Analysis of Federated Learning Paradigm in Medical Domain: Taking COVID-19 as an Application Use Case

Abstract

Federated learning (FL) has emerged as one of the de-facto privacy-preserving paradigms that can effectively work with decentralized data sources (e.g., hospitals) without acquiring any private data. Recently, applications of FL have vastly expanded into multiple domains, particularly the medical domain, and FL is becoming one of the mainstream technologies of the near future. In this study, we provide insights into FL fundamental concepts (e.g., the difference from centralized learning, functions of clients and servers, workflows, and nature of data), architecture and applications in the general medical domain, synergies with emerging technologies, key challenges (medical domain), and potential research prospects. We discuss major taxonomies of the FL systems and enlist technical factors in the FL ecosystem that are the foundation of many adversarial attacks on these systems. We also highlight the promising applications of FL in the medical domain by taking the recent COVID-19 pandemic as an application use case. We highlight potential research and development trajectories to further enhance the persuasiveness of this emerging paradigm from the technical point of view. We aim to concisely present the progress of FL up to the present in the medical domain including COVID-19 and to suggest future research trajectories in this area.

1. Introduction

Federated learning (FL) is a promising paradigm that assists distributed clients to jointly train/learn a shared AI model (machine learning (ML) or deep learning (DL)) without disclosing personal data [1]. For example, multiple hospitals or research institutes can collaboratively train an AI model to enhance predictive accuracy for a particular problem/task without sharing the data on which the model is trained. FL has revolutionized the privacy-preserving domain with this unique concept (i.e., no acquisition of personal data). Specifically, FL focuses on bringing algorithms close to the data, rather than bringing the data to the algorithms. Before the advent of FL, the centralized learning (CL) domain always gathered personal data at a central server first, and machine learning algorithms were trained subsequently. However, the key problems in CL are more privacy breaches, a lack of guidance on managing privacy settings, and market monopolization by the tech giants. The concrete difference between FL and CL is provided in Figure 1. As shown in Figure 1, FL is a decentralized approach, meaning algorithms are brought close to the data, not data brought to the algorithms. It empowers users to preserve privacy effectively by not sharing any kind of data with a central server [2]. The application areas of FL are continuously expanding into many sectors, including energy management, supply chains, smart healthcare, biomedical apps, finance, and motion control.

Apart from other sectors, the use of FL in the medical domain is highly demanding as it can assist in developing new treatment procedures and clinical practices [3,4,5], helping our society combat deadly/rare diseases. Also, FL is the need of modern times because most hospitals are making the transition from general treatment procedures to personalized ones. Modern healthcare systems are generating a mammoth amount of data that can unlock many financial opportunities for diverse stakeholders (e.g., drug makers, insurance providers, doctors, etc.) in medical science [6]. However, the CL-based approaches cannot be employed to unlock the potential of medical data as there is a risk of data manipulation or misuse [7]. To this end, the FL is crucial as it can resolve one of the critical challenges of privacy preservation while at the same time developing many data-driven products utilizing diverse hospital data [8]. In the recent past, FL has also played a vital role in addressing many aspects (e.g., testing, drug discovery, severity estimation, etc.) concerning the COVID-19 pandemic. FL is reshaping the healthcare industry with many innovative use cases that were not possible in the past due to restricted data access and regulations.

Since FL is a privacy-by-design approach, it stays legally compliant because data do not leave the hospitals (or client devices). Furthermore, a powerful model can be trained collaboratively in the FL setting, and therefore, the data island problem can be solved effectively [9]. FL use in healthcare is highly beneficial because data often sit on an island, distributed massively, highly imbalanced, and non-IID (non-independent and identically distributed) in nature. Similarly, by applying FL in the COVID-19 era, detailed insights were obtained on this disease’s progression, spread, and dynamics, which can vary from site to site. Although FL has shown remarkable improvements in addressing the data island, privacy, and ML model performance issues, it can be subjected to various attacks due to its distributed nature [10]. Commonly executed attacks on FL environments are data and model poisoning, the backdoor attack, gradient inversion, and class/feature leakage [11]. Due to the distributed nature of FL, many such attacks can go undetected, and many efforts are underway to secure the FL pipeline from attack.

