Cryptography-based privacy-preserving large language models: a lifecycle survey of frameworks, methods, and future directions¶
Authors: Jinglong Luo; Yehong Zhang; Zhuo Zhang; Shiyu Liu; Ye Dong; Haoran Li; Yue Yu; Hui Wang; Xun Zhou; Zenglin Xu
Year: 2025
Venue: Artificial Intelligence Review
type: survey collection: sok-references status: auto
Links: DOI
Summary¶
Auto-extracted summary (unreviewed)
This survey reviews cryptography-based privacy-preserving large language models (CPLMs), i.e., work that applies FHE and MPC across the LLM lifecycle of data selection, fine-tuning and inference. It categorizes privacy leakage and privacy attacks on LLMs, defines threat models, systematically classifies existing CPLM methods and optimization strategies, tabulates open-source libraries, and outlines future research directions.
Abstract¶
The rapid development of Transformer-based large language models (LLMs) has made them one of the most critical technological infrastructures in modern society. However, this rapid deployment has transformed the risk of privacy breaches from a theoretical concern into a systemic threat spanning the entire lifecycle of LLMs. These risks continu ally challenge existing data compliance and regulatory frameworks, directly limiting the large-scale adoption of LLMs in highly sensitive and heavily regulated industries. Crypto graphic technologies, such as fully homomorphic encryption (FHE) and secure multi-party computation (MPC), have garnered significant attention due to their provable security guarantees, theoretically safeguarding the privacy of sensitive data and LLMs weights. These cryptographic techniques have rapidly permeated key stages of LLMs, including data selection, fine-tuning, and inference. Despite these advancements, there is currently no comprehensive survey summarizing the work related to cryptography-based privacypreserving LLMs (CPLMs), leaving their research isolated and fragmented. To fill this gap, We provide a comprehensive review of existing CPLMs research and systematically classifies them, enabling researchers to effectively coordinate optimization strategies for the efficient design of CPLMs algorithms. Finally, based on the limitations of current CPLMs research, we outline several promising directions for future exploration.
Source: pdf
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | first comprehensive review of cryptography-based privacy-preserving LLM research | llm | 0.80 |
| contribution | lifecycle-oriented classification covering data selection, fine-tuning and inference | llm | 0.80 |
| contribution | taxonomy of privacy leakage and privacy attacks (model inversion, membership inference, model extraction, backdoor, attribute inference) and threat models | llm | 0.80 |
| contribution | summary of open-source CPLM libraries | llm | 0.80 |
| contribution | future research directions | llm | 0.80 |
| schemes | FHE | llm | 0.80 |
| schemes | MPC | llm | 0.80 |
| schemes | BFV | llm | 0.80 |
| libraries | SEAL | llm | 0.80 |
| libraries | tfhe-rs | llm | 0.80 |
| libraries | OpenBumbleBee | llm | 0.80 |
| libraries | liberate-fhe | llm | 0.80 |
| libraries | Pyfhel | llm | 0.80 |
| tasks | private LLM data selection | llm | 0.80 |
| tasks | private LLM fine-tuning | llm | 0.80 |
| tasks | private LLM inference | llm | 0.80 |
| threat model | passive (privacy leakage) and active (privacy attack) adversaries over the LLM lifecycle | llm | 0.80 |
| code url | https://github.com/zama-ai/tfhe-rs (URL found in the full text) | regex | 0.50 |
| code url | https://github.com/AntCPLab/OpenBumbleBee (URL found in the full text) | regex | 0.50 |
| code url | https://github.com/Desilo/liberate-fhe (URL found in the full text) | regex | 0.50 |
| llm relevance | high | llm | 0.80 |
SoK reference [31]¶
J. Luo, Y. Zhang, Z. Zhang, S. Liu, Y. Dong, H. Li, Y. Yu, H. Wang, X. Zhou, and Z. Xu, “Cryptography-based privacy-preserving large language models: a lifecycle survey of frameworks, methods, and future directions,” Artificial Intelligence Review, 2025.
Listed on the reference page of the site.
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- scheme: MPC 91 [p.1], Secret sharing 20 [p.3], Function secret sharing 9 [p.25], TEE / confidential computing 8 [p.3], CKKS 7 [p.5], Differential privacy 6 [p.5], Oblivious transfer 5 [p.9], TFHE 3 [p.35], Garbled circuits 3 [p.10], BFV 2 [p.35], Homomorphic secret sharing 2 [p.22]
- library: EzPC / CrypTFlow 5 [p.36], Microsoft SEAL 4 [p.35], Pyfhel 4 [p.35], CrypTen 4 [p.35], TFHE-rs 3 [p.35], SecretFlow-SPU 3 [p.37], HEaaN (library) 2 [p.35], Liberate.FHE 2 [p.35], MP-SPDZ 2 [p.10], PyTorch 2 [p.35]
- model: BERT 42 [p.24], Transformer 21 [p.1], GPT 20 [p.1], LLaMA 12 [p.1], LoRA 12 [p.22], ViT 10 [p.26], RNN / LSTM 4 [p.7], MLP 2 [p.21], AlexNet 1 [p.29], VGG 1 [p.29], CNN 1 [p.7], GAN 1 [p.6], Diffusion model 1 [p.6]
- dataset: GLUE 19 [p.22], ImageNet 6 [p.26], CIFAR-10 5 [p.22], GLUE tasks (QNLI/RTE/MRPC/CoLA/QQP/MNLI/STS-B) 3 [p.25], MNIST 2 [p.22], SST-2 2 [p.30], WikiText 2 [p.26], SQuAD 1 [p.25], IMDB 1 [p.26], MedMNIST 1 [p.23]
- hardware: GPU 2 [p.35], CUDA 1 [p.35], Multi-GPU 1 [p.35]
- technique: Fine-tuning 61 [p.1], Softmax 59 [p.8], GELU 48 [p.8], LayerNorm 31 [p.8], Semi-honest model 20 [p.10], Attention 15 [p.2], Matrix multiplication 14 [p.8], ReLU 13 [p.24], Malicious model 13 [p.10], Membership inference 12 [p.5], Model inversion 7 [p.6], Federated learning 6 [p.22], Quantization 5 [p.12], Two-party computation 5 [p.9], Model extraction / stealing 5 [p.12], Rotations 4 [p.24], Convolution 4 [p.7], Pruning 4 [p.23], Knowledge distillation 4 [p.12], KV cache 3 [p.12], Polynomial approximation 2 [p.31], Goldschmidt 2 [p.24], SIMD packing 2 [p.19], RNS 2 [p.35], Non-interactive 2 [p.24]
- task: Medical / healthcare 11 [p.2], Language modeling / generation 4 [p.7], Finance 3 [p.2]
- topic: Large language models 149 [p.1], Homomorphic encryption 75 [p.1], Private / secure inference 27 [p.3], Open source 15 [p.35], Privacy-preserving machine learning 9 [p.5], Benchmark 3 [p.4]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/sok-references/031_cryptography-based-privacy-preserving-large-language-models.pdf |
sok-references | 45 | 3.4 MB | d352c0d9f6a3 |
Database key: cryptography-based-privacy-preserving-large-language-models-a-lifecycle · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set cryptography-based-privacy-preserving-large-language-models-a-lifecycle COLUMN VALUE