[Paper] Differential Privacy for Secure Machine Learning in Healthcare IoT-Cloud Systems

Source: arXiv - 2512.10426v1

Overview

The paper presents a multi‑layer IoT‑Edge‑Cloud framework for health‑care applications that need both real‑time response (e.g., emergency alerts) and strong privacy guarantees. By weaving differential privacy (DP) into several common machine‑learning (ML) models and coupling the system with blockchain‑based auditability, the authors show how to keep patient data safe while still delivering accurate analytics in a distributed health‑IoT environment.

Key Contributions

• Hybrid IoT‑Edge‑Cloud architecture that routes tasks according to latency‑criticality and storage permanence.
• Differential‑privacy‑enabled ML pipeline covering K‑means, Logistic Regression, Random Forest, and Naive Bayes, with a novel adaptive Laplace‑Gaussian noise mechanism.
• Comprehensive threat model that distinguishes three adversary classes (attribute inference, data reconstruction, and model inversion).
• Empirical evaluation showing supervised models retain 82‑84 % accuracy at a practical privacy budget (ε = 5.0) while cutting inference attacks by up to 18 % and reconstruction correlation by 70 %.
• Blockchain integration for immutable logging, time‑stamping, and traceability of analytics results.
• Edge‑level latency reduction of ≈8× for emergency scenarios, confirming the benefit of hierarchical processing.

Methodology

System Design

The authors split the health‑IoT workflow into three layers:

• IoT devices (wearables, sensors) collect raw vitals.
• Edge nodes (hospital gateways, local servers) perform fast, latency‑sensitive preprocessing and emergency detection.
• Cloud hosts the heavy‑weight ML training and long‑term storage.

Differential Privacy Injection

For each ML algorithm, noise is added to either the training data (input perturbation) or the model parameters (output perturbation). Three mechanisms are compared:

• Laplace – good for low‑dimensional data, heavy tails.
• Gaussian – better for high‑dimensional data, lighter tails.
• Hybrid Laplace‑Gaussian with adaptive budget allocation that distributes the privacy budget (ε) across features based on their sensitivity.

Threat Modeling

Three adversary profiles are defined:

• Class 1: tries to infer specific patient attributes.
• Class 2: attempts to reconstruct raw data from released models.
• Class 3: performs model inversion to extract training records.

Blockchain Auditing

Every analytics request and result is recorded on a permissioned blockchain, providing tamper‑evident logs and enabling traceability for compliance (e.g., HIPAA).

Experimental Setup

Public health datasets (both low‑ and high‑dimensional) are used to train the four ML models under varying ε values (1–10). Accuracy, attack success rates, and latency are measured across the three system layers.

Results & Findings

Takeaway: The hybrid Laplace‑Gaussian mechanism delivers the best privacy‑utility trade‑off, preserving most of the model’s predictive power while dramatically weakening the adversary’s ability to infer or reconstruct patient data. Edge processing cuts emergency‑response latency by an order of magnitude, validating the hierarchical design.

Practical Implications

• Fast Emergency Alerts: Hospitals can deploy edge gateways that run lightweight DP‑protected classifiers locally, triggering alarms within milliseconds—crucial for cardiac events, falls, or drug‑overdose detection.
• Compliance‑Ready Analytics: The built‑in DP guarantees meet regulatory thresholds (e.g., GDPR’s “reasonable risk” standard) without sacrificing the usefulness of predictive models for chronic‑disease management.
• Secure Data Sharing Across Providers: The blockchain audit trail lets multiple clinics or insurers verify who accessed which analytics results, simplifying cross‑institutional collaborations while preserving patient consent records.
• Scalable Cloud Training: Data scientists can train richer models in the cloud on DP‑noised datasets, knowing that downstream deployments (edge or mobile) inherit the same privacy guarantees.
• Developer Toolkit: The paper’s noise‑allocation algorithm can be wrapped into a library (e.g., Python’s dpprivacy), enabling developers to plug‑in DP with a single function call across common ML frameworks (scikit‑learn, TensorFlow).

Limitations & Future Work

• Fixed Privacy Budget: The study uses a static ε = 5 for most experiments; dynamic budgeting based on real‑time risk assessment is left for future exploration.
• Dataset Diversity: Experiments are limited to a few benchmark health datasets; validation on large‑scale, heterogeneous IoT streams (e.g., continuous ECG, multi‑modal imaging) is needed.
• Blockchain Scalability: While the overhead is low for logging, the paper does not address consensus performance under high transaction volumes typical of nationwide health networks.
• Model Generalization: Only classic ML algorithms are examined; extending the DP framework to deep learning (CNNs, RNNs) and federated learning scenarios is an open research direction.

Bottom line: By marrying a hierarchical IoT‑Edge‑Cloud design with a smart hybrid differential‑privacy scheme and blockchain‑backed auditability, the authors chart a practical path toward secure, low‑latency, data‑driven health care—a blueprint that developers can start experimenting with today.

Authors

• N Mangala
• Murtaza Rangwala
• S Aishwarya
• B Eswara Reddy
• Rajkumar Buyya
• KR Venugopal
• SS Iyengar
• LM Patnaik

Paper Information

• arXiv ID: 2512.10426v1
• Categories: cs.CR, cs.DC
• Published: December 11, 2025
• PDF: Download PDF