In this article, we provide a brief overview of the FL paradigm and its potential applications in the medical field including the COVID-19 pandemic. Then, we demonstrate the recent synergies/integration of the FL paradigm with other latest technologies (differential privacy, homomorphic encryption, blockchain, internet of things, ChatGPT, etc.) in order to either secure FL or extend the horizons of FL applications into multiple domains. We discuss advanced technologies originating from the FL and their benefits in general medical applications as well as COVID-19 that remained unexplored in the recent literature. We discuss major taxonomies of the FL systems and enlist technical factors in the FL ecosystem that are the foundation of many adversarial attacks on these systems. Finally, we discuss the key challenges in the FL paradigm under emerging technologies and in the COVID-19 era, such as performance on non-IID data and in data reconstruction. This extended knowledge can pave the way to clearly understanding the research status of FL in the medical domain and COVID-19, and to devising more secure methodologies in order to meet privacy and security expectations in the FL pipeline.

The rest of this perspective paper is organized as follows. Section 2 provides the fundamental concepts about the FL paradigm including major taxonomies of the FL system. Section 3 discusses the fusion of FL with the medical domain including the need for FL in the medical domain, architecture, applications, and challenges in the medical domain. Section 4 pinpoints and summarizes the potential applications of FL in the COVID-19 arena. Section 5 discusses the integration of FL with other digital technologies in the medical context. Section 6 discusses potential and hot research trajectories that require further research and developments. We sum up this perspective in Section 7.

2. Fundamental Concepts Related to FL

2.1. Fundamentals of FL

A conventional FL ecosystem encompasses N participants/clients, one central server, and a training protocol/algorithm. In each round, the global model, denoted with $\Delta W$, is assessed with respect to performance, refurbished, and distributed to all parties (e.g., servers and clients) encompassed in the ecosystem. The training procedure of FL keeps repeating until the convergence criterion is met. There are N rounds in the FL, and the analysis of one round of FL is explained in our recent work [12]. In the beginning, participants acquire a $\Delta W$ (global model update) from the centralized server. Subsequently, each client independently performs training on the local data and computes a local weight. Afterward, the local updates from each participant are forwarded to the central server for aggregation purposes. The function/aggregator, g, used at the central server for averaging at time/epoch t is in Equation (1):

$$g\left(t\right)=\frac{1}{N}\sum _{i=1}^N\Delta W_i^t$$

(1)

where N shows the # of clients, $g\left(t\right)$ shows the global weight at t epoch/time, and $\Delta W_i^t$ is the updated gradient at epoch t for the ith client.

2.2. Activities of Server and Clients in FL

In the FL system, servers and clients perform diverse functions to collaboratively train an AI model. The main functions executed by clients are: (i) acquiring parameters of $\Delta W$ from the server, (ii) model training with $\Delta W$ parameters and local data, and (iii) forwarding the gradients/param to the central server. In contrast, the functions performed by the server are: (i) sharing $\Delta W$ with all clients/participants, (ii) collecting all participants’ gradients in each round, (iii) computing the g (e.g., aggregated global model), (iv) updating model parameters, and (v) filtering malicious gradients/updates. It is worth noting that some of the activities may change depending on the application scenario or setting of FL systems. For example, in the poisoning attack detection scenario, some additional algorithms are executed on the client side to filter malicious local models.

2.3. Major Taxonomies in FL

On the client side, the data on which local models are usually trained can be either independent and identically distributed (IID) or non-IID. The major taxonomies of the FL system are demonstrated in Figure 2.

Referring to Figure 2, there are six different types of classifications of FL systems. The data-based classification of FL has three main types [13]. In the horizontal FL (HFL), the sample space is different but the feature space is the same. In contrast, vertical FL (VFL) has the same sample space but feature space is different [2]. In hybrid FL (also known as federated transfer learning (FTL)), both feature and sample spaces are different [14]. An example of all three classifications of FL (e.g., HFL, VFL, and hybrid/FTL) is given in Figure 3.

The communication architecture is either centralized, decentralized, and/or depending upon the application scenario [15,16]. The AI models used in the FL setting can be linear like linear regression, and more complex non-linear like convolutional neural networks. In some cases, hybrid AI models can also be employed to achieve faster convergence [17]. There are two main branches of the FL in terms of data and resources, which is called federation. To this end, FL systems can be categorized into three key settings. In a cross-silo setting, the size of participating devices is large, but the # of clients is small [18]. In contrast, the cross-device setting has small-sized devices, but the clients can be large [19]. In the hybrid setting, the device sizes and number of clients are balanced. The updates from the server can be shared in synchronous (server shares the global model after receiving updates from all clients) and asynchronous (the server does not wait for all clients in order to share the global model) ways [20]. Lastly, the communication from the local devices is either frequent vs. less frequent [21,22].

These taxonomies can pave the way to clearly understand the dimensions of FL. Apart from these taxonomies, some studies have classified the FL systems based on the privacy mechanism used as well [23]. The rest of this paper discusses the significance/use of FL in the general medical domain and COVID-19 context.

