ACL 2025long0 citations

ObfusLM: Privacy-preserving Language Model Service against Embedding Inversion Attacks

Yu Lin, Ruining Yang, Yunlong Mao, Qizhi Zhang, Jue Hong, Quanwei Cai, Ye Wu, Huiqi Liu

Abstract

As the rapid expansion of Machine Learning as a Service (MLaaS) for language models, concerns over the privacy of client inputs during inference or fine-tuning have correspondingly escalated. Recently, solutions have been proposed to safeguard client privacy by obfuscation techniques. However, the solutions incur notable decline in model utility and mainly focus on classification tasks, rendering them impractical for real-world applications. Moreover, recent studies reveal that these obfuscation, if not well designed, is susceptible to embedding inversion attacks (EIAs). In this paper, we devise ObfusLM, a privacy-preserving MLaaS framework for both classification and generation tasks. ObfusLM leverages a model obfuscation module to achieve privacy protection for both classification and generation tasks. Based on (k, 𝜖)-anonymity, ObfusLM includes novel obfuscation algorithms to reach provable security against EIAs. Extensive experiments show that ObfusLM outperforms existing works in utility by 10% with a nearly 80% resistance rate against EIAs.

BibTeX
@inproceedings{lin-etal-2025-obfuslm,
    title = "{O}bfus{LM}: Privacy-preserving Language Model Service against Embedding Inversion Attacks",
    author = "Lin, Yu  and
      Yang, Ruining  and
      Mao, Yunlong  and
      Zhang, Qizhi  and
      Hong, Jue  and
      Cai, Quanwei  and
      Wu, Ye  and
      Liu, Huiqi  and
      Chen, Zhiyu  and
      Duan, Bing  and
      Zhong, Sheng",
    editor = "Che, Wanxiang  and
      Nabende, Joyce  and
      Shutova, Ekaterina  and
      Pilehvar, Mohammad Taher",
    booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.acl-long.58/",
    doi = "10.18653/v1/2025.acl-long.58",
    pages = "1160--1174",
    ISBN = "979-8-89176-251-0"
}
ObfusLM: Privacy-preserving Language Model Service against Embedding Inversion Attacks · ACL 2025