Implementing federated learning over VPN-based wireless backhaul networks for healthcare systems

Federated learning in healthcare.

Healthcare data often appear fragmented due to the intricate structure of the healthcare system and its associated procedures. For instance, various hospitals may have access exclusively to the clinical records of their respective patient groups. These records contain highly sensitive information, including protected health data (PHI) of individuals. To govern the access and analysis of such data, stringent regulations like the Health Insurance Portability and Accountability Act (HIPAA) have been established (Gostin, 2001). This presents a significant obstacle for contemporary data mining and machine learning (ML) methodologies, including deep learning (LeCun, Bengio & Hinton, 2015), as they typically rely on substantial volumes of training data.

Federated learning shows significant potential in the field of healthcare data analytics. It offers promising prospects for both provider-oriented tasks, such as creating models to predict hospital readmission risks using patient Electronic Health Records (EHR) (Min, Yu & Wang, 2019), and consumer-focused applications, like detecting atrial fibrillation using electrocardiograms recorded by smartwatches (Perez et al., 2019). This methodology ensures that sensitive patient data remains confidential within local institutions or with individual consumers throughout the federated model learning process, thereby preserving patient privacy effectively. Federated learning involves the challenge of training a top-notch shared global model using a central server, drawing data from decentralized sources spread across a multitude of diverse clients (see Fig. 1).

Challenges and opportunities in FL over wireless network.

In federated learning, local models are trained on end-devices and then combined on a central server. The global model is continuously sent back to end-devices until convergence is achieved. Federated learning relies on iterative interaction between end-devices and the aggregation server, necessitating efficient optimization schemes like FedAvg and FedProx. Subsequently, the aggregated weights (global model) are returned to the end-devices. Despite its benefits, federated learning encounters challenges such as privacy issues, resource optimization, incentive mechanisms, and statistical and system heterogeneity (McMahan et al., 2017). Malicious aggregation servers or end-devices could potentially extract sensitive information from others’ model updates, emphasizing the need for effective privacy preservation mechanisms. In summary, the discussion highlights the importance of addressing key questions regarding federated learning over wireless networks (Hong et al., 2021).

How can we design a federated optimization scheme to handle statistical and system heterogeneity effectively? How can we optimize computing and communication resources to shorten the global convergence time of federated learning? How do we enhance privacy in federated learning despite the existence of a malicious aggregation server?

To answer the above questions, we present a proposed network framework based on Wireless Backhaul Network.

Aim of the study

Federated learning (FL) over wireless networks incorporates ML and Internet of Everything (IoE) applications, but there is room for improvement in resource allocation, security, and addressing statistical variations in FL aggregation algorithms. We have identified two key areas for improvement: (1) Enhancing wireless channels with terahertz spectrum, and (2) improving privacy in FL-based healthcare networks through a secure cross-silo solution using fixed wireless backhaul networks.

In this study, we explore a novel approach in the FL Briefing that leverages terahertz wireless channels and fixed wireless backhaul networks to address the challenges of resource efficiency and security. Additionally, we propose a design for implementing FL for healthcare using VPN-based wireless backhaul networks, including terahertz. This contribution is novel in the area of federated learning for healthcare improvement.

Article structure

The article is structured as follows: It begins with an introduction and motivation for the study. “Proposed Solution” provides comprehensive details of the proposed solution and its key factors, while “Methodology” covers the methodology of the study. “Experiment Setup” and “Results and Discussion” explain the experiments, results, and discussion. Finally, the conclusion and future work of the study are presented.

Key factors of proposed architecture

In this section, we study the communication and security aspects of the proposed solution, focusing on several key considerations on which our proposed network is based and implemented:

• Private versus public network

• Licensed versus unlicensed channel

• Cross-silo versus cross-device

• Enhanced communication wireless channel

Enhance wireless communication channel

We are dealing with a private network, specifically a wireless backhaul network. This network operates only under licensed spectrums such as microwave, mmWave, and the newly studied spectrum, terahertz (THz). In our previous study We propose terahertz-based wireless backhaul as integral to 6G due to its numerous benefits: significantly higher bandwidth in the terabits per second (Tbps) range, exceptional user data rates up to 100 Gbps, ultra-low latency under 0.1 ms, improved spectrum efficiency, and a tenfold increase in bandwidth compared to traditional solutions. This technology promises unprecedented speeds and is ideal for next-gen applications and services. The suggested design is versatile and perfectly suited for various wireless backhaul networks operating at frequencies starting from 38 GHz and beyond. Its capabilities, combined with the benefits of terahertz technology, establish it as a promising and forward-looking solution to meet the changing connectivity requirements of the 6G era (Mahmood et al., 2024).

Flow of experiments

In this section, we explain the flow of the experiment and how federated ML performs the classification of signals. In Fig. 6, the stepwise approach for federated ML is illustrated. In Step 1, the central server sends the latest model parameters to the nodes. Step 2 involves data collection at each node. In Step 3, each local model is trained based on the latest parameters. Step 4 consists of communicating the updated model parameters back to the global model. In Step 5, the updates from each model are combined to retrain the global model, resulting in a new model. Finally, Step 6 restarts the process from Step 1.