3. Fusion of FL and Medical Domain

This section discusses the fusion of FL and the medical domain from a technical perspective.

3.1. Need of FL in Medical Domain

With the proliferation of wearable devices and IoT-powered architecture in the medical domain, the amount of medical data is steadily increasing, which can offer many opportunities for innovation and knowledge extraction. However, the data owned by one hospital cannot be easily outsourced to another similar hospital/medical institute due to privacy and regulatory concerns. Moreover, each hospital can own diverse data, and training AI models over it cannot be generalized to other hospital data, leading to a data island problem. However, in some cases, the model training on diverse data owned by different hospitals is imperative when a new pandemic emerges or a rare disease occurs. However, sharing data for great good or training an AI model that can generalize to unseen data is a long-standing problem in the AI community. Thanks to the emergence of the FL, data exchange across hospitals is no longer required, and therefore, the privacy/regulatory requirements can be easily met whereas high-quality AI models can still be trained over scattered data. With the help of FL, high-quality AI models can be trained that exhibit higher generalization power and accuracy in real-world cases because models deployed in hospitals can benefit from the experience of other models. Furthermore, the population from one hospital to another is different, which can significantly contribute to a gain in the generalization power of the respective AI models. The use of FL in the medical domain can contribute to lower treatment costs, improved treatment, discovery of clinical practices, and lowering the burden of healthcare professionals.

3.2. Architecture of FL in Medical Domain

Figure 4 depicts the generic architecture of FL in the medical domain, where multiple hospitals/institutes jointly train the AI model for certain medical tasks.

The key steps in the overall process are task definition (e.g., image analysis or motion detection), client selection (hospitals or medical institutes), desired objectives (detection or classification), AI model selection, AI model parameters, etc. The model training process (local training, aggregation, and distribution of global models) is akin to the traditional FL process. Usually, the global model is trained on diverse data owned by many hospitals, and the training process stops when the desired accuracy is achieved, or loss converges. The trained AI model can be deployed in real settings for prediction/classification tasks. In FL-based medical applications, the AI model is exposed to diverse medical data, and therefore, the models are more generalizable and robust.

3.3. Applications of FL in Medical Domain

Recently, many promising applications of FL have been identified in the medical domain by utilizing data from diverse modalities (e.g., images, time series, signals, etc.) [24]. Furthermore, the recent FL developments have led to higher collaboration between the institutes that share common interests, and therefore the application stack of FL is swiftly expanding. Through detailed analysis of high-quality literature, we identified and organized the FL application in the medical domain into ten different categories as shown in Figure 5. To the best of the authors’ knowledge, none of the previous studies has grouped FL applications into such broad categories. The analysis given in Figure 5 can help understand the FL applications in the medical domain from a broader perspective. We believe the future applications of FL in the medical domain can likely fall into one of these categories.

3.4. Challenges of FL in Medical Domain

Although FL has many promising applications in the medical domain, at the same time, there are numerous challenges while applying this technology to real medical settings. These challenges can be associated with training data, clients (hospitals in medical cases), servers, aggregation algorithm (global model), network structure, training procedures, model and parameters, inference, convergence, social (privacy), and deployment. In this work, we focus mainly on data-related challenges as data are the cornerstone of AI developments. Figure 6 presents challenges/issues related to FL in medical and COVID-19. In Figure 6, we categorize the FL challenges/issues into two types: general medical domain (13 challenges) and COVID-19 (6 challenges). However, it is worth noting that both the “general medical domains” and the “COVID-19 scenario” mostly share the same issues/challenges when it comes to FL application, but the last six issues were commonly observed during the COVID-19 pandemic. For instance, the spatial data of diverse types were collected and processed with AI/FL models as a non-pharmaceutical intervention to curb the virus spread, which is unique to communicable diseases like COVID-19. In contrast, poisoned data and privacy issues are very common in both the “general medical domains” as well as the “COVID-19 scenario”. In conclusion, most data-related challenges are common between both whereas there exist some challenges that solely apply to either “general medical domains” or “COVID-19 scenario”.

