TL;DR

Naresh and Reddi build a CKKS-based logistic regression framework (HELR) that trains and infers on encrypted heart-disease records using TenSEAL/PyTorch on Google Colab, achieving 84% accuracy versus 85% for the plaintext baseline while preserving patient privacy across patient / hospital / cloud-service-provider roles [Abstract][§Experimentation and results].

Problem and motivation

The paper targets privacy-preserving heart disease prediction in cloud settings where hospitals want to outsource training/inference of logistic regression to a Cloud Service Provider (CSP) without exposing sensitive patient records [§Introduction][§Proposed system]. Traditional privacy techniques (anonymization, perturbation, differential privacy) either degrade utility or inject accuracy-harming noise; the authors motivate full HE as a way to avoid this trade-off [§Introduction]. Threat model: the CSP is treated as “authentic and curious” (honest-but-curious) and must perform model training and prediction over encrypted data; patients hold their own private keys and the hospital holds keys for the training pipeline [§Proposed system / System architecture]. The security analysis additionally considers active adversaries mounting poisoning, evasion, membership inference, model inversion, and model extraction attacks [§Security analysis].

Key contributions

FHE setup

ML setup

Datasets

Dataset Task Size (train/test) Modality Notes
Kaggle Heart Disease (johnsmith88) Binary heart-disease classification Not reported Tabular EHR features Primary dataset; used for the headline 84%/85% accuracy comparison [§Experimentation and results, ref 27]
Statlog Heart Binary heart-disease classification Not reported Tabular Used in HELR vs HESVM comparison [§Experimentation and results, ref 36]
Framingham Heart Study Binary CHD risk classification Not reported Tabular EHR Used in HELR vs HESVM comparison [§Experimentation and results, ref 35]

Pipeline diagram

flowchart LR
    subgraph Patient
        P1[Health record z] --> P2[Encrypt with hospital pk]
    end
    subgraph Hospital
        H1[Training set + params] --> H2[Encrypt with hospital pk]
        H3[Decrypt prediction] --> H4[Re-encrypt with patient pk]
    end
    subgraph CSP
        C1[Train HELR on enc data]
        C2[Encrypted inference]
    end
    subgraph Patient_dec [Patient]
        D1[Decrypt with patient sk] --> D2[Disease label]
    end
    H2 -. enc training .-> C1
    C1 -. enc weights .-> H3
    P2 -. enc test .-> C2
    C2 -. enc prediction .-> H3
    H4 -. enc result .-> D1

Pipeline steps (text)

  1. Patient and hospital each generate CKKS key pairs via Keygen() [§Proposed model, Alg. 1].
  2. Hospital encrypts the training dataset and initial model parameters with its public key and ships them to the CSP [§Proposed system].
  3. CSP runs gradient-descent logistic regression on the encrypted data, updating an encrypted weight vector each epoch [§Logistic regression over encrypted data, Alg. 3].
  4. CSP returns the encrypted updated weights to the hospital [§Proposed system].
  5. Patient encrypts their record (test sample) with the hospital public key and sends it to the hospital, which forwards it to the CSP [§Proposed model].
  6. CSP computes the encrypted prediction σ(Enc(z)·Enc(w)) and returns it [§Security analysis, Eq. 24].
  7. Hospital decrypts the prediction, re-encrypts it under the patient’s public key, and sends it to the patient [§Proposed model].
  8. Patient decrypts to learn the disease status [§Proposed model].

Architecture diagram

flowchart LR
    I([Input z, D features]) --> L1[Linear w^T z + b]
    L1 --> S[sigmoid]
    S --> O([Threshold 0.5 -> label])

The model is a single-layer logistic regression with D input features, one sigmoid output unit, and a 0.5 (or tuned 0.8 for ROC) decision threshold [§Logistic regression][§Experimentation and results].

Results

Headline numbers, taken directly from the paper [§Experimentation and results, Figs. 4-8]:

Metric This paper (HELR) Baseline Hardware
Accuracy on Heart (Kaggle) 84% (HELR_4096), 84% (HELR_8192) LR plaintext: 85% Google Colab (CPU)
Accuracy delta vs plaintext “between 0.01 and 0.03” n/a Google Colab
Training-set encryption time, HELR_4096 1061 / 1093 / 1301 / 1731 ms at 5 / 10 / 20 / 50 epochs n/a Google Colab
Training-set encryption time, HELR_8192 4550 / 5242 / 5542 / 12225 ms at 5 / 10 / 20 / 50 epochs n/a Google Colab
Test-set encryption time, HELR_4096 219 / 220 / 222 / 229 ms n/a Google Colab
Test-set encryption time, HELR_8192 1013 / 1019 / 1033 / 1059 ms n/a Google Colab
HELR vs HESVM on Heart HELR 81.7 (4096), 85 (8192) HESVM 67 (4096), 79 (8192) Google Colab
HELR vs HESVM on Statlog HELR 100 (both configs) HESVM 69 (4096), 88 (8192) Google Colab
HELR vs HESVM on Framingham HELR 62 (4096), 65 (8192) HESVM 63 (4096), 61 (8192) Google Colab
Best ROC threshold 0.8 (HELR_4096 has higher AUC than HELR_8192) n/a Google Colab

Single-encrypted-sample inference latency is not separately reported, so single_inference_seconds is set to N/A in the comparison block; the test-set encryption times above are dataset-level, not per-sample [§Experimentation and results].

Limitations and assumptions

Code and artifacts

Not released. The paper states “No datasets were generated or analysed during the current study” and does not provide a repository URL; implementation is described as TenSEAL + PyTorch on Google Colab [§Availability of data and materials][§Experimentation and results].

Extra diagrams (optional)

Threat model

flowchart LR
    P[Patient] -- enc(test record, hospital pk) --> H[Hospital]
    H -- enc(train set + params, hospital pk) --> S[CSP - honest but curious]
    S -- enc(weights) --> H
    H -- enc(test record) --> S
    S -- enc(prediction) --> H
    H -- enc(prediction, patient pk) --> P

Three parties (patient, hospital, CSP); the CSP performs training and inference on ciphertexts and never sees plaintext data, weights, or predictions [§Proposed system / System architecture].

Open questions