Machine learning for classification: state of the art

Machine learning a broad topic in the field of artificial intelligence (AI) that involves developing algorithms and models that enable computers to learn from and make predictions or decisions based on data. Machine learning has the potential to revolutionize various industries by making predictions, automating tasks, and extracting valuable insights from data In the realm of machine learning, numerous situations present difficulties when it comes to transferring data to a central server for model training. To tackle these issues, an alternative approach is to employ distributed machine learning, also detailed in Verbraeken’s survey (Verbraeken et al., 2020).

Neural network—federated learning

Distributed machine learning involves training numerous models across servers that are located in different geographical locations, and these models are later combined to create a unified machine learning model. Federated learning is a key modern-day solution for addressing the challenges of training a machine learning model on extremely large datasets, such as astronomical data, which include substantial computational energy consumption and extended training durations (Li et al., 2020). It is a machine learning method enabling training models on decentralized devices like smartphones, IoT devices, or local servers, without transferring data to a central server. In contrast to traditional machine learning, where data is usually gathered and sent to a central server for model training. However, in federated learning, the model is sent to the data sources, and model updates are performed locally on these devices. In this way, FL earns many advantages over other approaches such as privacy preservation, reduced data transfer, decentralization and model customization.

In recent times, there has been a notable increase in attention towards the field of healthcare data analytics. This heightened interest is attributed to the growing availability of healthcare data from diverse sources, such as clinical institutions, individual patients, insurance firms, and pharmaceutical companies. This influx of data presents an unparalleled opportunity to develop computational methods aimed at extracting valuable insights from the data. These insights, driven by data, have the potential to enhance the quality of healthcare delivery significantly (Wang & Preininger, 2019).

In applying federated learning referred to Fig. 1, to classify diseases on computer systems, we prioritize patient data privacy and network efficiency. Instead of sending raw data to a central server, a decentralized approach is adopted. Remote devices periodically communicate with a central server to obtain a global model. During each communication round, a specific subset of devices conducts local training on their distinct patient data, sending updates to the server. The server adjusts the global model and transmits it to another subset of devices. This iterative process continues until convergence or a predefined stopping criterion is met.

Description of datasets

The data is in the form of a signal extracted from a biofunctionalized SPR-TFBG sensor. To clarify and simplify, we explained how the biofunctionalized SPR-TFBG sensor works. The primary function of the SPR-TFBG biosensor is to differentiate the spectral details between a virus-free solution and a virus-containing solution. The SPR mode’s spectral responses vary with different concentrations of the EV-A71 virus in both clean solutions and solutions with varying levels of impurities. Readers are encouraged to review the additional information on the development and experiments of the SPR-TFBG biosensor found in Udos et al. (2021).

The horizontal axis corresponds to the wavelength, and the vertical axis represents the amplitude of each wavelength in decibels (dBm). Figure 9 illustrates the relationship between wavelength and amplitude for each solution. Referring to the raw data, the signal waveform spans from 1,490 nm to 1,570 nm, with the y-axis representing the output spectrum in dBm, ranging from −21 to −15. Each signal comprises 4,000 individual features.

In other words, one sample or single patient data has 4,000 individual features based on the waveform, and their values vary from −21 to −15. The matrix of the complete dataset for all impurities and a clean solution has a dimension of n×4000, where 4,000 are the number of features and n is the number of samples. In the next step of dimension reduction, the region of interest (ROI) is applied. The region of interest ranges from 1,510 nm to 1,540 nm, meaning the ROI has 1,496 features of different wavelengths, and the dimension has now been reduced to n×1,496. There are five different classes available in the datasets based on expert opinion and analysis: PBS, 02 µL, 08 µL, 32 µL, and 128 µL. PBS refers to the clean solution, meaning there are no virus strains present, while 02 µL to 128 µL refer to impure solutions with varying virus intensities and types. The sample dataset is presented in Table 1. There are a total of four clients taking part in the classification process. Client 1 contributes 598 samples, Client 2 contributes 287 samples, Client 3 has 275 samples, and Client 4 has 162 samples. There are 264 samples tested in each solution: PBS, 02 µL, 08 µL, 32 µL, and 128 µL. The total number of tested samples is 1,320 when all the samples from all the clients are consolidated.

The merged dataset for all the impurities and a clean solution is presented in Table 1. In the final dataset, each wavelength is treated as the individual feature of the samples. In the final image of the dataset presented in Table 2, there are around 4,000 features for every sample. Each sample is labeled as 0 ul, 02 ul, 08 ul, 32 ul, and 128 ul.

Feature analysis

The section is divided into feature extraction and feature engineering for both centralized and federated ML approaches. In each part, we explained the technique that is used, and then their results for both approaches are presented separately.

