A data sharing method for remote medical system based on federated distillation learning and consortium blockchain

Abstract

With the development of Medical Internet of Things (MIoT) technology and the global COVID-19 pandemic, hospitals gain access to patients’ health data from remote wearable medical equipment. Federated learning (FL) addresses the difficulty of sharing data in remote medical systems. However, some key issues and challenges persist, such as heterogeneous health data stored in hospitals, which leads to high communication cost and low model accuracy. There are many approaches of federated distillation (FD) methods used to solve these problems, but FD is very vulnerable to poisoning attacks and requires a centralised server for aggregation, which is prone to single-node failure. To tackle this issue, we combine FD and blockchain to solve data sharing in remote medical system called FedRMD. FedRMD use reputation incentive to defend against poisoning attacks and store reputation values and soft labels of FD in Hyperledger Fabric. Experimenting on COVID-19 radiography and COVID-Chestxray datasets shows our method can reduce communication cost, and the performance is higher than FedAvg, FedDF, and FedGen. In addition, the reputation incentive can reduce the impact of poisoning attacks.

KEYWORDS:

• Remote medical equipment
• consortium blockchain
• federated learning
• knowledge distillation
• reputation incentive

1. Introduction

With the ongoing COVID-19 pandemic transmission, people's health is being seriously threatened (Rahman et al., 2020). Traditional medical technology requires patients to go to the hospital in person. With the development of the remote medical system, it can reduce patients’ commuting time to the hospital and improve doctors’ diagnostic efficiency (Bashshur et al., 2020). While benefiting from remote medical system infuses new vitality into clinical scientific research, epidemic prevention and public health management, it also brings privacy leakage. According to the OCR HHS, the number of victims of medical data leakage increased over 1.5 times compared to 2020 (Maria, 2022). Due to policy and legal restrictions and the awareness of patients to protect their privacy and security, it is difficult to share data from the hospital's local server with other institutions. Lack of sufficient data samples in hospital servers for machine learning (Zheng et al., 2022), which results in isolated data islands. Therefore, there is an urgent need for a secure method of data sharing in remote medical systems.

Fortunately, Federated learning (FL) is a distributed computing technique (McMahan et al., 2017), that solves isolated data islands. Traditional machine learning transmits collected medical data to a cloud server for training. FL reduces the risk of data leakage by deploying the model on a local device and transferring the model training parameters. Instead of patients’ medical data directly to a cloud server with high computing power. This way guarantees the privacy of the user's data and solves the data isolations problem.

However, the latest research shows that there are still some critical issues for FL. For example, first, Melis et al. (2019) proof gradient updates may reveal important information about customer training. Attackers can recover data from local servers uploaded gradients (Hitaj et al., 2017). In addition, the FL training process contains millions of parameters because of the large CNN model. Transmitting these massive model parameters will produce a vast communication cost and increase the risk of malicious attacks during the model training process. Medical data is often heterogeneous in the remote medical system. Second, Yang and Sun (2022) believe that only interacting common information from training results on different devices can improve local personalised training results more effectively.

In order to solve the problem of medical data heterogeneity in the remote medical system and the high communication cost of FL. Knowledge distillation combined with FL provides a new solution (Hard et al., 2018). Wang et al. (2020) proposed a knowledge distillation method for object recognition. They achieved this by deploying the model on the cloud and fog nodes and the online bidirectional model heterogeneous training on the distributed data set. Oh et al. (2020) proposed a distributed machine learning framework with high communication efficiency and privacy protection to solve the problem of asymmetric uplink and downlink capacity. The output of the local model using the FD algorithm is uploaded to the server in the uplink, the global model parameters of FL are downloaded to the downlink. Sattler et al. (2021) used auxiliary data for unsupervised pre-training and deterministic weighted integrated distillation. The differential privacy method is used to make a weighted ensemble prediction for the auxiliary data according to the determinism of each client model. Kumar et al. (2021) proposed a secure framework for FL in a hostile environment. Knowledge distillation and model inversion methods to ensure security and privacy features and reduce communication cost.