In the general medical domain, data heterogeneity across hospitals is the most pressing challenge because it can lead to deficient model performance and prolong convergence time. Also, the data can be non-IID, which is a serious challenge in FL-based systems because it is hard to distinguish whether the low performance of the AI model is due to non-IID data or poisoned data. In some cases, the data can be poisoned via label flip, feature drop, and/or wrong labels, which can lead to wrong results in local model training and disparity in the global model. The data at different hospitals can be low in quality (outliers, missing data, low-resolution images, noisy signals, etc.), which can lead to poor convergence and low accuracy. In many medical applications, the data can be enclosed in different formats (e.g., different MRI manufacturers), posing a serious challenge to connecting dots or curating fused knowledge. In some cases, all data are not available at the training time as it can be pulled in real-time from wearable devices, so the local model can obtain different data in each phase, leading to inconsistent results. In some cases, it is hard to classify data based on disease as the characteristics of each disease are different, and analysis of diverse data about the same disease is very challenging. In some cases, data are not available in commodity for conducting disease analysis and obtaining more data is challenging due to privacy results.

In some cases, the data can overlap across hospitals, and training AI models over the same data can lead to a waste of computing resources. Data valuation is also very hard as the data quality valuation metrics can vary across hospitals, which may lead to higher inconsistencies in data and can contribute to information loss. Privacy preservation of medical data is challenging in the FL setting because the sensitive properties of data can still be leaked in model parameter exchange. The unavailability of benchmark datasets for some diseases hinders the proper evaluation of AI models. There are no procedures and mechanisms to identify the underperforming subpopulations to rectify them during the training process because the data sit in silos. Lastly, the data at each site can be biased toward some major populations/communities, which can degrade the local and global model performance.

The recent COVID-19 pandemic posed quite different data-related issues in the framework of COVID-19-oriented FL. For instance, data governance was one of the biggest challenges amid the pandemic, which hindered the application of AI and FL, owing to data misuse and privacy concerns. In some countries, the work on data governance, particularly personal data handling and privacy laws started after the pandemic [25]. Furthermore, the urgency of the situation and mutation leads to higher concept and data drifts in the FL models, and therefore, the performance of most models was fragile. Cross-silo FL, which involves multiple hospitals, is a widely used architecture in the medical domain. However, the data disparity among these hospitals may need a higher number of iterations to accomplish the desired accuracy on a global server [26]. This increased computation time poses challenges in maintaining responsiveness and efficiency in medical applications. The aggressive use of digital technologies in the pandemic era led to an explosion in clinical (and general) data, and therefore, it was very challenging to apply FL to extract useful knowledge. Furthermore, the existence of less diverse data of one entity (1:M) across multiple hospitals posed a serious challenge to generalizability gain in FL.

The diverse medical data such as laboratory measurements, clinical data, omics, etc., and its integration with existing knowledge bases posed a serious challenge in terms of data truthfulness/fairness. The limited availability of medical data concerning the pandemic and the lack of methods to evaluate the fairness/truthfulness, also known as the non-IID nature of data, hindered the FL-based model development. This was already a major obstacle in the general medical domain and biomedical research, but a pandemic of a poorly understood novel disease that overburdened hospitals’ capacity has divulged the importance of this problem [27]. Also, the incomplete and redundant data that are heavily biased towards fewer aspects (e.g., mostly negative test data) of the pandemic or certain populations led to poor convergence of the FL-based model or imbalanced learning issues. To curb the spread of the pandemic, both clinical and non-clinical (location, mobility, etc.) data were supposed to be processed with FL models. However, it was very challenging to simultaneously process both these kinds of data with FL to either restrain the spread of the virus or determine clinical complexities in patients. The solution to the above challenges is urgent and vital to unlocking the hidden potentials of FL in the medical domain.

4. Potential Applications of FL Technology in the COVID-19 Era

In this section, we present some promising applications of FL that have helped societies across the globe to lower the effects of COVID-19 via technology. Some promising applications of FL are virus detection from image/X-ray data, diagnosis based on heterogeneous source data fusion, and outcome prediction [28]. Furthermore, FL has also played a key role in understanding the dynamics of COVID-19 via institute information/model sharing of COVID-19 patient data with privacy preservation. Although work has been published related to COVID-19 detection, AI usage in the COVID-19 context, and generic AI applications, a holistic picture of FL applications from a broader perspective, has not been discussed in the published literature. In Figure 7, we cover this gap by presenting recently proposed/developed promising applications of FL in detail.

