Source-linked AI summary
Federated Learning for Privacy Preservation in Smart Healthcare Systems: A Comprehensive Survey
Mansoor Ali, Faisal Naeem, Muhammad Tariq, Geroges Kaddoum
TL;DR
Smart healthcare IoMT systems create privacy risks because centralized AI can expose confidential information. This survey synthesizes IoMT privacy issues, FL-based protection, advanced DRL, digital-twin, and GAN architectures, and practical applications, concluding with open research challenges. It identifies FL as a privacy-preserving approach while noting data limitations, communication demands, and residual privacy risks.
Problem
Centralized healthcare learning and limited data sharing expose confidential patient information, while single-institution datasets may be insufficient and imbalanced for accurate models.
Method
The survey analyzes IoMT privacy issues, FL architectures and applications, and advanced schemes incorporating DRL, digital twins, and GANs.
Results
The survey presents FL as a viable privacy-preserving approach for smart healthcare and reviews applications including medical image processing and COVID-19 detection.
Takeaways & Limitations
FL can support collaborative healthcare-model training without directly accessing participants’ confidential local data.
Takeaways & Limitations
Healthcare FL remains constrained by insufficient, imbalanced, and feature-limited data, while communication-resource scheduling is challenging in large IoMT networks.
Abstract
from arXiv · showhide
Recent advances in electronic devices and communication infrastructure have revolutionized the traditional healthcare system into a smart healthcare system by using IoMT devices. However, due to the centralized training approach of artificial intelligence (AI), the use of mobile and wearable IoMT devices raises privacy concerns with respect to the information that has been communicated between hospitals and end users. The information conveyed by the IoMT devices is highly confidential and can be exposed to adversaries. In this regard, federated learning (FL), a distributive AI paradigm has opened up new opportunities for privacy-preservation in IoMT without accessing the confidential data of the participants. Further, FL provides privacy to end users as only gradients are shared during training. For these specific properties of FL, in this paper we present privacy related issues in IoMT. Afterwards, we present the role of FL in IoMT networks for privacy preservation and introduce some advanced FL architectures incorporating deep reinforcement learning (DRL), digital twin, and generative adversarial networks (GANs) for detecting privacy threats. Subsequently, we present some practical opportunities of FL in smart healthcare systems. At the end, we conclude this survey by providing open research challenges for FL that can be used in future smart healthcare systems
I. INTRODUCTION
IoMT enables personalized and intelligent healthcare, but centralized AI creates scalability, security, and privacy risks. The survey examines FL and emerging architectures for privacy-preserving smart healthcare.
- IoMT devices support continuous monitoring and AI-enabled applications including disease prediction and remote health monitoring.
- Centralized healthcare AI faces scalability problems and exposes sensitive patient data to security threats.
- FL trains models by averaging local updates without accessing local data, reducing risks to user preferences and patient information.
- Earlier surveys addressed FL applications or healthcare implementation, but did not comprehensively cover privacy-preserving advances using RL, digital twins, and GANs.
- The survey covers IoMT security and privacy, FL motivation and architecture, advanced privacy-oriented architectures, applications, and research directions.
II. SECURITY AND PRIVACY OF IOMT
Smart healthcare systems face serious cybersecurity and privacy threats because connected medical infrastructure can be attacked and sensitive patient information exposed. The paper emphasizes risk analysis and adaptable protections before deployment.
- Healthcare cybersecurity remains a major concern because electronic healthcare systems are vulnerable targets for attackers.
- Attackers may paralyze healthcare systems, steal patient information, sell it, or blackmail patients into releasing personal data.
- Security and privacy protections should follow risk analysis and satisfy system requirements while adapting to changes in healthcare systems.
- Privacy protection requires attention alongside network security, supported by dedicated threat descriptions and protection schemes.
1) Private data:
IoMT private data includes information linked to individuals and can be compromised during communication, storage, or analysis. Protection approaches combine access control, anonymization, encryption, and privacy-aware processing practices.