Although many excellent solutions to solve the problem of heterogeneous FL models and communication overhead, we have noticed that they can't control the quality of data transmitted by participants in the process of FL training. The attacker will upload low-quality data and affect the performance of the global model. Thus it is affected by poisoning attacks. And it relies on a centralised server aggregation model, but the central server is prone to failure (Fung et al., 2018). Blockchain technology allows model aggregation and task training to be securely distributed. Blockchain is defined as a peer-to-peer distributed ledger for recording transactions, stored in a chain of connected blocks (Crosby et al., 2016). The development of blockchain technology provides FL with a decentralised global model convergence environment in which all model updates are verified. It is stored in the blockchain in a distributed manner through the blockchain consensus mechanism. Smart contract calls are used to select nodes in FL, thus controlling the uploaded data to be of high quality. Hyperledger Fabric as a representative of the consortium blockchain. It is an open-source, enterprise-grade implementation platform maintained by IBM and the Linux Foundation (Androulaki et al., 2018). Unlike Bitcoin and Ethereum, Hyperledger Fabric does not have any cryptocurrency.

In this paper, we propose a data sharing method in remote medical system called FedRMD based on FD and consortium blockchain, respectively. To summarise, our major contributions are as follows.

• We propose an FD approach called FedRMD to solve the problem of patient data heterogeneity in remote medical systems. While traditional FL transmits model parameters, our approach transmits the model output logit, thus greatly reducing the communication cost.

• We propose a reputation incentive based consortium blockchain to improve the quality of data in FL. Blockchain provides a decentralised global model convergence environment, and store the global model of the task publishers and the reputation values of the data providers.

• We experiment with FedRMD on COVID-19 Radiography and COVID-Chestxray datasets. Extensive experiments show that our algorithm achieves higher accuracy compared with the FedAvg, FedDF, FedGen. And our method can resist the influence of poisoning attack in FL model training.

The rest of the paper as follows: Section 2 details the current state of research federated distillation learning and blockchain-based data sharing. Section 3 introduces problem definition. Section 4 elaborates our proposed FedRMD algorithm, and the reputation incentive of blockchain. Section 5 presents the experimental setting of federated distillation learning and blockchain. Section 6 analysis the results of our experiments. Section 7 summarises our research and future work.

2. Related work

2.1. Knowledge distillation in federated learning

Knowledge distillation in FL has recently emerged as an effective method for addressing user heterogeneity. We have conducted extensive research on the current area of combining FL with distillation learning. Li and Wang (2019) proposed FedMD using transfer learning and knowledge distillation. This framework allows clients to design different network structures based on their computational power. It protects the privacy security of the dataset and the model under the condition of jointly training a model. However, it requires less computation than centralised FL. Chang et al. (2019) in Cronus, each client uses the local and soft-label public datasets jointly for local training. However, the model he used is large, and the high-dimensional parameter vectors are vulnerable to privacy and security attacks. Li et al. (2019) designed a personalised distillation environment for each client rather than solving a global model. Lin et al. (2020) proposed FedDF training models. This knowledge distillation method combined with FL reduces privacy risks and cost and allows flexible aggregation of heterogeneous client models without many restrictions on the model's university and data structure. Zhu et al. (2021) proposed a data-free knowledge distillation method to address heterogeneous FL, called FedGen, in which a lightweight generator was learned to integrate user information in a data-free manner. The server uses local training to coordinate and broadcast to users. Cheng et al. (2021) proposed a framework that can prevent negative and malicious knowledge transfer. This framework further improves the robustness of FL against poisoning attacks. It reduces the cost of communication between the server and the client, but this framework has some limitations, such as the dependency on marking as public datasets. Gong et al. (2021) proposed a single-issue FL framework, called FedAD. They theoretically preserve the efficiency of using only untagged and domain name public data and efficiently available network bandwidth resources, demonstrating the applicability of FedAD to real-world cross-institutional learning through medical image data. Nguyen et al. (2022) proposed a distributed FL approach that uses knowledge distillation to ensure data privacy and protection, with each node operating independently and without access to external data. Xing et al. (2022) proposed a multitask time series classification system for multitasking using the FD. The scheme matches each user in the system with a partner. In order to their weights are the closest of all uploaded weights. However, the framework is still composed of a central server and multiple mobile users, which is vulnerable to the failure of the central server and can cause the system to crash. Han et al. (2022) proposed an unsupervised FL framework that does not require clients to share data features. He used knowledge distillation methods to learn bi-directionally from dispersed and heterogeneous data. They also proved that the framework could enhance the current unsupervised FL algorithm. He et al. (2022) proposed a selective self-distillation method for FL. It imposes adaptive constraints on local updates by self-distilling knowledge of the global model and selectively weighting it by evaluating class and sample level confidence.