Because the future of healthcare can benefit from FL to a great extent, many FL-related projects are underway across the globe to determine its potential for future endeavors. A remarkable initiative in this regard is the Trustworthy Federated Data Analytics (TFDA) https://tfda.hmsp.center/ (accessed on 29 January 2024) project. TFDA adapts the core concept of FL (algorithms → data) in a regulatory-compliant and trustworthy way, instead of using a data-centric approach (data → algorithms). The TFDA project will be validated in medical research from both theoretical and technical perspectives before utilization on a wider scale. The Joint Imaging Platform (JIP) https://jip.dktk.dkfz.de/jiphomepage/ (accessed on 31 January 2024) is another promising initiative of the German Cancer Consortium (DKTK), fostering image-based analysis without disclosing actual image data. This platform assists in performing image-related tasks with an easy-to-use toolkit developed solely based on the FL paradigm. In line with these works, Federated Tumor Segmentation (FeTS) https://www.med.upenn.edu/cbica/fets/ (accessed on 1 February 2024) is a third promising initiative for detecting a tumor boundary by leveraging diverse patient data without accessing private data. All these developments indicate the ever-increasing interest of researchers in the FL domain, especially in medical environments. Similarly, other developments https://cordis.europa.eu/project/id/826078/reporting (accessed on 3 February 2024), https://owkin.com/connect (accessed on 5 February 2024), are underway concerning COVID-19 in an attempt to eradicate this deadly pandemic from societies as soon as possible. The details of each FL application can be learned by searching for relevant studies using application names. The analysis demonstrated in Figure 7 can pave the way to learning the potential of FL in the medical domain, especially in the context of COVID-19. Apart from the main applications, in some cases, FL has been utilized to improve critical parts of ML algorithms. For example, some studies have demonstrated FL use as pre-processing and on-demand data provisioning in ML environments.

5. FL Synergy with Other Digital Technologies in the Medical Context

In this section, to attain multiple goals, we discuss the established integration of FL with other technologies in the context of medical aspects. For example, FL works in a distributed manner and is vulnerable to different kinds of privacy attacks (e.g., parameter disclosure, data/gradient leakage, and update theft). To resist these attacks, differential privacy is increasingly used with FL to safeguard privacy issues [29]. In the DP model, algorithm $\mathcal{A}$ satisfies $ϵ$-DP if for all subsets $X\subseteq Range\left(\mathcal{A}\right)$ and all D and $D^{\prime }$, $d(D,D^{\prime })=1$ (e.g., D differs from $D^{\prime }$ by just one record):

$$\frac{Pr\left(\mathcal{A}\right(D)\in X)}{Pr(\mathcal{A}\left(D^{\prime }\right)\in X)}\le exp\left(ϵ\right)$$

(2)

where $ϵ$ denotes the privacy loss budget, and its value is usually higher than 0 (i.e., $ϵ>0$).

To improve utility, ($ϵ$, $\delta$)-DP has also been widely integrated with FL to improve both privacy and utility [30]. There are two main types of DP: local DP and global DP. In local DP, noise is added to each data item, $x_i$, before sharing it with the server, because the server is considered untrustworthy. In contrast, noise is added to $x_i$ at the server in global DP, and the server is considered trustworthy. Apart from DP, FL has been amalgamated with many other technologies to improve its technical deficiencies. In

Table 1. Overview of FL integration with digital techniques/technologies from medical context.

Paradigm	Synergy with	Purpose of Synergy
FL	DP	Privacy protection of sensitive medical data
	Local DP	Privacy protection of FL parameters/gradients
	HE	Privacy preservation of data/parameter in transfer
	Blockchain	Privacy infrastructure to share COVID-19 information [35], resource allocation, self-testing, EWS, poisoning detection [36], etc.
	Edge CC	Secure analytics of COVID-19 image data
	Multilayer perceptron	Robust mortality predictive models
	Support vector machine (SVM)	Prediction of daily COVID-19 cases
	IoT	COVID-19 screening, medical data acquisition and efficient analysis [37]
	IoT big data	Fast detection of virus, Enhancing data quality [38]
	5G-enabled architecture	Classifying severity of COVID-19
	IIoT	Accurate detection of COVID-19, personalized healthcare [39]
	RL	Identifying COVID-19 from medical images, image segmentation [40], decision making [41]
	NAS	Face mask detection
	CC	Diagnosis [42], analytics of medical data [43], disease monitoring [44], QoS enhancement of FL-based systems [45]
	IoMT	Medical DSS for tracking COVID-19, collaboration among medical institutes [46]
	Transfer learning	Classifying COVID-19 from lung scans, breast cancer classification [47]
	Access control	To safeguard medical data in BDE
	Social IoT	Robust contact tracing, on-board training of medical devices [48]
	SMC	Construction of virus vulnerability map, data protection in dynamic scenarios [49]
	Watermarking	Privacy protection of gradient information, privacy-preserved data sharing [50]
	Knowledge distillation	Identifying normality, COVID-19, and pneumonia from X-rays, medical image segmentation [51]
	ZKPs	Mobile healthcare, sensitive data protection [52]
	Split learning	Collaborative healthcare analytics [53]
	Fog computing	Performance enhancement of medical devices [54]
	ChatGPT	Knowledge enhancement and better QoS [55,56]

Abbreviations: DP = differential privacy, HE = homomorphic encryption, CC = cloud computing, QoS = quality of service, EWS = early warning system, IoT = Internet of things, RL = reinforcement learning, IIoT = industrial Internet of things, NAS = neural architecture search, IoMT = Internet of medical things, DSS = decision support system, BDE = big data environments, SMC = secure multiparty computation, ZKPs = zero-knowledge proofs.

