# Federated Learning with Adaptive Sensitivity Hybrid Differential Privacy for Health Internet of Things Systems
This repository contains the implementation of the **Adaptive Sensitivity Hybrid Differential Privacy (AS-HDP)** framework for healthcare federated learning.
The project focuses on improving privacy protection in:
- Healthcare machine learning
- Health Internet of Things (HIoT) systems
- Distributed healthcare analytics
while preserving predictive utility under:
- Non-IID federated learning
- Client heterogeneity
- Dynamic training behaviour
- Class imbalance
The framework integrates:
- Federated Learning (FL)
- Adaptive Sensitivity Estimation
- Adaptive Gradient Clipping
- Hybrid Differential Privacy (Local + Central DP)
- Dynamic Privacy Budget Allocation
- Rényi Differential Privacy (RDP) Accounting
The framework was evaluated using:
- Breast Cancer Wisconsin Diagnostic Dataset
- Diabetes Dataset
- Glioma Grading Dataset
# Research Motivation
The rapid growth of the Health Internet of Things (HIoT) has enabled continuous patient monitoring and data-driven healthcare. However, healthcare data remains highly sensitive and vulnerable to privacy leakage.
Although Federated Learning allows institutions to collaboratively train models without sharing raw patient data, exchanged gradients and model updates can still leak sensitive information through:
- Membership inference attacks
- Gradient leakage
- Reconstruction attacks
Traditional Differential Privacy (DP) and Hybrid Differential Privacy (HDP) methods typically rely on:
- Fixed clipping thresholds
- Static sensitivity assumptions
- Fixed noise calibration
These assumptions often degrade model utility in heterogeneous healthcare environments.
This project introduces an adaptive privacy-preserving federated learning framework that dynamically aligns privacy protection with:
- Observed gradient behaviour
- Sensitivity dynamics
- Update magnitudes
- Training progress
# Main Contributions
The AS-HDP framework introduces:
## 1. Adaptive Sensitivity Estimation
The framework estimates client sensitivity dynamically instead of relying on fixed worst-case assumptions.
## 2. Adaptive Gradient Clipping
Gradient clipping bounds are updated dynamically using:
- Exponential moving averages
- Quantile-based update statistics
## 3. Hybrid Differential Privacy
The framework combines:
- Local Differential Privacy (LDP)
- Central Differential Privacy (CDP)
to strengthen privacy protection.
## 4. Dynamic Privacy Budget Scheduling
Privacy budgets are allocated adaptively across communication rounds based on learning behaviour.
## 5. Rényi Differential Privacy (RDP) Accounting
The framework tracks cumulative privacy expenditure across training rounds using Rényi Differential Privacy accounting.
# Framework Architecture
The AS-HDP framework contains the following main components:
| Component | Function |
|---|---|
| Local Model Training | Client-side federated learning |
| Adaptive Clipping | Dynamic update norm control |
| Client-Level Sensitivity Estimation | Dynamic privacy calibration |
| Local DP Noise Injection | Client-side privacy protection |
| Secure Aggregation | Aggregation of protected updates |
| Adaptive Budget Allocation | Dynamic privacy scheduling |
| Central DP Noise Injection | Server-side privacy protection |
| RDP Accounting | Multi-round privacy tracking |
The architecture aligns privacy enforcement with real gradient behaviour during training.
# Datasets Used
The framework was evaluated using publicly available healthcare datasets.
| Dataset | Task |
|---|---|
| Breast Cancer Wisconsin Diagnostic | Binary cancer classification |
| Diabetes Dataset | Diabetes prediction |
| Glioma Grading Dataset | Brain tumour grading classification |
The datasets simulate realistic heterogeneous healthcare federated learning environments.
