Skip to content
Crypto.world

Research

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

Healthcare wearables and connected medical devices generate sensitive patient data that is stored and analysed in cloud systems by many stakeholders — clinicians, insurers, researchers…

Differential Privacy for Secure Machine Learning in Healthcare IoT-Cloud Systems
arXiv
2512.10426
Published
2025-12-11
Authors
N Mangala, Murtaza Rangwala, S Aishwarya, B Eswara Reddy, Rajkumar Buyya, KR Venugopal, SS Iyengar, LM Patnaik

AI summary

Overview

Research area: Privacy-preserving machine learning for healthcare Internet of Things (IoT)–cloud systems, combining Differential Privacy (DP), edge computing, and blockchain.

Technical level: Intermediate — readers should be comfortable with basic machine learning classifiers and the idea of a privacy budget, but the paper explains DP, noise mechanisms, and architecture layers from first principles.

Scope (one sentence): The paper proposes and evaluates a three-tier IoT–Edge–Cloud healthcare architecture that applies differential privacy (Laplace, Gaussian, and a hybrid of the two) to four machine learning models, adds blockchain for integrity, and measures the resulting privacy–utility–latency trade-offs under a three-class adversary threat model.

What This Paper Is About

Healthcare wearables and connected medical devices generate sensitive patient data that is stored and analysed in cloud systems by many stakeholders — clinicians, insurers, researchers, pharmaceutical companies. Removing names and addresses does not protect patients, because combinations of other fields can re-identify them. The paper's goal is to build a system that keeps emergency response fast (by pushing time-critical processing to the edge) while mathematically guaranteeing that machine learning models trained on patient data do not leak information about any individual.

Key Contributions

  1. A multi-layer IoT–Edge–Cloud architecture that distributes healthcare tasks according to response criticality and data permanence requirements, in order to improve response times for emergency scenarios.
  2. An adaptive hybrid Laplace–Gaussian noise mechanism for differentially private machine learning, which selects noise distributions based on the sensitivity characteristics of individual data features and the specific algorithm used, rather than applying one noise type uniformly.
  3. A systematic empirical evaluation of the privacy–utility trade-off across four machine learning algorithms — K-Means (unsupervised), Logistic Regression, Random Forest, and Naive Bayes (supervised) — within the multi-layer architecture, yielding algorithm-specific recommendations for privacy budget allocation.
  4. Blockchain integration for data integrity, tamper resistance, trust, and provenance through time-stamping, traceability, and immutability, analysed jointly with DP and hierarchical access control under a single threat model.

Main Findings

  • Supervised accuracy ceiling: Supervised algorithms achieved up to 83.6% accuracy under differential privacy across the evaluated privacy budgets.
  • Accuracy at the practical operating point: At ε = 5.0, supervised algorithms achieved 80–81% accuracy.
  • Attack mitigation at ε = 5.0: Attribute inference attacks were reduced by up to 18%, and data reconstruction correlation was reduced by 70%.
  • Latency gains from edge computing: Edge computing demonstrated an 8× latency reduction for emergency scenarios, which the authors present as validation of the hierarchical architecture for time-critical operations.
  • Hybrid noise is the balanced choice: The hybrid Laplace–Gaussian mechanism with adaptive budget allocation offered moderate distribution tails and better privacy–utility trade-offs than either pure mechanism, for both low- and high-dimension datasets. The paper frames this as exploiting Laplace noise's strength on low-dimensional, sparse features and Gaussian noise's strength on high-dimensional, dense features with composition properties.
  • Threat model: The authors define three adversary classes and analyse inference attacks, data tampering, and consensus attacks. The paper claims the three security mechanisms (DP, blockchain, hierarchical access control) provide complementary protections without conflicting requirements or degrading system performance.

Methodology in Plain English

The authors start from a cautionary case: Sweeney re-identified the medical records of the Governor of Massachusetts using only three queries on an anonymised hospital dataset released by the Massachusetts Group Insurance Commission, cross-referenced against public voter registration records. That motivates a formal privacy guarantee rather than anonymisation by deletion.