2.2. Blockchain-based data sharing

Nguyen et al. (2019) proposed a medical data-sharing solution using mobile cloud computing and blockchain. The data is collected by interacting with the medical sharing system through mobile devices, storing the patient's data address in the blockchain and the data set in the local server. They also implement an access control mechanism based on medical personnel's public key and data address to achieve decentralised data storage and sharing. Lu et al. (2019) employ blockchain to form a data-sharing platform that benefits data privacy and security. On this basis, Lu et al. (2020) utilise a shard-based blockchain protocol for preserving FL's convergence when large-scale nodes are involved. Zhou et al. (2020) proposed a decentralised multi-community training framework, which utilises blockchain to maintain a global model within each community. Communication between the community and other newer models follows the full reduction protocol. El Rifai et al. (2020) integrated FL and blockchain in a healthcare environment to use diabetes data sets to predict decision support tools. They proposed a smart contract to achieve transparency and immutability in information sharing. Tan et al. (2021) proposed a blockchain to control patients’ COVID-19 medical records. When the user meets the attribute access policy of the blockchain, the key can be obtained through the cloud to decrypt the information and control malicious users through tracking lists. Kumar et al. (2021) proposed a FL model based on blockchain by invoking smart contracts to upload the weights of local training models to the blockchain network and train a deep learning model on the decentralised network to obtain the latest information of COVID-19 patients. Egala et al. (2021) proposed a blockchain-based electronic health record through smart contracts to provide convenient services. However, the data is stored in an on-chain database, which makes the system unstable and less scalable. Zheng et al. (2022) proposed a mutual authentication and key agreement scheme based on the Internet of Vehicles and blockchain. Storing the vehicle's authentication information on the blockchain reduces the time for cross-domain authentication of the vehicle and uses lightweight computing to protect the user's private information. Yao et al. (2022) proposed a multi-dimension traceable privacy-preserving health code scheme based on blockchain. It prevents health code information from being tampered with and supports the traceability of the virus transmission chain. Qu et al. (2022) designed a quantum electronic medical record system and proposed a blockchain network based on quantum cryptography. The use of quantum authentication systems and the automatic formation of time stamps by linking quantum blocks to controlled activities reduce the amount of storage space required.

Finally, Table 1 gives a comparison of related work with FedRMD. The results show that our method is designed for COVID-19 and efficient communication and defends against poisoning attacks and single node failures.

Table 1. Comparison of related work with FedRMD.

Methods	Efficient communication	Decentralised training	Defended poison attacks	Designed for COVID-19
Nguyen et al. (2022)	✓	✓	✗	✗
Cheng et al. (2021)	✓	✗	✓	✗
Tan et al. (2021)	✓	✗	✗	✓
Kumar et al. (2021)	✗	✓	✓	✓
Zhou et al. (2020)	✗	✓	✓	✗
Gong et al. (2021)	✓	✗	✗	✓
Lu et al. (2020)	✓	✓	✗	✗
FedRMD	✓	✓	✓	✓

