TL;DR

A short survey by Cheng Hong (Ant Group) that organizes recent FHE-based privacy-preserving machine learning (PPML) work by the four “generations” of FHE schemes plus mixed-mode and hardware-acceleration efforts, aimed at helping practitioners pick a method for a given model/data combination [§1, p. 1-2].

Problem and motivation

FHE allows computation on encrypted data without decryption and is a candidate technology to protect data confidentiality during ML training or inference [§1, p. 1]. But “there are many kinds of available FHE schemes and way more FHE-based solutions in the literature, and they are still fast evolving, making it difficult to get a complete view” [Abstract, p. 1]. The article tries to answer: given a model X and data Y, which FHE-based PPML method should I pick and what efficiency should I expect? [§1, p. 1-2]. Threat model is not explicitly formalized; the survey notes that FHE-based PPML is “a fully noninteractive procedure run on the server side”, contrasting it with interactive MPC approaches [footnote 1, p. 2].

Key contributions

FHE setup

ML setup

Datasets

Dataset Task Size (train/test) Modality Notes
MNIST Image classification Not reported (survey) Images Used by CryptoNets, FHE-DiNN, TAPAS, Chillotti et al. [§2, §3, p. 2-4]
CIFAR-10 Image classification Not reported Images Used by Lola, Lee et al. ResNet-110, Park et al. GPU ResNet-20 [§2, §4.2, §6, p. 3-5]
iDASH2017 gene dataset Logistic regression training 1579 samples × 18 features Tabular / genomics Kim et al. trained in 6 minutes via CKKS [§4.1, p. 4]
iDASH2018 SNPs p-value / multi-model training Not reported Tabular / genomics Won by CKKS-based teams [§4.1, p. 4]
Han et al. logistic regression set LR training 422,108 samples × 200 features Tabular 200 iterations in 1060 minutes [§4.2, p. 4]
FAB FPGA LR set LR training 11,982 samples × 196 features Tabular Trained in 0.1 s on FPGA [§6, p. 5]
BERT-base input Transformer inference 128 tokens NLP NEXUS, 1103 s end-to-end [§4.2, p. 4]

Pipeline diagram

Generic FHE-PPML server-side ciphertext flow the survey contrasts against interactive MPC [§1 footnote, p. 2].

flowchart LR
    subgraph Client
        A[Plaintext input] --> B[Encrypt]
    end
    subgraph Server
        C[Homomorphic ML eval]
        C --> D[Bootstrap if needed]
        D --> C
    end
    subgraph Client_dec [Client]
        E[Decrypt] --> F[Read prediction]
    end
    B -. ciphertext .-> C
    D -. ciphertext .-> E

Pipeline steps (text)

  1. Client encrypts input data with public key (pixel-wise, bitwise, or SIMD-packed depending on scheme) [§2, p. 2].
  2. Server runs homomorphic inference or training on the chosen ML model, with no decryption [§1, p. 1].
  3. Server optionally bootstraps to refresh ciphertext noise (3rd-gen fast bootstrap / PBS, or 4th-gen CKKS bootstrap) [§3, §4.2, p. 3-4].
  4. Server returns the encrypted result.
  5. Client decrypts with the secret key.

Architecture diagram

Taxonomy of the survey: FHE generations and the representative PPML works mapped to each, plus mixed-mode and hardware-acceleration branches [§1-§6, p. 2-5].

flowchart LR
    R([FHE-based PPML])
    R --> G1[Gen 1: Gentry, DGHV]
    G1 --> G1n[Too slow for PPML]

    R --> G2[Gen 2: BFV / BGV / YASHE]
    G2 --> G2a[CryptoNets]
    G2 --> G2b[Lola]
    G2 --> G2c[Nandakumar et al. NN training]

    R --> G3[Gen 3: DM / CGGI / TFHE]
    G3 --> G3a[FHE-DiNN]
    G3 --> G3b[TAPAS BNN]
    G3 --> G3c[SHE boolean ReLU]
    G3 --> G3d[Chillotti et al. PBS NN]
    G3 --> G3e[Frery et al. trees]