They then design a three-tier system. Edge nodes handle data that needs fast analysis — blood pressure, oxygen saturation, cardiac rhythms — so clinicians can act quickly. The cloud handles long-term storage and heavy analytics such as longitudinal research across institutions. Differential privacy is applied to the machine learning pipeline by injecting calibrated noise, where the noise scale is set from the query or model sensitivity and the privacy budget: for Laplace noise the scale is b = Δf/ε, and for Gaussian noise the standard deviation is σ = √(2 log(1.25/δ)) · Δ₂f/ε, giving (ε, δ)-DP with L2 sensitivity. They compare Laplace, Gaussian, and their hybrid adaptive variant over a range of privacy budgets, train the four algorithms on the noised data, and measure both accuracy and the success of simulated attacks. Blockchain provides an immutable, time-stamped, traceable log of healthcare transactions. The evaluation is structured around a threat model with three adversary classes. The specific datasets and their sizes are not specified in the available text.

Why This Matters

Research impact: Prior work studied DP–ML integration, blockchain-based distributed ML, and blockchain healthcare systems largely in isolation, and DP–ML frameworks have focused mostly on deep learning in centralised clouds. This paper's distinctive move is to characterise how edge-layer computational constraints and latency requirements affect the practical achievability of privacy guarantees, and to analyse DP, blockchain, and hierarchical access control as one integrated system rather than pairwise combinations.

Real-world applications:

  • Remote patient monitoring and emergency response — wearable and connected device data processed at the edge for time-critical alerts, with 8× latency reduction reported.
  • Clinical research on shared records — multi-institution analysis of electronic health records and clinical trial data without exposing individual patients, which the paper links to HIPAA and GDPR compliance obligations.
  • Insurance and actuarial analytics — aggregate risk modelling where individual policyholders cannot be re-identified, relevant given the paper's note that USA healthcare expenditure reached nearly 20% of GDP in 2021.
  • Auditable medical record systems — blockchain-backed traceability and time-stamping for provenance of treatments and hospitalisations.

Industry relevance: The paper's context is concrete: 692 large healthcare data breaches were reported between July 2021 and June 2022, and organisations must satisfy both HIPAA (Security, Privacy, and Breach Notification Rules) and GDPR, which classifies health data as a special category requiring explicit consent and data protection by design. A framework that offers a tunable privacy budget is directly usable by cloud healthcare providers, device manufacturers, and analytics vendors who need a defensible privacy argument rather than ad hoc de-identification.

Future Directions

  • Feature-aware noise selection as an automated policy: The paper motivates choosing Laplace versus Gaussian noise from feature sensitivity and dimensionality, but leaves open how to determine those mappings automatically for new datasets and algorithms.
  • Extending beyond the four evaluated models: The evaluation covers K-Means, Logistic Regression, Random Forest, and Naive Bayes; whether the hybrid mechanism's advantages hold for deep learning models — the focus of much prior DP-ML work — is an open question.
  • Composition across repeated queries and model retraining: The paper invokes composition properties of Gaussian noise and discusses query-level privacy budgeting, but the cumulative privacy loss from continuous IoT data streams is a natural next problem.
  • Blockchain overhead versus latency gains: Blockchain consensus adds communication overhead, which the literature review notes as a limitation in related work; quantifying that cost against the reported 8× edge latency reduction in a deployed system remains to be done.

Target Audience

Best suited for: graduate students and applied researchers in privacy-preserving machine learning and IoT/edge systems; healthcare cloud architects and security engineers evaluating differential privacy for regulated workloads; and compliance-oriented practitioners who need to connect technical privacy mechanisms to HIPAA and GDPR obligations. Readers wanting a rigorous DP theory treatment or a large-scale benchmark against state-of-the-art DP libraries will find the paper is more of an applied system evaluation than a theoretical contribution.

Authors’ abstract

Healthcare has become exceptionally sophisticated, as wearables and connected medical devices revolutionize remote patient monitoring, emergency response, medication management, diagnosis, and predictive and prescriptive analytics. Internet of Things and Cloud computing integrated systems (IoT-Cloud) facilitate sensing, automation, and processing for these healthcare applications. While real-time response is crucial for alleviating patient emergencies, protecting patient…

The full abstract on arXiv
Read the original paper