Skip to content

Towards efficient privacy-preserving machine learning: A systematic review from protocol, model, and system perspectives

Authors: W. Zeng; T. Xu; Y. Chen; Y. Zhou; M. Zhang; J. Tan; C. Hong; M. Li
Year: 2025

type: survey collection: sok-references status: auto

Links: ePrint 2025/1365

Summary

Auto-extracted summary (unreviewed)

A systematic review of cryptographic privacy-preserving machine learning (mainly two-party client-server inference) organized by cross-level optimizations: protocol level (HE- and secret-sharing-based linear and non-linear layers, computation-graph optimizations), model level (PPML-friendly linear and non-linear layers, softmax/ReLU/GELU approximation, quantization, pruning, LoRA) and system level (compilers, GPU acceleration, libraries). It gives qualitative and quantitative comparisons, surveys open-source libraries, discusses future directions, and maintains a public GitHub paper list.

Extracted details

Field Value Source Confidence
contribution Taxonomy of PPML optimizations across protocol, model and system levels llm 0.80
contribution Review of HE-based and secret-sharing-based protocols for linear and non-linear layers llm 0.80
contribution Review of PPML-friendly model designs for CNNs and Transformers/LLMs llm 0.80
contribution Survey of compilers, GPU acceleration and open-source PPML libraries with comparisons llm 0.80
contribution Discussion of future directions and a continuously updated GitHub repository llm 0.80
schemes secret sharing llm 0.80
schemes homomorphic encryption (CKKS/BFV) llm 0.80
schemes TFHE llm 0.80
schemes oblivious transfer llm 0.80
schemes garbled circuits llm 0.80
schemes function secret sharing llm 0.80
schemes replicated secret sharing llm 0.80
libraries Lattigo llm 0.80
libraries TFHE-rs llm 0.80
libraries HElib llm 0.80
libraries python-paillier llm 0.80
libraries HElayers llm 0.80
libraries OpenFHE llm 0.80
tasks private CNN inference llm 0.80
tasks private Transformer/LLM inference llm 0.80
tasks two-party privacy-preserving machine learning llm 0.80
threat model two-party client-server (2PC); server learns nothing about input, client learns nothing about weights llm 0.80
code url https://github.com/PKU-SEC-Lab/Awesome-PPML-Papers.CCS (URL found in the full text) regex 0.50
code url https://github.com/PKU-SEC-Lab/Awesome-PPML-Papers.Taxonomy (URL found in the full text) regex 0.50
code url https://github.com/homenc/HElib (URL found in the full text) regex 0.50
code url https://github.com/PKU-SEC-Lab/Awesome-PPML-Papers llm 0.80
limitations Primarily focuses on two-party computation; three or more parties only briefly discussed llm 0.80
llm relevance medium llm 0.80

SoK reference [32]

W. Zeng, T. Xu, Y. Chen, Y. Zhou, M. Zhang, J. Tan, C. Hong, and M. Li, “Towards efficient privacy-preserving machine learning: A systematic review from protocol, model, and system perspectives,” Cryptology ePrint Archive, Paper 2025/1365, 2025. [Online]. Available: https://eprint.iacr.org/2025/1365

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: CKKS 38 [p.6], MPC 38 [p.1], TFHE 29 [p.4], BFV 17 [p.6], Function secret sharing 17 [p.8], Secret sharing 14 [p.2], BGV 13 [p.6], FHEW 5 [p.28], Oblivious transfer 5 [p.2], Paillier 4 [p.30], RSA 2 [p.34], TEE / confidential computing 2 [p.1], PIR 2 [p.29], Leveled / somewhat HE 2 [p.16]
  • library: EVA / CHET 19 [p.24], EzPC / CrypTFlow 12 [p.7], CrypTen 11 [p.21], Microsoft SEAL 9 [p.18], ABY / ABY3 9 [p.7], TFHE-rs 6 [p.28], Cheddar 6 [p.28], Fhelipe 6 [p.24], nGraph-HE 6 [p.16], Orion 5 [p.24], HElib 3 [p.30], Lattigo 3 [p.21], Pyfhel 3 [p.30], Phantom 3 [p.27], MP-SPDZ 3 [p.30], OpenFHE 2 [p.30], HEaaN (library) 2 [p.22], Zama 2 [p.28], SecretFlow-SPU 2 [p.37], Concrete (Zama) 1 [p.24], HEonGPU 1 [p.29]
  • model: CNN 49 [p.3], Transformer 26 [p.1], ViT 23 [p.17], ResNet 18 [p.7], GPT 13 [p.5], RNN / LSTM 10 [p.28], LLaMA 7 [p.5], Logistic regression 2 [p.28], LoRA 2 [p.4], LeNet 1 [p.26], MobileNet 1 [p.11], MLP 1 [p.29]
  • dataset: MNIST 3 [p.7], CIFAR-10 3 [p.27], ImageNet 3 [p.7]
  • hardware: GPU 45 [p.3], NVIDIA A100 5 [p.28], CUDA 5 [p.27], NVIDIA RTX 2 [p.29], NVIDIA V100 1 [p.29]
  • technique: Bootstrapping 71 [p.4], Convolution 59 [p.3], Matrix multiplication 53 [p.9], ReLU 50 [p.2], Quantization 47 [p.2], GELU 43 [p.4], Softmax 38 [p.4], Rotations 30 [p.6], Pruning 29 [p.2], SIMD packing 27 [p.4], Compiler 25 [p.3], Two-party computation 19 [p.1], Rescaling 17 [p.4], Polynomial approximation 13 [p.2], Lookup table 13 [p.6], Non-interactive 10 [p.2], KV cache 9 [p.5], Semi-honest model 8 [p.7], Programmable bootstrapping 7 [p.15], Attention 7 [p.5], NTT 7 [p.4], RNS 5 [p.21], Fine-tuning 5 [p.17], Key switching 4 [p.4], Three-party computation 4 [p.7]
  • assumption: RLWE 5 [p.6], LWE 2 [p.6]
  • task: Language modeling / generation 6 [p.5], Medical / healthcare 2 [p.1]
  • topic: Homomorphic encryption 174 [p.1], Privacy-preserving machine learning 147 [p.1], Private / secure inference 35 [p.13], Large language models 34 [p.1], Open source 28 [p.1], Benchmark 6 [p.28]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/sok-references/032_towards-efficient-privacy-preserving-machine-learning-a-syst.pdf sok-references 39 9.7 MB 99b928153e63

Database key: towards-efficient-privacy-preserving-machine-learning-a-systematic · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set towards-efficient-privacy-preserving-machine-learning-a-systematic COLUMN VALUE