, we demonstrate the synergy of FL with different technologies in the context of the medical domain. These synergies are imperative to improving FL in different contexts, as well as to controlling the pandemic’s impact on the general public. As shown in Table 1, synergies between FL and other technologies benefit each other from many technical aspects. The synergy between FL and other emerging technologies is likely to grow in the coming years. In addition, due to greater abilities in working with images and video data, FL will be a mainstream technology in future data-driven applications. It is worth noting that FL has integration with more than one technology also to effectively meet the security and privacy requirements [31,32,33,34]. In some cases, three different techniques have been integrated with the FL paradigm for strong privacy preservation [31]. In the future, integration with multiple technologies will be imperative in order to accomplish multiple tasks or to overcome FL performance issues.

In the past few years, the FL concept has significantly evolved. In Figure 8, we demonstrate the evolution of FL from local learning to federated analytics. The general AI concept started with local learning (LL). In LL, data and computations are usually in isolated sites, as seen in Figure 8a. Later, LL evolved into central learning where data moved to central locations to perform computations (also known as cloud-based ML). See Figure 8b. Later, CL evolved into FL, in which data are kept with their contributors, and computing is performed on the client side, but parameter/updates are orchestrated by the central server.

See Figure 8c. Later, FL evolved into swarm learning (SL) in which there is no need for a central server because both data and parameters stay at the edge, as seen in Figure 8d. SL is one of the technologies resulting from FL.

In 2017, Google introduced the concept of federated analytics (FA). In FA, the analytics can be performed on local devices with local data akin to FL. See Figure 8e. The FA concept was derived from the FL and belongs to the category of distributed computing paradigm [57]. Soon, FA will become one of the key technologies in the domain of collaborative learning. In the COVID-19 era, FA has been used for conducting analytics of COVID-19’s impact and other diseases, analysis of vaccine efficacy on different subgroups, the categorization of COVID-19’s effects based on demographics, for identification of risk factors related to COVID-19, forecasting vulnerability indexes, mortality/case predictions, forecasting disease trends, laboratory data analytics (cough samples, breathing patterns, oxygen levels, infected regions of lungs, etc.), fMRI analysis, telemedicine, vulnerable group identification, candidate vaccine developments, and protein sequence analysis [58,59,60,61].

6. Future Research Trajectories for FL Ecosystem

Since FL’s inception, a substantial number of technical deficiencies in the FL ecosystem have been observed from multiple perspectives, and many developments are underway to enhance the persuasiveness of the FL ecosystem. For example, in the FL ecosystem, attackers can compromise clients’ data or updates, corrupt the server-side global model, infer/reconstruct training data, slow the convergence speed, degrade the accuracy of global/local models, and/or predict sensitive information about clients [62]. In COVID-19 scenarios, the FL ecosystem can be trapped into slow convergence as well as lower prediction differences due to greater differences in data distributions/imbalances at each site [63]. Furthermore, deployment of the FL ecosystem for COVID-19 classification/prediction tasks can be subjected to adversarial attacks and data breaches. Hence, securing the FL ecosystem from a variety of adversarial attacks is an important concern in FL ecosystems [23]. In Figure 9, we present a detailed overview of the technical factors in the FL ecosystem that are the foundation of many adversarial attacks on these systems. The detailed analysis in Figure 9 can pave the way to improving the FL ecosystem from a technical perspective. In future endeavors, efficient solutions for all the challenges in Figure 9 are imperative in order to truly benefit from the potential of FL. Apart from the challenges in Figure 9, below, we pinpointed eight main future research aspects that need urgent solutions in the context of general FL and general medical domain including COVID-19.

