Skip to content

A Survey of Deep Learning Architectures for Privacy-Preserving Machine Learning With Fully Homomorphic Encryption

Authors: Robert Podschwadt; Daniel Takabi; Peizhao Hu; Mohammad H. Rafiei; Zhipeng Cai
Year: 2022
Venue: IEEE Access

type: survey collection: ppml status: auto

Links: DOI

Summary

Auto-extracted summary (unreviewed)

Podschwadt et al. survey the literature on deep learning architectures for privacy-preserving machine learning with fully homomorphic encryption in the MLaaS setting. They analyse how neural network models and architectures are modified to be HE-compatible (e.g., handling multiplicative depth and the restriction to additions and multiplications) and how these changes affect performance, catalogue challenges such as computational overhead, usability and scheme limitations, discuss candidate solutions, and propose evaluation metrics for more meaningful comparison of PPML solutions.

Abstract

Outsourced computation for neural networks allows users access to state-of-the-art models without investing in specialized hardware and know-how. The problem is that the users lose control over potentially privacy-sensitive data. With homomorphic encryption (HE), a third party can perform computation on encrypted data without revealing its content. In this paper, we reviewed scientific articles and publications in the particular area of Deep Learning Architectures for Privacy-Preserving Machine Learning (PPML) with Fully HE. We analyzed the changes to neural network models and architectures to make them compatible with HE and how these changes impact performance. Next, we find numerous challenges to HEbased privacy-preserving deep learning, such as computational overhead, usability, and limitations posed by the encryption schemes. Furthermore, we discuss potential solutions to the HE PPML challenges. Finally, we propose evaluation metrics that allow for a better and more meaningful comparison of PPML solutions.
Source: pdf

Extracted details

Field Value Source Confidence
contribution review of HE-based privacy-preserving deep learning literature llm 0.75
contribution analysis of architectural changes needed for HE compatibility and their performance impact llm 0.75
contribution catalogue of challenges: computational overhead, usability, scheme limitations llm 0.75
contribution discussion of potential solutions to HE PPML challenges llm 0.75
contribution proposed evaluation metrics for comparing PPML solutions llm 0.75
schemes homomorphic encryption llm 0.75
libraries SEAL llm 0.75
libraries HEAAN llm 0.75
tasks private neural network inference llm 0.75
tasks Machine Learning as a Service llm 0.75
security level 128-bit or more in many surveyed works; some choose lower levels llm 0.75
threat model client-server MLaaS; server computes on encrypted client data llm 0.75
interaction non-interactive FHE llm 0.75
code url https://github.com/Microsoft/SEAL (URL found in the full text) regex 0.50
code url https://github.com/snucrypto/HEAAN (URL found in the full text) regex 0.50
limitations source code of surveyed works is often unavailable, hindering reproduction and comparison llm 0.75
llm relevance medium llm 0.75

Keyword mentions

Counts of vocabulary terms in the full text (first page of occurrence in brackets).

  • scheme: CKKS 25 [p.4], Leveled / somewhat HE 25 [p.4], TFHE 17 [p.4], BFV 9 [p.4], MPC 8 [p.1], BGV 6 [p.4], YASHE 6 [p.4], Functional encryption 5 [p.1], Differential privacy 3 [p.3], Zero-knowledge proofs 3 [p.2], NTRU 2 [p.14], Paillier 2 [p.22], Oblivious transfer 2 [p.22], TEE / confidential computing 2 [p.3]
  • library: EVA / CHET 9 [p.7], nGraph-HE 6 [p.7], Microsoft SEAL 4 [p.5], HElib 2 [p.5], PALISADE 1 [p.23], PyTorch 1 [p.7], TensorFlow 1 [p.24]
  • model: RNN / LSTM 36 [p.6], CNN 8 [p.6], ResNet 2 [p.13], AlexNet 1 [p.23], MLP 1 [p.8], Logistic regression 1 [p.23]
  • dataset: MNIST 25 [p.6], CIFAR-10 15 [p.9], ImageNet 3 [p.17], IMDB 1 [p.16]
  • hardware: GPU 20 [p.6], FPGA 1 [p.21], ARM 1 [p.3]
  • technique: Bootstrapping 57 [p.4], SIMD packing 35 [p.4], Convolution 22 [p.6], ReLU 18 [p.5], Quantization 15 [p.6], Compiler 10 [p.7], Matrix multiplication 8 [p.6], Softmax 7 [p.5], Polynomial approximation 6 [p.6], Rotations 6 [p.7], Encrypted training 5 [p.14], Minimax approximation 4 [p.12], RNS 4 [p.6], Pruning 4 [p.6], Sigmoid 3 [p.5], Chebyshev 2 [p.11], Goldschmidt 2 [p.12], Lookup table 2 [p.12], Non-interactive 2 [p.19]
  • assumption: Lattice-based 7 [p.14]
  • task: Speech recognition 3 [p.6]
  • topic: Homomorphic encryption 175 [p.1], Privacy-preserving machine learning 75 [p.1], Open source 3 [p.23]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/ppml/surveys/podschwadt-2022_survey-of-deep-learning-architectures-for-ppml-with-fhe.pdf ppml/surveys 24 3.0 MB fff137cd7828 was Podschwadt et al. - 2022 - A Survey of Deep Learning Architectures for Privacy-Preserving Machine Learning With Fully Homomorph.pdf

Database key: a-survey-of-deep-learning-architectures-for-privacy-preserving-machine · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set a-survey-of-deep-learning-architectures-for-privacy-preserving-machine COLUMN VALUE