Feature selection

Following feature extraction, the subsequent stage involves feature selection. Numerous techniques exist to tackle the issue of diminishing irrelevant and redundant features that can complicate a demanding task. Generally, the feature selection methods fall into categories such as filters, wrappers, and embedded methods, as depicted in Fig. 10.

Filter methods —mutual information.

Filter techniques utilize variable ranking strategies as the fundamental basis for selecting variables by means of ordering. The selection of ranking methods stems from their simplicity and proven effectiveness within practical contexts. A fitting ranking criterion is applied to assign scores to the variables, and subsequently, a threshold is employed to exclude variables that do not meet the threshold. Ranking methods operate akin to filter methods, given that they are implemented prior to classification, with the aim of sieving out less pertinent variables. A fundamental quality of a distinctive feature lies in its capacity to encompass valuable insights concerning the various classes inherent in the data. This attribute can be defined as feature relevance, offering a metric for assessing the feature’s capability to distinguish among the diverse classes present. In this investigation, we have chosen mutual information as the filter approach to designate the selected features (Chandrashekar & Sahin, 2014).

Utilizing the concept of mutual information (MI) introduced by Deng, Luo & Yu (2014), a term weighting scheme was developed. For a given term ti and a document set Dc (comprising documents within a specific category), the MI between them is computed using the subsequent formula:

M I t i , D c = l o g P t i , D c P t i x P D c

Figures 11, 12, 13, 14 and 15 illustrate the feature set along with their MI Score using the method for both centralized ML and federated ML. From the selected features, we targeted the top 80% and compare the results with the sequential feature selection method. These selected feature sets are further employed for classification purposes. In the classification phase, we selected all five feature sets and run classifier, evaluating the performance and accuracy based on the established criteria.

The top eight out of 10 feature sets used for centralized machine learning classification and for distributed clients in federated networks refer to specific attributes or characteristics of data selected based on mutual information, as presented in Table 4.

Performance evaluation

The classification process serves as the ultimate and crucial stage for the identification and categorization of both pure and impure solutions, each characterized by varying degrees of impurities. According to the established methodologies, the features identified and listed in Table 4 via employing MI will be employed to classify the signals. Consequently, the signals will be grouped into virus-free or virus-present categories, depending on the impurity level. The dataset was divided into separate testing and training sets. In order to assess the results obtained across different classifiers, the following metrics were employed:

• True positives (TP) correspond to cases where the expert-annotated positive class coincides with the model’s positive class prediction.

• False positives (FP) represent instances where the expert-annotated negative class is erroneously predicted by the model as the positive class.

• True negatives (TN) indicate instances where the expert-annotated negative class is accurately predicted as the negative class by the model.

• False negatives (FN) denote cases where the expert-annotated positive class is incorrectly classified by the model as the negative class. These performance metrics were employed in this study to assess the outcomes achieved through various classifiers.:

Experiment with neural network—centralized ML

This study implemented a neural network using Python and the Keras library, which is a user-friendly high-level Python library designed for deep learning. Keras seamlessly integrates with TensorFlow (or Theano or CNTK), focusing on essential deep learning concepts such as creating neural network layers while managing tensor details. TensorFlow (or Theano or CNTK) acts as the backend for Keras. By utilizing Keras in deep learning applications, practitioners can avoid dealing directly with the complexities of TensorFlow (or Theano or CNTK). Keras offers two main frameworks: the sequential API, which arranges layers in a linear sequence, and the functional API. The sequential model represents a straightforward stack of layers.

Experiment results

Referring to Table 1 and ‘Artificial neural network (ANN)’, our consolidated dataset consists of a total of 1,320 samples, each characterized by eight distinct features, as described earlier. The centralized machine learning architecture previously described is applied to this dataset, and the results achieved are presented in Fig. 16. In this scenario, the dataset is split into training and testing sets using a 70–30 percent ratio, with 70 percent allocated for training and the remaining 30 percent for testing.

Federated ML

In this proposed approach, the regular backhaul traffic remains undisturbed, employing a straightforward VPN architecture. A dedicated VPN is established over the conventional traffic, creating a distinct tunnel for transmitting and receiving data within the federated network. The implementation of the FL framework for a wireless backhaul network, utilizing VPN, has been successfully accomplished through the GNS3 virtual environment emulator. This emulator simplifies the tasks of designing, constructing, emulating, configuring, testing, and resolving issues in both virtual and actual networks, eliminating the necessity for physical hardware. The network controller is constructed based on the OpenDaylight (ODL) architecture, utilizing Open vSwitches (OvS) and layer three switches/routers to replicate the proposed topology in the simulation environment. The experiments, illustrated with a set of four clients, involve each client contributing its own HFMD biosensor data. To implement the federated learning strategy, the Flower framework (Beutel et al., 2022) has been installed and properly configured on both the client and server sides, executing the FedAvg, Q-FedAvg, and FedProx algorithms. The training process on the clients involves utilizing the HFMD dataset with neural networks, as extensively explained in Mahmood et al. (2022). In Fig. 17 the simulation setup is explained.