    R --> G4[Gen 4: CKKS approximate]
    G4 --> G4a[iDASH Kim et al. LR]
    G4 --> G4b[Han et al. LR at scale]
    G4 --> G4c[Lee et al. ResNet-110]
    G4 --> G4d[NEXUS BERT]

    R --> M[Mixed-mode]
    M --> Ma[Chimera]
    M --> Mb[Pegasus]
    M --> Mc[HE3DB]

    R --> H[Hardware acceleration]
    H --> Ha[GPU: Park et al.]
    H --> Hb[FPGA: FAB]
    H --> Hc[ASIC simulators]

Results

Headline numbers reported in the survey for the works it cites. Hardware is given where the survey states it; otherwise “Not reported”.

Metric This paper Baseline Hardware
CryptoNets MNIST throughput 4096 images in 200 s, 99% accuracy [§2, p. 3] — Not reported
Lola speedup vs CryptoNets >90× faster, supports CIFAR-10 [§2, p. 3] CryptoNets Not reported
Nandakumar et al. NN training 1.5 days for 3-layer model on minibatch of 60 images [§2, p. 3] — Not reported
FHE-DiNN MNIST inference ~1 image / second [§3.1, p. 3] — Not reported
TAPAS BNN inference Hours per inference [§3.1, p. 3] FHE-DiNN Not reported
SHE latency 3-5× higher than Lola, better accuracy (lossless ReLU) [§3.1, p. 3] Lola Not reported
Chillotti et al. 100-layer MNIST “Tens of seconds” [§3.2, p. 3-4] — AWS cloud server
Zama Concrete-ML rounding optimization [25] >10× faster than [23]; ~2% accuracy drop on MNIST [§3.2, p. 4] Chillotti et al. [23] Not reported
Frery et al. tree (depth 5) Within 5 s; sub-second possible; accuracy matches plaintext [§3.2, p. 4] — Not reported
Kim et al. CKKS LR (iDASH2017) 6 min on 1579×18 [§4.1, p. 4] BFV/BGV: tens-to-hundreds of min Not reported
Han et al. CKKS LR at scale 1060 min for 200 iterations on 422,108×200 [§4.2, p. 4] — Not reported
Lee et al. ResNet-110 (CKKS) ~13,000 s per CIFAR-10 image; 75% in bootstrap; accuracy ~ plaintext [§4.2, p. 4] — Not reported
NEXUS BERT-base inference 1103 s for 128 tokens [§4.2, p. 4] — Not reported
Pegasus arbitrary functions 1-2 s; PBS accuracy <8-bit limit [§5, p. 5] Chimera Not reported
Park et al. GPU ResNet-20 (CIFAR-10) 8.5 s; 267× over CPU [31]; 1.4 s with low-deg poly ReLU NAS [§6, p. 5] CPU CKKS [31] NVIDIA A100
FAB FPGA LR 0.1 s training on 11,982×196; 370× over CPU [§6, p. 5] CPU Xilinx Vivado FPGA @ 300 MHz

Limitations and assumptions

The article positions itself against MPC-based PPML surveys (Ng & Chow, SoK [15]) and against the FHE accelerator survey by Zhang et al. [43]. Within FHE, it discusses CryptoNets, Lola, FHE-DiNN, TAPAS, SHE, Chillotti et al., Frery et al., Kim et al., Han et al., Lee et al., NEXUS, Chimera, Pegasus, HE3DB, Park et al., FAB, HEAX, FPT, CHAM, TensorFHE [§2-§6, p. 2-5].

Code and artifacts

Not released (this is a survey article). Zama’s Concrete-ML announcement is cited as a public artifact for [25] (https://www.zama.ai/post/making-fhe-faster-for-ml-beating-our-previous-paper-benchmarks-with-concrete-ml) [§3.2, p. 4]. iDASH competition site cited: http://www.humangenomeprivacy.org [§4.1, p. 4].

Extra diagrams (optional)

Threat model

flowchart LR
    C[Client] -- enc(x) --> S[Server runs FHE ML]
    S -- enc(y) --> C

The survey contrasts this non-interactive FHE flow against MPC-style interactive protocols [§1 footnote 1, p. 2].

Open questions