Exploring the future of privacy-preserving heart disease prediction: a fully homomorphic encryption-driven logistic regression approach¶
Authors: Vankamamidi S. Naresh; Sivaranjani Reddi
Year: 2025
Venue: Journal of Big Data
type: application collection: ppml status: auto
Links: DOI
Summary¶
Auto-extracted summary (unreviewed)
The authors build HELR, a logistic-regression model trained on CKKS-encrypted data with TenSEAL and PyTorch, so that heart-disease prediction can be outsourced to a cloud without revealing patient records. On three Kaggle healthcare datasets HELR reaches accuracy within 1-3% of plaintext logistic regression (84% vs 85%), outperforms an HE-based SVM, and a polynomial degree of 4096 is cheaper than 8192 at some cost in security; a security analysis argues robustness against poisoning, evasion, membership-inference, model-inversion and model-extraction attacks.
Abstract¶
Homomorphic Encryption (HE) offers a revolutionary cryptographic approach to safeguarding privacy in machine learning (ML), especially in processing sensitive healthcare data. This study aims to address the critical issue of privacy-preserving heart disease prediction by developing a novel Homomorphic Encryption-Driven Logistic Regression (HELR) framework, leveraging the Cheon-Kim-Kim-Song (CKKS) encryption scheme. The framework was implemented using the TenSeal and Torch libraries and evaluated on encrypted heart disease datasets with varying polynomial degrees. The study’s design involved applying the HELR model to three healthcare datasets and comparing its performance with Support Vector Machines (SVM). The major findings revealed that the HELR model achieved high accuracy, within 1% to 3% of its non-HE counterpart, while maintaining competitive computational efficiency. Furthermore, the HELR framework demonstrated robust security against various privacy attacks, including poisoning, evasion, member inference, model inversion, and model extraction, at different ML stages. Notably, the HELR model outperformed SVM in terms of accuracy, showcasing its effectiveness for secure healthcare predictions. The results suggest that HE-enhanced models can offer secure, accurate predictions, paving the way for advancements in privacy-preserving healthcare analytics. However, the study identified limitations related to the computational overhead introduced by HE and the scalability of the model for large datasets. Future work will focus on optimizing encryption techniques and exploring parallel processing methods to address these challenges.
Source: pdf
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | HELR: CKKS-based logistic regression trained on encrypted healthcare data | llm | 0.80 |
| contribution | evaluation across polynomial modulus degrees (4096 vs 8192) and epoch counts | llm | 0.80 |
| contribution | comparison against an HE-based SVM on three datasets | llm | 0.80 |
| contribution | security analysis against poisoning, evasion, membership inference, model inversion and model extraction | llm | 0.80 |
| schemes | CKKS | llm | 0.80 |
| libraries | TenSEAL | llm | 0.80 |
| libraries | PyTorch | llm | 0.80 |
| models | logistic regression | llm | 0.80 |
| models | SVM | llm | 0.80 |
| datasets | heart disease datasets from Kaggle (three) | llm | 0.80 |
| datasets | Statlog (Heart) | llm | 0.80 |
| tasks | heart disease prediction | llm | 0.80 |
| tasks | binary classification | llm | 0.80 |
| tasks | healthcare analytics | llm | 0.80 |
| threat model | training outsourced to an untrusted cloud service; attacks considered include poisoning, evasion, membership inference, model inversion and model extraction | llm | 0.80 |
| interaction | non-interactive FHE | llm | 0.80 |
| limitations | computational overhead introduced by homomorphic encryption | llm | 0.80 |
| limitations | limited scalability to large datasets | llm | 0.80 |
| limitations | binary classification only; multi-class extension left to future work | llm | 0.80 |
| llm relevance | low | llm | 0.80 |
| key result | accuracy: 84% (HELR) vs 85% (plaintext logistic regression) (heart disease dataset) | llm | 0.80 |
| key result | accuracy gap: 1% to 3% below the non-HE model (0.01-0.03) (across the evaluated datasets) | llm | 0.80 |
| key result | accuracy: perfect scores, HELR outperforming HE-SVM (Statlog dataset) | llm | 0.80 |
| key result | AUC: HELR_4096 higher than HELR_8192 (ROC analysis) | llm | 0.80 |
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- scheme: CKKS 41 [p.1], Differential privacy 4 [p.2], BFV 3 [p.4], Leveled / somewhat HE 3 [p.9], MPC 2 [p.23]
- library: Microsoft SEAL 1 [p.3], TenSEAL 1 [p.20], PyTorch 1 [p.20]
- model: Logistic regression 81 [p.1], SVM 5 [p.1], Gradient boosting / XGBoost 2 [p.2], Linear regression 1 [p.6], Decision tree 1 [p.18], k-means 1 [p.4]
- dataset: Heart disease 12 [p.1], COVID 2 [p.23], MNIST 1 [p.23]
- technique: Sigmoid 7 [p.5], Model inversion 6 [p.1], Federated learning 4 [p.2], Model extraction / stealing 4 [p.1], Membership inference 3 [p.3]
- assumption: Composite residuosity 4 [p.20]
- task: Medical / healthcare 81 [p.1], IoT / edge 7 [p.2], Clustering 3 [p.4], Genomics 3 [p.5], Language modeling / generation 2 [p.8]
- topic: Homomorphic encryption 83 [p.1]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/ppml/applications/naresh-2025_privacy-preserving-heart-disease-prediction-with-fhe.pdf |
ppml/applications | 27 | 3.0 MB | 8107ee535ba7 |
was Naresh and Reddi - 2025 - Exploring the future of privacy-preserving heart disease prediction a fully homomorphic encryption-.pdf |
Database key: exploring-the-future-of-privacy-preserving-heart-disease-prediction-a · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set exploring-the-future-of-privacy-preserving-heart-disease-prediction-a COLUMN VALUE