• Robust solutions for statistical heterogeneity: In the FL, the data used in local model training can be highly diverse at each site (e.g., different languages used in next-word-prediction tasks) leading to poor convergence of the global model [64]. Since the data mostly violate the IID assumption, the accuracy of the global model minimally improves in each round. Although some strategies (data augmentation, data sharing, etc.) for addressing the non-IID problem in the FL setting have been proposed [65], this topic still needs rigorous work to eliminate all limitations stemming from all categories (i.e., attributes, labels, distributions, and temporal skews) of non-IID data.
• Robust strategies for client and server behavior analysis: In the FL paradigm, it is quite challenging to distinguish between benign and abnormal clients/servers because, in some cases, abnormal clients/servers can behave like real clients/servers, and vice versa. However, clients and servers are two of the most critical components of the FL paradigm, so analysis of their behavior to ensure reliable results and fairness is imperative in the FL setting. To this end, integration of the latest technologies, such as blockchain, RL, and anomaly-detection algorithms with FL, is handy for restricting manipulations of data or parameters, tampering with the training phase, and the possibility of launching collaboration attacks. In the future, it will be interesting to devise robust strategies for client/server behavior analysis to ensure the smooth operation of FL.
• Credible incentive mechanisms for good data contributors: In the FL setting, most clients can have a free ride and can leave the FL system at any time, leading to poor convergence of the global model. To ensure smooth mechanisms, incentive mechanisms for clients have been introduced recently in the FL paradigm. However, the selection of a pool of suitable clients that can contribute good data and stay active throughout the model training process is tricky. To this end, credible incentive mechanisms for selecting good data contributors (and performance analyses) are needed. Furthermore, devising multicriteria-based incentive mechanisms (e.g., activeness and availability of clients, data quality, resource availability, motivation) is a vibrant area of research.
• Intelligent methods for inference-time vulnerability mitigation: After multiple rounds of training, the trained model can be deployed to serve people. However, there are plenty of interference-time vulnerabilities that stem from trained model deployment. For example, the model can yield low accuracy on different versions of the test data carefully crafted by an attacker. The credibility and fairness of the final results produced by the FL model can be low due to the higher diversity in data. Furthermore, scalability and verifiability are two main challenges in this context. To alleviate such issues, intelligent methods are required to lower interference-time vulnerabilities and make FL models trustworthy.
• Practical methods for enhancing security and privacy in FL ecosystems: FL ecosystems are constantly targeted by active/passive attackers to impair model performance [66,67]. Attacks that corrupt/damage global model performance by manipulating either model updates or samples of training data are referred to as poisoning attacks (PAs) [68]. PAs can be classified into two types:

  –

  Data PA: integrity breach in training data → global model corruption.

  –

  Model PA: manipulation of model updates (i.e., training procedure → global model corruption).

Both PAs can cause severe damage to global model performance (i.e., wrong classifications/predictions) and are difficult to detect. We present both these attacks with practical examples in Figure 10.

Some developments, such as federated distillation, moving target defense, trusted execution environments, DP-powered FL for parameter/update security, anomaly detection, pruning, ZKPs, adversarial training, legitimate client recognition, clipping gradients, federated multi-task learning, and SMC, have recently emerged to tackle these PAs and their variants. Although these approaches are handy, certain limitations exist regarding emerging adversarial threats [69]. Hence, proposing practical methods to enhance security and privacy in the FL ecosystem from major adversarial attacks without compromising accuracy is an interesting area of research.

• Simultaneous optimization of multiple types of trade-offs in FL systems: There exist multiple types of trade-offs in FL systems such as privacy-accuracy trade-offs, privacy-poising trade-offs, privacy-convergence trade-offs, and privacy-fairness trade-off [70,71]. Recently, researchers have resolved more than two types of trade-offs in FL environments. Wei et al. [72] explored the solution for three different types (e.g., robustness, privacy, and governance) of trade-offs in distributed AI systems. Zhang et al. [73] explored ways to solve three types (security, robustness, and privacy) of trade-offs to make FL more trustworthy. Kang et al. [74] explored ways to resolve privacy, utility, and efficiency objectives in FL environments. To this end, more methods are required that can effectively resolve the trade-offs of different kinds without degrading performance.
• Adoption of data-centric AI approaches to address data quality problems: The application of data-centric approaches https://landing.ai/data-centric-ai/ (accessed on 10 February 2024) to enhance data quality within the framework of COVID-19-oriented FL is very challenging due to the higher diversity in the amount and type of data across hospitals. In the COVID-19 era, there was a huge skew in the data across hospitals such as feature skew, labels skew, quality skew, and quantity skew. However, the solution for quantity skew poses more challenges in server-side aggregation as the local models trained with long-tail distributed data cannot fairly contribute to the global model’s performance [75]. In the conventional FL setting, the data are assumed to be fixed; however, the COVID-19 scenario was different, and data were originating in stream form, posing many challenges in terms of cleaning, transformation, integration, and reduction. It was very challenging to achieve convergence in a reasonable time, particularly when non-IID data and the straggler effect (e.g., some clients have poor computing resources) combined in COVID-19 handling systems [76]. It was very challenging to distinguish between the non-IID COVID-19 data and poisoned COVID-19 data as the effect of both these cases is roughly the same on FL model performance, and there is no way to analyze the local clients’ data. The end-to-end pipeline development (e.g., IoMT) under the FL paradigm posed many technical challenges concerning resource management and equitable use. The selection of suitable AI models for handling diverse aspects (e.g., screening, prediction, projections, diagnosis, classification, etc.) related to the pandemic was also very challenging. Exploring promising solutions for the above-cited pandemic data-related challenges is an interesting area of research.
• Harnessing the potentials of quantum computing to improve FL-based systems: Of late, a new paradigm, namely, quantum FL (QFL) has emerged [77,78], and its potential in the medical domain is yet to be explored. Therefore, it is also one of the promising research areas to be investigated in the context of the medical domain to address performance bottlenecks in the conventional FL paradigm. Lastly, high communication bandwidth, a large number of devices to participate in the training process to achieve a good model, and scalability issues also require low-cost solutions.

