TL;DR

The authors evaluate the performance overhead of CKKS-encrypted SVM inference (linear and polynomial kernels) using OpenFHE on the Iris dataset, finding that ring dimension and first-modulus size are the dominant cost factors while classification accuracy (96.7%) is preserved versus plaintext [§6, §7].

Problem and motivation

ML pipelines often expose personally identifiable information (PII) during in-process computation, even when data are encrypted at rest and in transit [§1]. The paper targets this gap by running SVM inference directly over CKKS-encrypted feature vectors, so the cloud-side classifier never sees plaintext. The threat model is implicit and consistent with standard FHE inference: the server is honest-but-curious (no collusion or malicious-client discussion is given in the text).

Key contributions

FHE setup

ML setup

Datasets

Dataset Task Size (train/test) Modality Notes
Iris (UCI) 3-class classification 120 train / 30 test (80/20 split) Tabular, 4 features (sepal length/width, petal length/width) Standardized to zero mean, unit variance; 150 samples, 3 balanced classes [§4.2.1, §4.2.2]

Pipeline diagram

flowchart LR
    subgraph Client
        A[Iris features] --> B[Standardize]
        B --> C[Encrypt with pk CKKS]
    end
    subgraph Server
        D[Enc dot product]
        D --> E[Add bias]
        E --> F{Polynomial kernel?}
        F -- yes --> G[Raise to power d]
        F -- no --> H[Skip]
    end
    subgraph Client_dec [Client]
        I[Decrypt with sk] --> J[Apply sign and label]
    end
    C -. ciphertext .-> D
    G -. ciphertext .-> I
    H -. ciphertext .-> I

Pipeline steps (text)

  1. Train SVM (linear or polynomial) on plaintext Iris data with scikit-learn; export weights w and bias b [§3.3, §4].
  2. Client standardizes the test feature vector x and CKKS-encrypts it with the public key pk (Eq. 15) [Algorithm 1, line 1].
  3. Server computes the encrypted decision function: Enc(w^T x + b) for linear, or Enc((w^T x + b)^d) for polynomial [Algorithm 1, lines 5-10].
  4. Server returns the resulting ciphertext c_f to the client.
  5. Client decrypts c_f with the private key sk to recover f(x) (Eq. 22) [Algorithm 1, line 11].
  6. Client applies sign(f(x)) and maps to the predicted class label (Eq. 23) [Algorithm 1, line 12].

Architecture diagram

flowchart LR
    I([Input x : 4 features]) --> S[Standardize]
    S --> E[Encrypt CKKS]
    E --> DP[Enc dot product w x]
    DP --> AB[Add bias b]
    AB --> K{Kernel}
    K -- linear --> O1([Enc f x])
    K -- poly d --> P[Raise to power d]
    P --> O2([Enc f x])
    O1 --> DEC[Decrypt and sign]
    O2 --> DEC
    DEC --> Y([Class label])

Results

Headline: encrypted accuracy matches plaintext exactly at 96.7% on Iris; encrypted single-sample inference is around 1000x slower than plaintext on the t3.medium instance, with ring dimension being the dominant scaling factor [§6.1.1, §6.1.2, §6.1.4].

Metric This paper Baseline Hardware
Encrypted accuracy (SVM, Iris) 96.7% 96.7% (plaintext SVM) AWS EC2 t3.medium, 2 vCPU Intel Xeon 3.1 GHz, 4 GB RAM [§4.1.3]
Feature encryption time 0.2029 s - t3.medium [Table 5]
Encrypted inference time 0.2029 s 0.0002 s (plaintext) t3.medium [Table 5]
Decryption time 0.0001 s - t3.medium [Table 5]
Encrypted inference scale-up (linear, MD=1..7, N=16384) 1032.8 -> 2324.5 (ratio vs plaintext) 1.0 t3.medium [Table 7]
Encrypted inference scale-up (poly, MD=1..7, N=16384) 915.2 -> 2248.4 1.0 t3.medium [Table 7]
Encrypted inference time vs ring dim (poly, MD=1) 915.2 s (N=16384) -> 7716.4 s (N=131072) - t3.medium; reported as “training time” in §7 but corresponds to the AET/scale-up rows in Tables 2-3 and 6 [§7, Table 6]
Accuracy with scaling factor S=10 81.7% (linear) / 78.4% (poly) 96.7% plaintext Drops only at S=10; recovers at S>=20 [Tables 2, 3]

Note: Tables 2 and 3 report Average Encryption Time (AET) per configuration in seconds, but it is not stated whether AET is per ciphertext, per test sample, or for the full 30-sample test set. The single-inference figure of 0.2029 s used in comparison.single_inference_seconds is taken from Table 5 (Runtime analysis) [§6.1.2, Table 5].

Limitations and assumptions

The paper does not benchmark numerically against other FHE-SVM systems. It positions itself relative to: Iezzi et al. (PPaaS/PTaaS taxonomy) [§3.1, ref 12]; Wood et al. (HE-for-ML survey) [§3.1, ref 9]; Blatt et al. (GWAS via OpenFHE, chi-squared and logistic regression) [§3.2, ref 13]; Cheon et al. (CKKS) [ref 6]; Brakerski-Vaikuntanathan (RLWE) [ref 5]; and Gentry (original FHE) [refs 3, 11].

Code and artifacts

Available at https://github.com/openfheorg/education/tree/main/examples/FHE_SVM_Examples (accessed 20 May 2025) [Data Availability Statement]. License: not stated in the text.

Extra diagrams (optional)

Threat model

flowchart LR
    C[Client owns x, sk, pk] -- Enc x with pk --> S[Server with plaintext w, b]
    S -- Enc f x --> C
    C -- Dec and sign --> Y[Label]

The paper does not formally state honest-but-curious, malicious, or collusion assumptions; the diagram reflects the implicit model where the server performs the homomorphic decision function and never sees plaintext features or the private key.

Open questions