3. Problem definition

Figure 1 shows the framework of FL in remote medical system. First, remote medical equipment transmit the encryption medical data to the local hospital server. We assume a set of $N$ hospitals, which stores a local dataset denoted as $\{{\mathcal{D}}_n\}n\in \{1,\dots ,N\}$. The local hospitals n experience loss over ${\mathcal{D}}_n$ is defined as: (1) ${\mathcal{L}}_n\left(\theta \right)={\displaystyle \frac{1}{\left\{{\mathcal{D}}_n\right\}}\sum \left\{{\mathit{x}}_n,y_n\right\}\in {\mathcal{D}}_{n}L\left(\theta ,x_n,y_n\right)}$(1) where $L\left(\theta ,x_n,y_n\right)$ is the loss function of the training medical data sample $x_n$ and $y_n$ be the label. $\theta$ be the model parameter in the global server and local hospitals. FL can be assumed to be the centralised data ${\mathcal{D}}_c$ distributed to $\text{N}$ hospitals, expressed as $\{{\mathcal{D}}_c\}=\cup \{{\mathcal{D}}_n{\}}_{n=1}^N$. Generally speaking, FL enables multiple hospitals to train a global model G without sharing their data, and through sharing of their training parameters θ. The aim of FL can be formulated as the following problem: (2) $\underset{\theta }{\min }{\displaystyle \frac{1}{N}\sum _{n=1}^N{\mathcal{L}}_n\left(\theta \right),{\mathcal{L}}_n\left(\theta \right)={\displaystyle \frac{1}{\{{\mathcal{D}}_n\}}\sum _{n=1}^{\{{\mathcal{D}}_n\}}L\left(\theta ,x_n,y_n\right)}}$(2)

However, in the remote medical system, the distribution of medical data stored in various hospital institutions differs. After executing SGD on each hospital server, the collected parameters are aggregated in the global server. However, the aggregation method cannot solve the problem of patient data heterogeneity. In addition, the quality of parameters uploaded by participants can't be guaranteed in the process of FL training, and it requires a centralised server for global aggregation. Hence, it is easy to crash the whole system due to a single node failure. Therefore, we propose our approach FedRMD combining FD and blockchain to solve the above problems.

Figure 1. Framework of FL in remote medical system.

4. Method

4.1. Overview

Our remote medical system comprises block nodes, task publishers for FL, and medical data providers. Task publishers such as hospital doctors and medical institutions release FL tasks according to their needs to find data providers and select appropriate data providers to participate in FL training by invoking smart contracts. Data providers, mainly medical institutions and hospitals, use remote medical equipment to supervise patient health and collect patient medical data. After that, they use local servers to train the FL model and transmit soft labels to the blockchain nodes. The blockchain of peer nodes is mainly responsible for accepting the FL tasks of the task publisher and the soft targets uploaded by the data provider's server. It also provides a robust aggregation environment for aggregating FL models and updating reputation values. Our proposed training process in one communication round is shown in Figure 2.

Figure 2. Overview of proposed method.

We provide a reliable data option scheme for FL tasks. Blockchain ensures that the data uploaded by the hospital can be instantly traceable and has greater transparency. We can use smart contracts to control the quality of the uploaded data. It consists of six steps. As shown in Figure 2, a clear training flow can be represented by the following steps:

Step 1: Publish FL task. Task publishers set reputation thresholds based on their requirements and invoke smart contracts to publish the FL task after consensus certification to produce the genesis block, which includes the initial global model, communication rounds, and reputation threshold.

Step 2: Select data providers. The reputation value of each hospital is stored in the blockchain. If the hospital meets the task publisher's reputation threshold, it will be selected to participate in the FL exercise as a data provider. The selected participants download the global model from the blockchain.

Step 3: Upload soft labels. The selected data provider downloads the global model from the distributed ledger to the local server and uses the local health data to train the model. After local training, the hospital server generates the extracted feature local knowledge, and the local server of the hospital uploads the soft labels to the peer node. New blocks are generated through consensus authentication.