In the realm of generative AI, almost all fields are expecting paradigm shifts, and a similar shift is expected in the FL systems. Some studies have readily explored ways of linking generative AI with the FL [79]. Therefore, it is vital to investigate avenues that can amalgamate these two latest technologies and improve the technical aspects (or QoS) of the FL paradigm. Lastly, FL may yield biased results in the training process, owing to the higher disparities in data among participating clients [80]. Hence, it is imperative to develop practical methods to ensure fairness and to prevent discrimination against any client during the training.

7. Concluding Remarks and Future Work

This article presented an overview of the FL paradigm in the context of the general medical domain and COVID-19. Specifically, we highlighted the need for FL in the medical domain, the architecture of FL in the medical domain, promising applications of FL in the medical domain including COVID-19, the integration of FL with emerging technologies to improve various service scenarios of FL and its privacy, and the technical challenges that practitioners face in FL design/deployment in the medical domain. We demonstrated the evolution of the AI domain (i.e., local learning → swarm learning), and the resulting technologies of FL (federated analytics and swarm learning). We discuss the major taxonomies of the FL systems in six aspects and enlist technical challenges in the FL ecosystem that are the foundation of many adversarial attacks on these systems. Finally, some key research areas that need further exploration from the AI community are suggested to increase the reliability of FL in the medical domain. This study’s findings can pave the way to discern the progress FL has made so far in the medical field, and the practical issues stemming from FL application in this line of work. In recent years, FL systems have widely contributed to training AI models of higher generalizability, which can assist doctors in many ways such as disease classification, disease prediction, prescribing medications, and severity assessment of various kinds of diseases. The FL has abilities to produce robust AI models by using diverse and large-scale datasets held by many hospitals (or medical institutes), and such AI models are in dire need in the medical domain considering the rapid rise in clinical complications of many diseases. Since the medical/healthcare domain belongs to safety-critical applications, and therefore, FL has huge prospects in it in terms of preventing wrong medication, prescriptions, and treatment procedures compared to CL by exposing underlying AI models to diverse datasets. FL can widely contribute to early and accurate diagnosis of some critical diseases such as Alzheimer’s, autism, lung cancer, etc. In addition, FL can foster collaboration across medical institutes while alleviating the stringent barrier of privacy which was not possible in the recent past (e.g., before the year 2016). It can also contribute to processing data of diverse modalities and identifying the critical parts in medical images, leading to faster and more reliable diagnosis with significantly reduced cost. Based on the above analysis, it is fair to say that FL has revolutionized the medical domain and it can contribute to developing data-driven healthcare products/solutions. In the recent pandemic, FL has also contributed to sharing medical knowledge across the institute to better plan treatment procedures. It is worth noting that FL is not a silver bullet, and there exist many technical issues in it when it comes to real-world deployments. For example, due to its distributed nature, it is prone to adversarial attacks of diverse types (e.g., backdoor, data poisoning, model poisoning, etc.). It cannot prevent privacy risks of all types, and personal information or data characteristics are still leaked when clients communicate with the server. Also, the convergence of the global model takes a long time, when either most clients do not share their updates or data at some clients is non-IID. Hence, more efforts from different stakeholders are needed to overcome these issues and to make this technology beneficial for people around the globe. In future work, we intend to explore the role/potential of two emerging technologies (i.e., federated analytics and swarm learning) that resulted from FL in the medical context. We also plan to investigate the applications and challenges of serverless FL in the financial sector, which is a promising research topic in the AI-driven era.