TL;DR

The paper presents a CKKS-based framework for privacy-preserving SVM inference on encrypted medical records and medical images, supporting linear, polynomial, RBF, and Sigmoid kernels at a 128-bit security level with single-sample latencies between 0.39 s and 12.80 s [Abstract, Conclusions p. 22].

Problem and motivation

Bioinformatics analyses on sensitive medical data (cardiology, oncology, medical imaging) need confidentiality guarantees when computation is outsourced to untrusted servers [Introduction p. 1-2]. The authors target an honest-but-curious server that follows the protocol but may try to learn information from encrypted data or the secret key; the server holds the SVM model in plaintext, and the client encrypts queries [§Threat model p. 8]. Client data privacy, insight privacy, insight accuracy, and unlinkability are the stated guarantees [§Threat model p. 8].

Key contributions

FHE setup

ML setup

Datasets

Dataset Task Size (train/test) Modality Notes
Cleveland Heart Disease (CHD) Binary classification (disease presence) 303 samples total, 80%/20% split Tabular, 13 numerical features Imbalanced; missing values present; min-max normalized [Table 1 p. 13, §Data splitting p. 15]
Wisconsin Breast Cancer (WBC) Binary classification (malignant/benign) 569 samples total, 80%/20% split Tabular, 9 numerical features (paper says 30 features for source but Table 1 reports 9) Imbalanced; no missing values [Table 1 p. 13]; Table 8 also lists 9 features for WBC
BreastMNIST Binary classification (normal/malignant) 780 breast ultrasound images, 70%/10%/20% train/val/test Grayscale images resized to 28x28 From MedMNIST; 128-D features extracted via Autoencoder [Table 2 p. 13, §Datasets p. 12-13]
PneumoniaMNIST Binary classification (normal/pneumonia) 5,856 chest X-rays, 90%/10% train + val/test Grayscale images center-cropped and resized to 28x28 From MedMNIST; 128-D features extracted via Autoencoder [Table 2 p. 13]

Pipeline diagram

flowchart LR
    subgraph Client
        A[Preprocess sample] --> AE[Autoencoder features<br/>images only]
        AE --> B[Encrypt with CKKS pk]
    end
    subgraph Server
        C[Homomorphic SVM<br/>kernel + dot product]
        C --> D[Chebyshev approx<br/>RBF/Sigmoid]
        D --> E[Sum alpha*y*K + b]
    end
    subgraph Client_dec [Client]
        F[Decrypt with sk] --> G[Apply sign<br/>output label]
    end
    B -. ciphertext .-> C
    E -. ciphertext .-> F

Pipeline steps (text)

  1. Client preprocesses the input (min-max normalization for tabular; for medical images, run the pretrained Autoencoder encoder to obtain a 128-D feature vector) [§Data preprocessing p. 14, §Feature extraction p. 14].
  2. Client encodes the resulting feature vector into a CKKS plaintext and encrypts under the public key [§FHSVM step 1 p. 9].
  3. Client transmits the ciphertext to the server [§FHSVM step 2 p. 9].
  4. Server evaluates the SVM decision function homomorphically: dot products with support vectors, kernel application (linear/polynomial direct; RBF/Sigmoid via Chebyshev polynomial approximation), then a weighted sum with alpha_i * y_i plus bias b [§Homomorphic prediction p. 15-16, Eq. (1) p. 6].
  5. Server returns the encrypted score to the client (the sign function is not evaluated server-side) [§Homomorphic prediction p. 16].
  6. Client decrypts and applies the sign function locally to obtain the final binary label [§Homomorphic prediction p. 16, §FHSVM step 5 p. 10].

Architecture diagram

The deployed model under FHE is an SVM (not a neural network). The diagram below shows the homomorphic SVM evaluation graph; for image inputs an offline CNN Autoencoder produces the 128-D vector that feeds it.

flowchart LR
    I([Encrypted input<br/>13 / 9 / 128 dims]) --> DOT["Dot products<br/>x . SV_i"]
    DOT --> K{Kernel}
    K -->|Linear| KL[x.SV_i]
    K -->|Poly| KP["poly x.SV_i + c"]
    K -->|RBF| KR["Chebyshev approx<br/>exp -gamma ||x-SV||^2"]
    K -->|Sigmoid| KS["Chebyshev approx<br/>tanh alpha x.SV + beta"]
    KL --> W["Weighted sum<br/>sum alpha_i y_i K"]
    KP --> W
    KR --> W
    KS --> W
    W --> B["Add bias b"]
    B --> O([Encrypted score<br/>sign applied client-side])

Autoencoder feature extractor (images only, plaintext, offline)

flowchart LR
    I([Input 1x28x28]) --> C1[Conv2D 16<br/>26x26x16]
    C1 --> P1[MaxPool<br/>13x13x16]
    P1 --> C2[Conv2D 8<br/>13x13x8]
    C2 --> P2[MaxPool<br/>7x7x8]
    P2 --> C3[Conv2D 8<br/>4x4x8]
    C3 --> F[Flatten 128]
    F --> O([128-D feature vector<br/>to SVM])

Results

Headline numbers from the paper [Table 6 p. 18, Table 7 p. 19, Table 8 p. 20]:

Metric This paper Baseline Hardware
WBC F1 (Sigmoid kernel, encrypted) 0.96 Comparable to plaintext SVM Intel Core i7-12700H, 64 GB RAM, multi-threaded OpenFHE
WBC F1 (RBF, encrypted) 0.96 Comparable to plaintext SVM same
CHD F1 (Sigmoid, encrypted) 0.75 Comparable to plaintext SVM same
BreastMNIST F1 (RBF) 0.80 MedMNIST auto-sklearn: 0.803 accuracy same
PneumoniaMNIST F1 (RBF) 0.85 MedMNIST auto-sklearn: 0.855 accuracy same
CHD latency (Linear) 0.39 s - same
CHD latency (Poly) 3.83 s - same
CHD latency (RBF) 4.81 s - same
CHD latency (Sigmoid) 11.23 s - same
WBC latency (Linear) 0.52 s - same
WBC latency (RBF) 5.73 s Sari et al. (Paillier, Linear): 47.00 s same
WBC latency (Sigmoid) 12.80 s - same
BreastMNIST latency (RBF) 2.09 s - same
PneumoniaMNIST latency (RBF) 2.03 s - same
WBC RBF vs Bajard et al. [32] 5.73 s (CPU) 21.42 s on Statlog Heart with RBF, GPU Authors note their CPU result outperforms [32]’s GPU result [§Performance comparison p. 20]

Approximate plaintext-vs-encrypted gap: the paper states “no notable precision loss is observed due to the introduction of homomorphic encryption” [§Introduction p. 3], and “comparable accuracy to the unencrypted SVM prediction” [§Main findings p. 21]. Exact deltas per kernel are not tabulated.

Limitations and assumptions

Code and artifacts

Open-source at https://github.com/caesaretos/svm-fhe [§Contributions p. 3, §Supporting Information p. 23]. License terms: “Clear terms of use and contribution guidelines” stated in repository per the paper, specific license name Not reported in the manuscript text.

Extra diagrams (optional)

Threat model

flowchart LR
    C[Client<br/>holds sk, query x] -- "Enc(x)" --> S[Server<br/>plaintext SVM model]
    S -- "Enc(score)" --> C
    C --> SIGN[Apply sign locally]

The server is honest-but-curious and never sees plaintext queries; SVM parameters are in the clear on the server [§Threat model p. 8].

Activation approximation

Chebyshev polynomial approximations of the kernel transcendentals [Table 4 p. 15]:

See Table 4 in paper for the exact ranges and degrees.

Open questions