Recoverability, utility, and source-content exposure in compressed multilingual text representations

Reusable text representations are routinely cached and compressed, yet downstream utility and source-content exposure are rarely measured together. We evaluate token sequences and pooled vectors across six languages, five encoders, classification and dense retrieval, matched attacker budgets, and transformation-aware attacks. In a five-seed mBERT experiment on parallel MASSIVE data, mean token F1 was \(0.9370\pm 0.0030\) from full token states and \(0.7211\pm 0.0333\) from one mean-pooled vector; the full-sequence advantage occurred in English, Korean, Arabic, Japanese, Swahili, and Telugu.

On six-language Belebele retrieval, clean macro nDCG@10 was 0.9027 for multilingual-E5-base and 0.9220 for BGE-M3; int8 retained 0.9025/0.9217, whereas PCA-64 fell to 0.7984/0.8043. Five-seed, content-disjoint attack curves show that rankings depend on paired-data and compute budgets. Fully transformation-aware strong-mT5 and Vec2Text attacks find little protection from int8 or masking and only partial reductions from PCA-256 or Gaussian noise.

A condition-blinded audit by the three study authors supports higher factual recoverability from full Korean token sequences and higher sensitive-content recovery from full English token sequences, while other contrasts remain uncertain. Actual payload accounting and Pareto analysis further show that storage reduction, utility, and leakage are not interchangeable. Granularity remains an important determinant under matched conditions, but exposure depends jointly on language, task, encoder, attacker access and adaptation, and representation budget.

Department of Military Operations Research, Korea National Defense University, Nonsan, Republic of Korea Interdisciplinary Program in Computational Sciences and Technology, College of Natural Sciences, Seoul National University, Seoul, Republic of Korea The authors declare no competing interests. The reported condition-blinded output-coding audit was conducted solely by the three study authors as part of model evaluation. Only their ratings are analyzed and reported; no external-rater data or personal data are used in this study.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it.

The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.