Skip to content

Approximate homomorphic encryption based privacy-preserving machine learning: a survey

Authors: Jiangjun Yuan; Weinan Liu; Jiawen Shi; Qingqing Li
Year: 2025
Venue: Artificial Intelligence Review

type: survey collection: sok-references status: auto

Links: DOI · link

Summary

Auto-extracted summary (unreviewed)

A survey of privacy-preserving machine learning built on approximate homomorphic encryption (CKKS-style schemes). It compares HE with other privacy techniques, introduces representative approximate-HE schemes and parameter/security considerations, and systematically reviews approximate-HE PPML works across four technical applications and three advanced applications together with their scenarios, models and datasets, closing with future directions such as combining HE with SMPC or DP.

Abstract

Machine Learning (ML) is rapidly advancing, enabling various applications that improve people’s work and daily lives. However, this technical progress brings privacy concerns, leading to the emergence of Privacy-Preserving Machine Learning (PPML) as a popular research topic. In this work, we investigate the privacy protection topic in ML, and show­ case the advantages of Homomorphic Encryption (HE) among different privacy-preserving techniques. Additionally, this work presents an introduction of approximate HE, emphasiz­ ing its advantages and providing the detail of some representative schemes. Moreover, we systematically review the related works about approximate HE based PPML schemes from the four technical applications and three advanced applications, along with their applica­ tion scenarios, models and datasets. Finally, we suggest some potential future directions to guide readers in extending the research of PPML.
Source: pdf

Extracted details

Field Value Source Confidence
contribution comparison of privacy-preserving techniques highlighting the advantages of HE llm 0.80
contribution introduction to approximate HE and representative schemes llm 0.80
contribution systematic review of approximate-HE PPML across four technical and three advanced applications llm 0.80
contribution summary of commonly used datasets, models and application scenarios llm 0.80
contribution future research directions for approximate-HE PPML llm 0.80
schemes CKKS llm 0.80
libraries TenSEAL llm 0.80
libraries SEAL llm 0.80
tasks privacy-preserving ML training llm 0.80
tasks privacy-preserving ML inference llm 0.80
security level 128-bit (discussed as the target, e.g. N=8192 with bounded q) llm 0.80
code url https://github.com/Microsoft/SEAL (URL found in the full text) regex 0.50
llm relevance medium llm 0.80

SoK reference [6]

J. Yuan, W. Liu, J. Shi, and Q. Li, “Approximate homomorphic encryption based privacy-preserving machine learning: a survey,” Artificial Intelligence Review, vol. 58, no. 3, p. 82, 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: CKKS 79 [p.3], Differential privacy 15 [p.10], Multi-key HE 11 [p.8], Paillier 7 [p.5], BFV 5 [p.6], MPC 5 [p.4], BGV 4 [p.6], Scheme switching 3 [p.10], FHEW 2 [p.6], RSA 2 [p.5], Secret sharing 2 [p.11], Garbled circuits 2 [p.11], Oblivious transfer 2 [p.11]
  • library: Microsoft SEAL 9 [p.8], nGraph-HE 4 [p.25], TensorFlow 3 [p.14], CrypTen 2 [p.13], SecretFlow-SPU 2 [p.13], PyTorch 2 [p.14], OpenFHE 1 [p.8], Lattigo 1 [p.8], TenSEAL 1 [p.8]
  • model: CNN 46 [p.21], RNN / LSTM 28 [p.21], Logistic regression 21 [p.14], ResNet 17 [p.25], k-NN 12 [p.15], SVM 9 [p.14], BERT 7 [p.20], Decision tree 7 [p.14], k-means 7 [p.17], Transformer 5 [p.29], MobileNet 5 [p.26], LeNet 4 [p.25], VGG 4 [p.25], MLP 4 [p.21], AlexNet 3 [p.25], LLaMA 1 [p.49], GNN 1 [p.43], Linear regression 1 [p.46], Random forest 1 [p.49], Naive Bayes 1 [p.49], GAN 1 [p.40]
  • dataset: MNIST 14 [p.11], Iris 11 [p.14], ImageNet 5 [p.24], CIFAR-10 4 [p.27], IMDB 3 [p.22], Breast cancer (Wisconsin) 2 [p.17], Fingerprint (FVC) 2 [p.46]
  • hardware: GPU 10 [p.11], NVIDIA Jetson 2 [p.35], Raspberry Pi 1 [p.34], ASIC 1 [p.26]
  • technique: Federated learning 14 [p.4], Rotations 13 [p.7], RNS 12 [p.8], Convolution 12 [p.25], Bootstrapping 10 [p.6], SIMD packing 10 [p.4], Softmax 9 [p.17], ReLU 9 [p.11], Non-interactive 7 [p.12], Rescaling 5 [p.7], Goldschmidt 4 [p.17], Sigmoid 4 [p.21], Polynomial approximation 3 [p.21], Matrix multiplication 3 [p.27], Relinearization 2 [p.8], Newton iteration 2 [p.16], GELU 2 [p.34], Lookup table 2 [p.10], Encrypted training 2 [p.26], Compiler 2 [p.25]
  • assumption: LWE 8 [p.6], RLWE 6 [p.6], Lattice-based 5 [p.6], IND-CPA 2 [p.8], 128-bit security 2 [p.8]
  • task: Medical / healthcare 12 [p.2], Text classification 11 [p.19], Sentiment analysis 9 [p.19], Genomics 5 [p.17], Language modeling / generation 4 [p.7], Finance 4 [p.14], IoT / edge 4 [p.35], Image classification 3 [p.24], Clustering 3 [p.17], Machine translation 2 [p.22], Speech recognition 2 [p.31]
  • topic: Homomorphic encryption 244 [p.1], Privacy-preserving machine learning 78 [p.1], Private / secure inference 31 [p.11], Large language models 19 [p.5], Open source 4 [p.8], Benchmark 3 [p.23]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/sok-references/006_approximate-homomorphic-encryption-based-privacy-preserving.pdf sok-references 49 3.9 MB e5a9d19f4ec6

Database key: approximate-homomorphic-encryption-based-privacy-preserving-machine · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set approximate-homomorphic-encryption-based-privacy-preserving-machine COLUMN VALUE