# Repository Structure
```text
├── Breast_Cancer_AS_HDP.py
├── Diabetes_AS_HDP.py
├── Glioma_AS_HDP.py
├── README.md
├── requirements.txt
└── outputs/The implementation supports:
- Federated learning simulation
- Non-IID Dirichlet client partitioning
- Multi-seed experimentation
- Adaptive clipping
- Adaptive sensitivity estimation
- Hybrid Differential Privacy
- Dynamic privacy scheduling
- Rényi DP accounting
- Contribution-aware aggregation
- Statistical evaluation
- Confidence intervals
- Privacy fairness verification
- Validation AUPRC tracking
The codes are designed for:
- Reproducibility
- Healthcare privacy research
- Experimental benchmarking
Typical experimental settings include:
| Parameter | Value |
|---|---|
| Number of Clients | 8 |
| Communication Rounds | 30 |
| Client Participation | 75% |
| Local Epochs | 2–3 |
| Dirichlet Alpha | 0.5–0.6 |
| Random Seeds | Multi-seed evaluation |
| Differential Privacy δ | 1e−5 |
The implementation also performs:
- Fairness checks
- Sequential privacy composition
- Confidence interval analysis
The framework includes several ablation configurations.
| Ablation | Description |
|---|---|
| NO_DP | Federated Learning without Differential Privacy |
| FIXED_HDP | Hybrid DP with fixed sensitivity assumptions |
| A1 | Adaptive clipping only |
| A2 | Adaptive clipping + adaptive budgeting |
| A3 | Adaptive clipping + adaptive sensitivity |
| A4 | Full AS-HDP framework |
The ablation studies isolate the contribution of each framework component.
The framework evaluates:
- Accuracy
- Precision
- Recall
- F1-score
- AUPRC (Area Under Precision Recall Curve)
AUPRC is emphasized because healthcare datasets are highly imbalanced.
The clipping bound is updated dynamically using:
- Exponential Moving Averages (EMA)
- Quantile-based update norm estimation
Sensitivity estimation combines:
- Feature variability
- Update magnitudes
- Client heterogeneity
The framework applies:
- Local DP noise at the client side
- Central DP noise during server aggregation
The framework tracks cumulative privacy expenditure using:
- Rényi Differential Privacy (RDP)
- Conversion of accumulated privacy cost into final ((\epsilon, \delta))-DP guarantees
The experiments show that:
- Static privacy mechanisms are often suboptimal in healthcare federated learning
- Adaptive sensitivity improves privacy calibration
- Adaptive clipping improves convergence stability
- Hybrid DP reduces privacy leakage risks
- Privacy–utility behaviour is highly dataset-dependent
- Sensitivity dynamics strongly influence model utility
- AS-HDP reduces seed variability compared with fixed HDP
- AS-HDP preserves clinically meaningful performance under privacy constraints
bash pip install numpy pandas scikit-learn matplotlib torch scipy
bash breast_cancer_as_hdp_colab_ready.py
bash diabetes_as_hdp_colab_ready.py
bash glioma_as_hdp_colab_ready.py
The framework automatically generates:
- Validation histories
- Test performance metrics
- Privacy ledgers
- Mean ± standard deviation tables
- 95% confidence intervals
- Privacy fairness statistics
- Convergence curves
- AUPRC plots
This repository is intended for:
- Healthcare AI research
- Federated Learning research
- Differential Privacy research
- Health Internet of Things (HIoT) systems
- Privacy-preserving machine learning
- Graduate and PhD research
- Experimental benchmarking
If you use this repository in your research, please cite:
bibtex @article{okaka2026ashdp, title={Federated Learning with Adaptive Sensitivity Hybrid Differential Privacy for Health Internet of Things Systems}, author={Okaka, Rebecca Adhiambo and Karanja, Evanson Mwangi and Oteyo, Isaac Nyabisa}, year={2026} }
- Federated Learning
- Differential Privacy
- Healthcare AI
- Privacy-Preserving Machine Learning
- Adaptive Privacy Mechanisms
This work builds upon research in:
- Federated Learning
- Differential Privacy
- Adaptive Clipping
- Rényi Differential Privacy
- Healthcare AI
Special thanks to the open-source research community for foundational contributions to privacy-preserving machine learning.