EMNLP Findings2023
TextMixer: Mixing Multiple Inputs for Privacy-Preserving Inference
Xin Zhou, Yi Lu, Ruotian Ma, Tao Gui, Qi Zhang, Xuanjing Huang
Abstract
Pre-trained language models (PLMs) are often deployed as cloud services, enabling users to upload textual data and perform inference remotely. However, users' personal text often contains sensitive information, and sharing such data directly with the service providers can lead to serious privacy leakage. To address this problem, we introduce a novel privacy-preserving inference framework called TextMixer, which prevents plaintext leakage during the inference phase. Inspired by k-anonymity, TextMixer aims to obfuscate a user's private input by mixing it with multiple other inputs, thereby confounding potential privacy attackers. To achieve this, our approach involves: (1) proposing a novel encryption module, Privacy Mixer, which encrypts input from three distinct dimensions: mixing, representation, and position. (2) adopting a pre-trained Multi-input Multi-output network to handle mixed representations and obtain multiple predictions. (3) employing a Privacy Demixer to ensure only the user can decrypt the real output among the multiple predictions. Furthermore, we explore different ways to automatically generate synthetic inputs required for mixing. Experimental results on token and sentence classification tasks demonstrate that TextMixer greatly surpasses existing privacy-preserving methods in both performance and privacy.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
1 top-tier paper has cited it since. Know what they found before your reviewers do.
Lune can draw on
the full paper · the top-tier papers that cite it · related work
Free to start. No credit card required.
Lineage
- ObfusLM: Privacy-preserving Language Model Service against Embedding Inversion Attacks cites Information Leakage in Embedding Models
- Information Leakage in Embedding Models cites Privacy Risks of General-Purpose Language Models
- Information Leakage in Embedding Models cites Overlearning Reveals Sensitive Attributes
Cited by top-tier papers1
Builds on14
- Privacy Risks of General-Purpose Language ModelsXudong Pan, Mi Zhang, Shouling Ji, Min YangS&P 2020 · 291 citations
- Training independent subnetworks for robust predictionMarton Havasi, Rodolphe Jenatton, Stanislav Fort, Jeremiah Zhe Liu et al.ICLR 2021 · 235 citations
- Iron: Private Inference on TransformersMeng Hao, Hongwei Li, Hanxiao Chen, Pengzhi Xing et al.NeurIPS 2022 · 209 citations
- Information Leakage in Embedding ModelsCongzheng Song, Ananth RaghunathanCCS 2020 · 200 citations
- Overlearning Reveals Sensitive AttributesCongzheng Song, Vitaly ShmatikovICLR 2020 · 177 citations
Related papers
- TextFusion: Privacy-Preserving Pre-trained Model Inference via Token FusionXin Zhou, Jinzhu Lu, Tao Gui, Ruotian Ma et al.EMNLP 2022 · 12 citations
- TextObfuscator: Making Pre-trained Language Model a Privacy Protector via Obfuscating Word RepresentationsXin Zhou, Yi Lu, Ruotian Ma, Tao Gui et al.ACL Findings 2023 · 21 citations
- Anti-adversarial Learning: Desensitizing Prompts for Large Language ModelXuan Li, Zhe Yin, Xiaodong Gu, Beijun ShenAAAI 2026
- MERGE: Fast Private Text GenerationZi Liang, Pinghui Wang, Ruofei Zhang, Nuo Xu et al.AAAI 2024 · 15 citations
- Privacy Preserving In-Context-Learning Framework for Large Language ModelsBishnu Bhusal, Manoj Acharya, Ramneet Kaur, Colin Samplawski et al.AAAI 2026 · 1 citation
