LLMs) such as BERT and GPT have been pre-trained at great expense on countless unlabeled datasets extracted from the Web. While these models represent incredible potential and promises, their large-scale deployment and their complexity, as well as the fact that they interact with and potentially influence individuals, raise multiple security and privacy concerns (Das et al., 2025) .  \nThe attack surface on models is still poorly understood (Weidinger et al., 2022 ; Lehman et al.,  \n2021 ; Duprieu and Berkouk, 2024 ; Chen et al., 2024) . A number of threats are related to the memorization and possible leakage of sensitive information used during model training, such as data reconstruction and membership inference (i.e., identifying elements used during the training or the fine-tuning) . Memorization of information by a model is not a problem in itself. However, this memorization becomes a problem when the training information is not generalized enough by the model which reproduces large portions of training data verbatim or discloses some sensitive information (Kassem et al., 2023 ; Wu et al., 2023) .  \nMost privacy risk assessment work has been conducted on English texts (Li et al., 2024 ; Shanmugarasa et al., 2025) . However, the language of the texts, their structures, and their characteristics inherently impact LLM memorization and, consequently, privacy risks. Although the language considered is well known to potentially introduce bias in some results, the impact of language on privacy risks has not yet been explored to our knowledge. To overcome this limitation, in this paper, we empirically explore the impact of language on privacy risks associated with LLMs. We also analyze the main characteristics and structures of each language and link them to various privacy vulnerabilities. Specifically, we comprehensively assess the privacy of LLMs fine-tuned on English, Spanish, French, and Italian medical corpora using an extraction attack, a membership inference attack and counterfactual memorization. Results show that privacy vulnerability scales with linguistic redundancy and tokenization granularity: Italian presents the highest leakage, while English has higher membership separability. In contrast, French and Spanish show greater","cbCainechtM9ugX7","https://ap.wps.com/l/cbCainechtM9ugX7","pdf",672594,11,"English","# Introduction\n# Background and Related Work\n## Extractable memorization","[{\"question\":\"What question does the paper investigate about multilingual LLMs?\",\"answer\":\"It examines how the linguistic structure of different languages influences privacy leakage in LLMs trained on language-specific medical corpora.\"},{\"question\":\"Which languages and data domains are evaluated?\",\"answer\":\"The study evaluates English, Spanish, French, and Italian using medical corpora to reflect sensitive-domain deployment conditions.\"},{\"question\":\"How is privacy vulnerability measured in the experiments?\",\"answer\":\"The paper quantifies six corpus-level linguistic indicators and evaluates attacks from three families: extraction, counterfactual memorization, and membership inference.\"}]","The Model’s Language Matters - A Comparative Privacy Analysis of LLMs | PDF",28]