- Personal Identifiable Information (PII) identifies a specific individual and can be classified as sensitive personal, general, or statistical data.
- PII owners retain authority over their information, while processors receive rights to access and use it for specified purposes.
- IoMT privacy threats include tampering with smart-device connections and accessing information through wearable health-monitoring devices.
- Privacy controls can reduce requested collection, shorten storage periods, and move processing to edge servers to reduce computation time.
- Users may protect identity through authentication, pseudonyms, or anonymity, while adversaries can infer information by analyzing data flows.
- Pseudonyms protect critical data, but context-based inference can still reveal user roles, locations, and identities.
- 3) Privacy of IoMT using identification and anonymity:: Cloaking areas, semantic obfuscation, and k-anonymity strengthen protection, although k-anonymity requires a trustworthy intermediate entity.
III. FL AND ITS PERSPECTIVE IN IOMT
Traditional ML centralizes healthcare data and incurs transmission costs, whereas FL distributes training across end devices while transferring model parameters to a global model. The section introduces FL architectures and identifies testing-model selection as an open issue.
- Traditional ML stores data centrally without necessarily addressing privacy protection or data-transmission costs.
- Selecting an appropriate testing model for IoMT privacy and security algorithms remains an open research issue between centralized and FL-based approaches.
- FL performs learning on end-device data and transfers model parameters to a central global model for privacy and traffic-management benefits.
- FL architectures include horizontal FL, vertical FL, and transfer FL.
A. Framework for FL based healthcare system
The FL-enabled healthcare framework trains models across IoMT end nodes while keeping local data at those nodes. This addresses privacy and data-sufficiency challenges, while reducing information leakage and communication demands.
- The central server selects the healthcare task, algorithm, learning rates, and configurable parameters before initiating federated training.
- End nodes train the shared model on local data and return updated parameters to the central server for aggregation.
- Federated averaging assigns weights to local model parameters according to available data size when computing the global model.
- Single-laboratory datasets may be insufficient, imbalanced, and feature-limited, restricting model accuracy and motivating collaboration across medical institutes.
- Keeping host data within local IoMT nodes increases privacy and reduces information leakage while collaborative training uses diverse data.
- Sharing gradients rather than larger datasets reduces communication cost, supports bandwidth utilization, and helps avoid congestion in massive IoMT networks.
IV. FEATURED FL DESIGN FOR IOMT
The survey presents advanced FL architectures for IoMT privacy, motivated by attacks that can extract sensitive information from global model updates and shared data.
- Advanced FL architectures are summarized from different perspectives to address privacy problems in IoMT networks.
- Construction and inference attacks can extract user information, including blood samples, disease types, and other shared data, from FL processes.
- GAN-generated attack patterns have been used to design protection mechanisms, including the robust FL mechanism EaSTFLy verified against IoT networks.
2) Poisoning attack:
The survey describes poisoning and related privacy attacks against FL, alongside defenses using GANs, blockchain, digital twins, privacy-preserving schemes, differential privacy, and encryption.
- Poisoning attack: Data poisoning tampers with training data, whereas model poisoning tampers with model parameters to reduce federated model performance.
- Poisoning attack: GAN-based prevention mechanisms perform data auditing to address poisoning attacks.
- Poisoning attack: Malicious nodes can share fake parameters or false data with neighbors, reducing model convergence time and accuracy; blockchain-incorporated FL targets Byzantine attacks.
- Poisoning attack: Digital-twin-enabled FL incorporates malicious data into consideration to help prevent privacy threats.
- Poisoning attack: Differential attacks can reveal whether end devices are dedicated to particular tasks, motivating defenses combining reinforcement learning, blockchain, and differential privacy.
- Poisoning attack: Differential privacy and homomorphic encryption are used in federated gradient boosting, while noise can be added to IoMT datasets to protect private information.
B. Incentive enable FL for IoMT
FL incentives address limited device participation caused by constrained IoMT computation and concerns about third-party platforms. The survey also situates FL within digital-twin applications and broader IoMT opportunities.
