TL;DR

A 2020 survey reviewing Fully Homomorphic Encryption (FHE) theory, scheme families, and tooling, with explicit focus on the intersection of FHE and Machine Learning as a Service (MLaaS); it argues that no specialized survey existed at this intersection and consolidates contributions for predicting/classifying confidential data under encryption [Abstract][§4.1].

Problem and motivation

Cloud computing introduces data-breach and trust threats: users delegate processing to a Cloud Service Provider (CSP) and must decrypt to compute, exposing data [§1]. Conventional encryption “cannot support effective ciphertext computing” [§1]. The threat model implicit throughout is a non-trustworthy third-party (honest-but-curious CSP) that should “blindly process encrypted information without disclosing confidential data” and never see the secret key [§1, §3]. FHE is positioned as the “holy grail” enabling computation on encrypted data and as a candidate against future quantum-computer attacks on factoring/discrete-log public-key crypto [§1].

Key contributions

FHE setup

ML setup

Datasets

Dataset Task Size (train/test) Modality Notes
Not applicable Survey Not reported N/A Survey paper; no experiments of its own. Application domains listed include medical, advertising, financial, genomics, smart cities, spam filtering, image processing [§5].

Pipeline diagram

The survey describes the canonical FHE-MLaaS message flow [§3, §4.2]: given a scheme ε, model α, and encrypted input Encrypt_ε(p), the server returns ciphertext c such that Decrypt_ε(c) = α(p).

flowchart LR
    subgraph Client
        A[Plaintext input p] --> B[Encrypt with pk]
    end
    subgraph Server [Non-trustworthy CSP]
        C[Evaluate model alpha] --> D[Bootstrap / key-switch]
    end
    subgraph ClientDec [Client]
        E[Decrypt with sk] --> F[Result alpha of p]
    end
    B -. ciphertext .-> C
    D -. ciphertext .-> E

Pipeline steps (text)

  1. Client runs KeyGen_ε(λ) to obtain (sk, pk) [§3.1].
  2. Client encrypts plaintext p with pk: c = Encrypt_ε(pk, p) [§3.1].
  3. Ciphertext c is sent to the non-trustworthy CSP [§1, §4.2].
  4. Server runs Evaluate_ε(pk, α, c) to homomorphically evaluate the ML model α [§3.1, §4.2].
  5. Server applies bootstrapping/key-switching as needed to manage noise growth [§3.2, §3.3].
  6. Server returns the resulting ciphertext to the client [§4.2].
  7. Client decrypts with sk to obtain α(p) [§3.1, §4.2].

Architecture diagram

Survey taxonomy of FHE families and the MLaaS approaches built on them [§2, §4.2].

flowchart LR
    R([FHE]) --> F1[Ideal-lattice family]
    R --> F2[Integer-based family]
    R --> F3[LWE / RLWE family]
    R --> F4[NTRU family]
    F1 --> P1[Gentry 2009]
    F2 --> P2[van Dijk et al.]
    F3 --> P3[BGV / GSW]
    F3 --> P4[CKKS / HEAAN]
    F4 --> P5[Rohloff-Cousins NTRU]
    R --> ML[FHE for MLaaS]
    ML --> M1[Logistic regression]
    ML --> M2[Bayesian / Decision trees]
    ML --> M3[CryptoNets NN]
    ML --> M4[Deep NN BGV]
    ML --> M5[CNN on GPU HCNN]
    ML --> M6[Poly-activation distributed NN]
    ML --> M7[Additively-HE + ASGD]
    ML --> M8[SecureNN 3-party]

Results

The paper presents no numerical experiments. The survey’s “result” is Table 1, a comparative matrix of 24 referenced works across operations supported (addition, multiplication, other), ML approach (logistic regression, NN, DNN, decision trees, DFT), scheme family (ideal-lattice, integer-based, (R)LWE, NTRU), party model (two-party / multi-party), and objective (security / efficiency) [§4.2, Table 1].

Metric This paper Baseline Hardware
Empirical accuracy / latency Not reported (survey) Not reported Not reported
Conceptual coverage 24 prior works tabulated across 4 FHE families and 5+ ML approaches [Table 1] n/a n/a

The authors note that bootstrapping is “the major bottleneck” and that the field’s objectives of security and efficiency are “in conflict” [§4.2].

Limitations and assumptions

CryptoNets [33], Chabanne et al. (batch-norm NN) [34], Badawi et al. (HCNN/AlexNet on GPU) [35], Zhang et al. (BGV deep computation) [36], Takabi et al. (decentralized poly-activation) [37], Phong et al. (additively-HE + ASGD) [38], Wagh et al. (SecureNN 3-party) [39], Naehrig et al. (logistic regression) [19], Khedr et al. (SHIELD: Bayesian / decision trees) [32], Bost et al. (ML classification over encrypted data) [31], Babenko et al. (RNS comparison) [40]. Compared against prior surveys: Armknecht et al. [18], Acar et al. [21], Martins et al. [22], and domain-specific surveys [27, 28, 29] [§4.1].

Code and artifacts

Not released (survey paper). Tools surveyed for practitioners: SEAL, HElib, TFHE, PALISADE, cuHE, HEAAN, HE-transformer/nGraph-HE [§5].

Extra diagrams (optional)

Threat model

flowchart LR
    C[Client - holds sk] -- Encrypt_pk of x --> S[Non-trustworthy CSP - holds pk, model alpha]
    S -- Encrypt_pk of alpha of x --> C
    note[CSP never sees sk or plaintext]

The third-party “does not learn anything about the data, or output of the computation” [§4.2].

Federated round

Not applicable in detail; the survey references multi-party schemes (Takabi et al. [37], Wagh et al. SecureNN [39], Phong et al. [38]) but does not specify a single federated protocol [§4.2].

Activation approximation

Not reported in detail. The survey notes that polynomial approximations replace sigmoid/non-polynomial activations (e.g., CryptoNets, Takabi et al.) but does not provide explicit polynomial coefficients or error plots [§4.2].

Open questions