Step 4: Update reputation values. The peer node updates the hospital's reputation value based on the result of each round. The number of FL tasks determines the data provider's reputation weight. After the training, the peer node combines the recent reputation value and reputation weight to form a final reputation value, which will be used as the reputation value for the selected data provider for the next task. After consensus certification, the peer node generates a new block, and the reputation values are added to the blockchain. Therefore, all task publishers can train the model by selecting hospitals with high-quality data through reputation values.

Step 5: Update the global model: Unlike traditional FL, the server for global model aggregation is replaced by blockchain nodes randomly selected by the blockchain system. The soft targets stored in the blockchain are used for global aggregation. After averaging the uploaded soft targets, the new aggregated global model is added to the blockchain through consensus authentication. If communication rounds are not completed, continue to the next round of training.

Step 6: Get optimised global model. The task publisher finally downloads the optimised global model from the consortium blockchain. Our approach ensures that hospitals can safely aggregate in a decentralised environment without contributing local data sets.

4.2. Main construction

To solve the problem of heterogeneous remote medical data and against poison attack, algorithm 1 summarises the proposed FedRMD algorithm.

Table

Algorithm 1: FedRMD
Input: global model G; communication rounds T; reputation threshold Γ; peer nodes set P; hospital n’s reputation value $Rvalue{}_n$ and datasets $\{D_n\}$
Output: final global model G
Initialise: global model G to $p\in P$
for each communication round $t=1,\dots ,T\mathit{d}\mathit{o}$
$H_t\leftarrow$ p select hospitals based on the Γ and $Rvalue{}_n$
M $\leftarrow$ G $H_t$ download local model from $p$
for $n\in H_t$ in parallel do:
Local model update: $M_t^n=M_{t-1}^n-\xi \tilde{\mathrm{\nabla }}{L}_i$($M_{t-1}^n$)
Compute the logits: $Z_t(\theta )=\text{ softmax }(\theta )$ for $\theta \in D_{logit}$
Send soft label $Z_t$ to $p$
Consensus certification and generate a new block
end for
$p$ calculates the average soft label for each hospital $Z_s^n=\frac{1}{|C_t|}\sum _{i\in S_t}^{i\ne n}{Z}_t^i$
Record positive and negative interactions with $n$
$p$ updates the global model: $G=G_{t-1}-\xi \tilde{\mathrm{\nabla }}{\ell }_g$($G_{t-1}$)
Consensus certification and generates a new block
end for

4.2.1. Knowledge distillation

During each distillation step $t$, we select a subset $H_t$ based on the reputation values of hospitals in the blockchain. This subset downloaded the global model from the blockchain and training with its own local dataset. Where $\xi$ is the learning rate. Such a training process often takes a huge number of update steps to converge. Parameters of the model on the local hospital after $t$ steps of stochastic gradient descent (SGD) iterations are denoted as follows: (3) $M_t^n=\{\begin{array}{ll}{M}_t^n-\xi \tilde{\mathrm{\nabla }}{L}_t\left(M_{t-1}^n\right)& t|\tau \ne 0\\ {\displaystyle \frac{1}{n}}\sum _{i=1}^n[{\theta }_{t-1}^i-\xi \tilde{\mathrm{\nabla }}{L}_i\left({\theta }_{t-1}^i\right)]& t|\tau =0\end{array}$(3) Knowledge distillation is the process of distilling knowledge from a large and well-trained teacher model to a small student model. We use the Kullback-Leibler (Van Erven & Harremos, 2014) divergence to ensemble all hospitals’ soft labels. Neural networks typically produce class probabilities by using a softmax function (Wang et al., 2018) as follows: (4) $Z_t(\theta )=softmax(\theta )=\frac{\exp \left(\frac{\gamma {({\mathit{x}}_i)}^n}{T}\right)}{\sum _{i=1}^n\exp \left(\frac{\gamma ({\mathit{x}}_i)}{T}\right)}forx\in D_{log}$(4) Let $x_i$ represent as input the logit of i-th health sample. Where $\gamma$ is the softmax function, which will output the prediction score of local model $Z_t$. To enable the peak probability congruence among hospitals, we require the refined peak probabilities of all hospitals to be a constant value $\tau$. After local knowledge distillation, hospitals send the soft label $Z_t$ of the updated local models to the peer node for global aggregation.

