TL;DR

The paper measures the performance overhead of running CKKS-encrypted SVM inference (both linear and polynomial kernels) on the Iris dataset using OpenFHE, sweeping multiplication depth, scaling factor, first modulus size, security level, batch size, and ring dimension [Abstract, §1]. Encrypted inference matches plaintext accuracy (96.7%) but is roughly 1,000x slower; ring dimension and modulus size dominate cost, while SVM-Linear and SVM-Poly perform similarly [Abstract, §6, §7].

Problem and motivation

Machine learning increasingly consumes PII, and while data can be encrypted at rest and in transit, it is typically processed in plaintext. The paper applies FHE (CKKS via OpenFHE) so an SVM can be evaluated on encrypted feature vectors, and characterises the resulting overhead across multiple cryptographic parameters [§1]. The threat model is not explicitly stated, but the workflow (client-side encryption with public key, server-side encrypted inference on a plaintext-trained model, client-side decryption with private key) implies a Private Prediction-as-a-Service setting with an honest-but-curious server [§3.1, §3.4, Algorithm 1].

Key contributions

FHE setup

ML setup

Datasets

Dataset Task Size (train/test) Modality Notes
Iris 3-class classification (setosa, versicolor, virginica) 120 / 30 (80%/20% split of 150 samples) Tabular (4 features: sepal length/width, petal length/width, cm) Standardised to zero mean, unit variance; UCI ML Repository; Appendix B notes the same script can also be used with a credit_approval dataset, but reported results are on Iris [§4, §4.2, Appendix A, Appendix B]

Pipeline diagram

flowchart LR
    subgraph Client_pre [Client]
        A[Standardise features] --> B[CKKS encrypt with pk]
    end
    subgraph Server
        C[Pack plaintext weights] --> D[EvalInnerProduct ct_x, pt_w]
        D --> E[Mask first slot]
        E --> F[EvalAdd bias]
    end
    subgraph Client_post [Client]
        G[Decrypt with sk] --> H[sign argmax label]
    end
    B -. ct_x .-> D
    F -. ct_res .-> G

Pipeline steps (text)

  1. Train SVM-Linear and SVM-Poly on plaintext Iris data with scikit-learn; export weights, intercept, dual coefficients, and support vectors [Appendix A].
  2. On the client, standardise the test feature vector and CKKS-encrypt it with the public key (ct_x = cc.Encrypt(keys.publicKey, pt_x)) [Appendix A].
  3. Send ct_x to the server.
  4. Server packs the plaintext weights and bias, then computes the encrypted inner product EvalInnerProduct(ct_x, pt_weights, n) [Appendix A, Eq. 25].
  5. Server multiplies by a one-hot mask to keep the first slot, then adds the bias via EvalAdd [Appendix A].
  6. For SVM-Poly, the decision function (w^T x + b)^d is evaluated under encryption [Eq. 28, Algorithm 1].
  7. Send the resulting ciphertext ct_res back to the client.
  8. Client decrypts with the secret key and applies sign(.) to produce the class label [Algorithm 1, Eq. 30].

Architecture diagram

flowchart LR
    I([Input x: 4 features]) --> E[CKKS Encrypt]
    E --> D1[Inner product wT x]
    D1 --> M[Mask slot 0]
    M --> B1[Add bias b]
    B1 --> P{Kernel}
    P -->|Linear| O1([Decrypt])
    P -->|Poly d=3| PW[Power d]
    PW --> O1
    O1 --> S([sign -> class])

Results

Metric This paper Baseline Hardware
Classification accuracy (encrypted SVM) 96.7% 96.7% (plaintext SVM) AWS EC2 t3.medium, 2 vCPUs Intel Xeon 3.1 GHz, 4 GB RAM [Table 1, Table 4]
Average encrypted inference time 0.2029 s 0.0002 s plaintext Same as above [Table 5, Appendix B]
Feature encryption time 0.2029 s n/a Same [Table 5]
Decryption time 0.0001 s n/a Same [Table 5]
Encrypted/plaintext slowdown ~1,000x - Same [§6]
Scale-up vs ring dimension (SVM-Linear, D=1, S=30, M=60, L=128, B=1024) 16K: 963.8x; 32K: 2,026.5x; 64K: 3,646.7x; 128K: 8,245.3x - Same [Table 6]
Scale-up vs ring dimension (SVM-Poly) 16K: 910.6x; 32K: 1,851.7x; 64K: 3,935.1x; 128K: 7,716.4x - Same [Table 6]
Scale-up vs multiplication depth, SVM-Linear (D=1..7) 1,032.8 / 1,172.7 / 1,397.2 / 1,794.5 / 2,056.4 / 2,097.5 / 2,324.5 - Same [Table 7]
Scale-up vs multiplication depth, SVM-Poly (D=1..7) 915.2 / 1,093.3 / 1,460.5 / 1,527.7 / 1,665.7 / 1,954.9 / 2,248.4 - Same [Table 7]
AET at S=10 (degraded accuracy) 0.817 (Linear) / 0.784 (Poly) 0.967 plaintext Same [Tables 2,3]

Note: the 0.2029 s figure is the average encrypted inference time per sample reported in Appendix B (Avg Encrypted Time: 0.2029 sec) and Table 5; it is used here as single_inference_seconds.

Limitations and assumptions

The paper positions itself relative to general FHE-ML literature rather than against specific encrypted-SVM baselines. Cited reference points include OpenFHE/PALISADE [§2.2], Microsoft SEAL [§2.2], the BGV/BFV [§2.6] and CKKS [§2.7] scheme papers, GWAS work by Blatt et al. via OpenFHE [§3.3], the PPaaS/PTaaS taxonomy by Iezzi et al. [§3.1], the Wood et al. FHE-for-medicine survey [§2.2, §3.1], and Bourse et al. on FHE-evaluated discretized NNs [§5]. No prior encrypted-SVM systems are directly benchmarked.

Code and artifacts

Not released as a public repository; the paper provides Python code snippets in Appendix A (using openfhe-python, scikit-learn, NumPy) and references scripts model_training.py, get_data.py, encrypted_svm_linear.py, encrypted_svm_poly.py without a hosting URL [§4.1, Appendices A and B]. License not stated.

Extra diagrams (optional)

Threat model

flowchart LR
    C[Client owns sk, pk] -- ct_x = Enc_pk(x) --> S[Server holds plaintext w, b]
    S -- ct_res = w*ct_x + b --> C
    C -- y = sign Dec_sk ct_res --> Y([Class label])

Open questions