- B. Incentive enable FL for IoMT: Vanilla FL requires IoMT devices to participate and share model updates, but limited computation, privacy, and trust concerns can reduce willingness to contribute.
- B. Incentive enable FL for IoMT: Incentive-based FL techniques are proposed to involve devices that otherwise may not participate in federated training.
- B. Incentive enable FL for IoMT: Game-theoretic mechanisms, including a Stackelberg game, model changing device computation resources and connection stability between IoMT devices and hospital aggregators.
- B. Incentive enable FL for IoMT: A digital twin creates a virtual patient environment where doctors can test prescribed medication before using it on real-world patients.
- B. Incentive enable FL for IoMT: The survey discusses FL applications in IoMT after reviewing privacy issues and prevention techniques.
A. Electronic health record management using FL
FL supports privacy-preserving healthcare record and medical-image analysis by training across institutions without sharing raw patient data. The surveyed applications combine FL with perturbation, differential privacy, and GAN-generated data to reduce exposure risks.
- EHR management using FL: FL enables collaborative EHR analysis across institutions without directly sharing local patient data.Perturbation of training data is used to protect local model parameters against memorization attacks.
- Medical image processing: 97% accuracy was observed on cancer datasets for an FL medical-image algorithm using GAN-generated raw image datasets and outperforming non-FL schemes.The approach combines images from different clients into a global-server image and shares raw imprints rather than actual images.
- Brain imaging: Differential privacy adds noise to local brain-imaging model updates, reducing information-leakage risk during parameter sharing.Clients train DNN models on brain-tumor data and share updated parameters for central aggregation.
- Advanced architectures: The survey identifies advanced FL architectures for privacy preservation in IoMT networks.The supplied table caption names advanced FL architecture as its subject but provides no row-level findings.
C. Role of FL in COVID- 19
FL is presented as a privacy-preserving approach for COVID-19 detection because hospitals can train on local X-ray data while sharing gradients with a global model. A dynamic variant further manages hospital participation and client selection during training.
- COVID-19 detection: FL supports COVID-19 detection by allowing hospitals to train locally on X-ray images and share only gradients with a global model.This addresses privacy concerns that make abundant medical-data collection and sharing difficult.
- COVID-19 detection: COVID-19 detection is motivated by the disease’s status as a major health threat and by the use of CNN-based chest-X-ray analysis.The passage describes CNNs as DNN-based algorithms used for early detection through chest-X-ray feature extraction.
- Dynamic FL: Dynamic FL uses hospital participation decisions and server-side client selection based on model performance and update timing.Hospitals decide whether to participate in each iteration, while the server selects models for aggregation.
- Communication requirements: FL deployment in smart healthcare requires sophisticated resource allocation because many IoMT devices exchange model parameters over dynamic uplink and downlink networks.The communication network is identified as a crucial deployment factor for transmitting model updates.
B. Universal FL architecture for IoMT
The survey frames FL as a privacy-preserving option for IoMT but identifies unresolved challenges in evaluation, heterogeneous client data, and future wireless-network conditions. It surveys architectures and enhancements involving DRL, DNN, and GANs.
- Universal FL architecture for IoMT: Universal standards and protocols are needed to evaluate FL algorithms consistently across IoMT settings.Comparative analysis is difficult because blockchain-based approaches use different datasets and application-specific model specifications.
- Universal FL architecture for IoMT: Real-world clients may hold heterogeneous images, video, or text with differing medical features, whereas many privacy-preserving FL algorithms assume similar features.The survey calls for more robust heterogeneous FL mechanisms.
- Universal FL architecture for IoMT: Future 6G healthcare deployments will generate abundant data, requiring study of how 5G/6G network conditions affect FL performance.The passage presents 6G adoption as relevant across applications including body area networks.
- Universal FL architecture for IoMT: The survey covers IoMT privacy issues, FL architectures, practical applications, and research directions, including DRL, DNN, and GAN enhancements.These techniques are introduced for improving FL performance from a privacy perspective.