4.2.2. Global aggregation

The global model loss function in the peer node. The global knowledge distillation loss encourage to bridge the gap between the task publisher’s soft target $Z_t$ and the soft label $Z_S$ received from the data providers. The global loss function is expressed as The global distillation loss of the task publisher’s dataset is $\{D_t\}$ the cross-entropy loss (Le et al., 2021) which is: (5) $\underset{g\theta }{\text{min}}{\ell }_g={\sum }_{i=1}^N\frac{{\ell }_g(Z_t(D_t),Z_S{(D_t)}^i)}{N}$(5) where $g\theta$ represents the parameter of the global model and $Z_t(D_t)$ is the soft output of the global model $G$. $Z_S(D_t{)}^i$ is the soft target obtained from hospital $i$. On the peer node, it trains the global model $G$ with the reference dataset $D_t$, and the outcomes from local hospital servers extract the generalised knowledge. Afterwards, the peer node broadcasts its knowledge, which as the soft target of the updated global model to data providers’ servers. During the FedRMD training process, the soft label instead of the model parameters can significantly reduce the communication cost of the training process.

4.2.3. Reputation incentive

First, we set that each hospital had a peer node exchanging information, which means that there are $N$ peer nodes in the blockchain system. We assume that $i$ is a new node added to the blockchain system. And the initial reputation value is determined by the reputation values of other nodes, which is (6) $Rvalu{e}_i^{init}=\frac{\sum _{n\in N}Rvalu{e}_n}{N}$(6) where $n$ is one of the hospitals selected to participate in FL. We use Beta reputation (Josang & Ismail, 2002) for a hospital $i$ is calculated as the recent reputation value, which is as follows: (7) $Rvalu{e}_{n\to ts}^{rect}=\frac{\alpha }{\alpha +\beta }$(7) where $\alpha =positive(n)+1$ and $\beta =negative(n)+1$ is the accumulated positive and negative reputation values at one point in time. Where $n\to ts$ is the hospital $n$ participates in one of FL tasks published by the system. We also design reputation weight to encourage hospitals to actively participate in FL, which is: (8) $Rweigh{t}_n=\frac{ts}{Ts}$(8) Let $t{s}_{}$ represent as the number of times $n$ accepts FL task. Where $Ts$ is the sum of FL tasks in the blockchain. So, the hospital $n$ final reputation value is (9) $Rvalue{}_n^{final}=\lambda Rvalue{}_n^{init}+Rvalue{}_n^{rect}\times Rweigh{t}_n$(9) where $\lambda$ is the weight of the node’s initial reputation. It gradually decreases with the progress of with FL training. To reduce the impact of low data quality submitted due to high initial reputation values.

5. Experimental settings

5.1. Datasets and learning model

We validate the efficacy of the FedRMD algorithm with the COVID-19 radiography (Chowdhury et al., 2020; Rahman et al., 2021) and COVID-Chestxray (Cohen et al., 2020) dataset. COVID-19 radiography dataset has 3616 COVID-19 positive cases along with 10,192 normal, 6012 Lung Opacity (Non-COVID lung infection), and 1345 viral pneumonia images. COVID-Chestxray dataset is created by assembling medical images from websites and publications and currently contains 123 frontal view X-rays. To build an open public dataset of chest X-ray and CT images of patients which are positive or suspected of COVID-19 or other viral and bacterial pneumonias. In all experiments, the neural network model is ResNet-18 (He et al., 2016).

5.2. Baseline methods

We select the classic FL algorithm and the most advanced FD algorithm for comparison. To demonstrate the high accuracy of our algorithm in the heterogeneous medical data.

