Sensitivity-Aware Hybrid Differential Privacy for Personalized Federated Learning under Class Imbalance and Data Heterogeneity This repository contains the implementation of the Sensitivity-Aware Hybrid Differential Privacy for Personalized Federated Learning under Class Imbalance and Data Heterogeneity (PR-AS-HDP) framework for healthcare federated learning. The framework was developed to address privacy preservation, class imbalance, and data heterogeneity in distributed healthcare machine learning environments. The project combines: • Personalized Federated Learning (PFL) • Adaptive Sensitivity Estimation • Hybrid Differential Privacy (Local + Central DP) • Adaptive Clipping • Adaptive Privacy Budgeting • Rényi Differential Privacy (RDP) Accounting The framework was evaluated on: • Cervical Cancer Dataset • Diabetes Dataset • Glioma Grading Dataset Research Motivation Healthcare datasets are: • Highly sensitive • Distributed across institutions • Non-IID (Non-Independent and Identically Distributed) • Frequently class-imbalanced Traditional centralized machine learning requires collecting patient data into one location, which raises privacy and governance concerns. Federated Learning (FL) addresses this by allowing collaborative model training without sharing raw data. However, model updates can still leak sensitive information. This project proposes a privacy-aware federated learning framework that dynamically adapts privacy protection according to: • client behaviour, • sensitivity dynamics, • training stability, • and class imbalance.

Main Contributions The project introduces:

1. Personalized Federated Learning o Shared global encoder o Client-specific local heads
2. Adaptive Sensitivity Estimation o Sensitivity is estimated dynamically instead of using fixed worst-case bounds.
3. Hybrid Differential Privacy o Local Gaussian noise at the client side o Central Gaussian noise at the server side
4. Adaptive Privacy Budgeting o Privacy budgets are allocated dynamically across communication rounds.
5. Adaptive Clipping o Clipping bounds change according to update distributions.
6. RDP Accounting o Rényi Differential Privacy is used for tighter privacy accounting.

Framework Architecture The PR-AS-HDP framework integrates the following components: Component Purpose Shared Global Encoder Learns transferable global representations Client-Specific Heads Supports personalization under heterogeneous data Adaptive Clipping Stabilizes updates dynamically Adaptive Sensitivity Estimation Calibrates privacy noise using client behaviour Hybrid DP Combines local and central privacy protection Adaptive Budget Scheduling Dynamically allocates privacy budgets RDP Accountant Tracks cumulative privacy loss

Datasets Used

1. Cervical Cancer Dataset • Highly imbalanced healthcare dataset • Used to evaluate unstable sensitivity behaviour and privacy degradation.
2. Diabetes Dataset • Moderately heterogeneous dataset • Used to evaluate privacy–utility stability.
3. Glioma Grading Dataset • More stable sensitivity dynamics • Used to evaluate personalized learning under strong privacy preservation.

Repository Structure ├── cervical_cancer_colab_ready.py ├── diabetes_colab_ready.py ├── glioma_grading_colab_ready.py ├── README.md ├── requirements.txt └── outputs/

Key Features of the Implementation The implementation includes: • Multi-seed experimentation • Federated client simulation • Non-IID Dirichlet partitioning • Personalized learning • Dynamic clipping • Adaptive sensitivity estimation • Hybrid differential privacy • RDP accounting • Early stopping • Statistical evaluation • Confidence intervals • Privacy fairness verification The codebase supports reproducible experiments across multiple healthcare datasets. Experimental Configuration The experiments typically use: Parameter Value Number of Clients 8 Communication Rounds 30 Client Participation 75% Local Epochs 2 Dirichlet Alpha 0.5 Random Seeds 42, 52, 62 Differential Privacy Delta 1e-5

The implementation uses adaptive clipping and adaptive privacy scheduling during training. Ablation Studies The framework includes four ablation settings: Ablation Description A1 Adaptive clipping only A2 Adaptive clipping + adaptive budgeting A3 Adaptive clipping + adaptive budgeting + adaptive sensitivity A4 Full PR-AS-HDP framework The ablation studies help isolate the contribution of each framework component. Evaluation Metrics The framework evaluates: • Accuracy • Precision • Recall • F1-score • AUPRC (Area Under Precision Recall Curve) • ROC-AUC AUPRC is emphasized because healthcare datasets are often highly imbalanced. Key Research Findings The experiments showed that: • Privacy–utility behaviour is highly dataset-dependent. • Stable sensitivity leads to better utility. • Unstable sensitivity increases privacy noise and degrades performance. • Adaptive mechanisms improve convergence stability. • Differential privacy impacts datasets differently. • Sensitivity dynamics influence performance more than the nominal privacy budget.

Running the Codes Install Dependencies pip install numpy pandas scikit-learn matplotlib torch scipy ucimlrepo Run Cervical Cancer Experiments python cervical_cancer_colab_ready.py Run Diabetes Experiments python diabetes_colab_ready.py Run Glioma Experiments python glioma_grading_colab_ready.py

Outputs Generated The framework automatically generates: • Validation histories • Test performance metrics • Privacy ledgers • Mean ± standard deviation tables • Confidence intervals • Privacy fairness statistics • AUPRC convergence plots

Privacy Mechanisms Local Differential Privacy (LDP) Noise is added locally before client updates are transmitted. Central Differential Privacy (CDP) Additional noise is added during server aggregation. Adaptive Sensitivity Sensitivity is estimated dynamically using: • representation dispersion, • clipped update magnitude, • and class imbalance. RDP Accounting Privacy loss is accumulated across communication rounds using Rényi Differential Privacy accounting.

Intended Use This repository is intended for: • Federated learning research • Differential privacy research • Healthcare AI research • Privacy-preserving machine learning • Graduate and PhD research • Experimental benchmarking

Citation If you use this repository in your research, please cite: Rebecca Adhiambo Okaka, Evanson Mwangi Karanja, Isaac Nyabisa Oteyo. Sensitivity-Aware Hybrid Differential Privacy for Personalized Federated Learning under Class Imbalance and Data Heterogeneity. 2026.

Author Rebecca Adhiambo Okaka Research Areas: • Federated Learning • Differential Privacy • Privacy-Preserving AI • Healthcare Machine Learning • Adaptive Privacy Mechanisms

Acknowledgements This work builds upon research in: • Federated Learning • Differential Privacy • Adaptive Clipping • Personalized Federated Learning • Rényi Differential Privacy • Healthcare AI Research Special thanks to the open-source healthcare machine learning and privacy research communities.