• FedAvg (McMahan et al., 2017): This method is a baseline of FL, after each round of the training process, all trained parameters in each participated client will be sent to the central server for aggregation.

• FedDF (Lin et al., 2020): This method consists of two chief components: ensemble and knowledge distillation via out-of-domain data. It uses unlabelled datasets from other domains to achieve model fusion in order to protect the privacy of FL.

• FedGen (Zhu et al., 2021): This method directly regulates local model updating using the extracted knowledge. It is a data-free FD method with flexible parameter sharing.

5.3. Evaluation environment

The hardware configuration is shown in Table 2. The software configuration is shown in Table 3.

Table 2. Hardware configuration.

CPU	AMD Ryzen 7 5800H
GPU	NVIDIA GeForce RTX 3060 Ti
SSD	1TB
Operating system	Ubuntu 20.04.3
Memory	DDR4 3200 8G*2

Table 3. Software configuration.

Python	3.8.3
Tensorflow	2.3.0
Hyperledger Fabric	2.3.1
Hyperledger Caliper	0.5.0
Docker	20.10.12
Go	1.15.11
numpy	1.22.2
Scikit-learn(sklearn)	1.0.3

5.3.1. Federated distillation learning details

We assume each computer to be a hospital server that stores the remote medical data. We also have a computer emulated as a global server. During the implementation, we use Dirichlet distribution Dir(α) to simulate the heterogeneous data among patients, where a smaller α shows a higher data heterogeneity.

We set α = 0.2, 0.5 and 1. The local hospital datasets are unbalanced and have a few training samples. Communication rounds T = 200. In each communication round, we randomly select hospitals to take part in the learning process. The local update and distillation, the batch size B = 32. The local update steps and the distillation steps as 15 epochs in each communication round. The step sizes of the local update and distillation are set to 0.02 and 0.002, respectively. The number of global epochs R = 4. And Table 4 shows the hyperparameter that is used in our experiments.

Table 4. Hyperparameter list.

	Hyperparameter	Value
Shared parameters	Learning rate	0.02
	Optimiser	sgd
	Local update steps	15
	Batch size(B)	32
	Communication rounds(T)	200
	Total hospitals	10
FedGen	Generator Optimiser	adam
	Generator learning rate	10^−4
	Generator inference size	128
	User distillation batch size	32
FedDF	Ensemble Optimiser	adam
	Generator learning rate	10^−4
	Ensemble batch size	128
FedRMD	Ensemble Optimiser	adam
	Generator learning rate	0.002
	Ensemble batch size	32

5.3.2. Consortium blockchain details

The experiment established a reliable transaction service by building blockchain in Hyperledger Fabric 2. 3.1. We use the PBFT as consensus algorithm, running on 10 peer nodes. And the SHA-256 encryption algorithm is used as the secure hashing algorithm. The data size for reputational values ranges from 100KB to 200KB.

5.4. Evaluation metrics

We use accuracy (ACC), Recall, and Precision as the evaluation indicators. Furthermore, we also use the Area Under Curve (AUC) to compare the classification results comprehensively. AUC is an average of true positive rates over all possible values of the false positive rate. To do this, we first obtain True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN). ACC can be calculated as follows: (10) $\text{ACC}=\frac{\text{TP}+\text{TN}}{\text{TP}+\text{TN}+\text{FP}+\text{FN}}$(10) Recall can be calculated as follows: (11) $\text{Recall}=\frac{\text{TP }}{\text{TP + FN}}$(11)

Precision can be calculated as follows: (12) $\text{Precision}=\frac{\text{TP}}{\text{TP + FP}}$(12)

We use transactions per second (TPS) as the indicator for performance of systems that handle routine transactions and record-keeping. TPS can be calculated as follows: (13) $\text{TPS}=\frac{\text{Number of transactions}}{\text{Number of seconds}}$(13)

At present, there is no international standard for blockchain performance indicators, according to existing Chinese Blockchain Application Security Technology (DB43/T 1842-2020), the test standards are shown in Table 5.

Table 5. Blockchain evaluation standards.

Success rate	>95%
Transaction per second (TPS)	>200
Average latency	<1 s
Average response time	<0.5 s

6. Results and analysis

6.1. Performance evaluation

To simulate realistic medical scenarios, the communication resources of various hospital servers differ. In data fusion with global services, only servers from some local hospitals are involved. Therefore, we set the number of local hospital services is N for each round of local servers taking part in the FedRMD process in each communication round. We use the COVID-Chestxray dataset as patients’ data stored in local hospitals, where the data heterogeneity α = 1.

We use the same colour to represent the number of hospitals participating in the training. It showed the experimental results in Figure 3. For all values of all N, the accuracy gradually increases as the number of communication rounds increases. When the training process converges, the accuracy of the training varies slightly. Furthermore, the reason is that with the increase in the number of hospitals participating in the training of the model, the more variety of knowledge the task requester's global model acquires, the better the performance of the model and the more accurate the results. However, in the case of low data quality will have a negative impact on the model training of FL. Therefore, we use the reputation mechanism to control the data quality and improve the quality of FL training.

Figure 3. Accuracy of different number of hospitals.

Although test accuracy is similar, the average time cost per round varies widely, as shown in Figure 4. For example, when 80% of hospitals are involved, the value of communication cost is 16.18s. However, 40% of hospitals are involved, and the communication cost is 8.25s. The more hospitals involved in the training, the greater the medical data that needs to be trained and the more extended time cost. While the amount of time spent training locally in FL does not increase with the number of other participants, the combination with blockchain requires consistent data stored in the distributed ledger. As the amount of data uploaded by the hospital to the blockchain increases, the time taken through the consensus mechanism increases. We hope to reduce the time of consensus in the future.

Figure 4. Time cost of different number of hospitals.

We set each hospital server with a different reputation value in the range of [0,1]. In the experiment, we set the reputation threshold to 0.2 and 0.5 to simulate the process of the blockchain dynamically screening hospital servers to participate in training. We used our proposed FedRMD algorithm for 200 rounds to study the effect of local reputation values on model convergence. Figure 5 shows that the training loss is larger when the reputation incentive is not used. The training loss decreases when the reputation value set to 0.2. The training loss is smaller when the reputation value set to 0.5. Therefore, using reputation inventive can improve the efficiency of FL, providing high-quality data sharing for FL training.

Figure 5. Training loss during poisoning attacks.

Caliper can easily interface with multiple blockchain platforms and shield the underlying details, and users only need to be responsible for designing specific test processes to obtain the visual performance test report. We use Caliper 0.5.0 to test the performance of Hyperledger Fabric, and we show the HTML report in Figure 6. According to the test report, we got the success rate, TPS, average latency, and so on. Compared with the test standard of blockchain, our system has passed the stress test, and the success rate has reached more than 95%. The system runs at a low load without breakdown. And all the test results meet the performance evaluation standards listed in Table 5.

Figure 6. Performance report of Hyperledger Fabric.

As shown in Figure 7, we test the average latency of 12,000 blocks. Respectively, and the average latency time also increased as the number of blocks increased, reaching an average latency of 519.6 ms when the number of blocks reached 12,000. Figure 8 shows that the system's TPS reaches a maximum of 281.9 when the number of blocks reaches 5000. But with the number of blocks increases, the system's TPS decreases slightly. The performance degradation is mainly due to system overwork. Key technologies such as sort execution verification and asynchronous block generation reduce the overhead of deterministic execution and block generation. Therefore, we can improve the TPS of the system through two improvements. On the one hand, it improves the performance of the consensus layer. For example, reduce the number of nodes, increase the consensus speed, modify the data consistency method, and broadcast rotation. On the other hand, the Sharding technique can be used, where the main chain only synchronises the block header and validation information, and the content method is completed with many sharding. Many high-frequency trades are placed off-chain, with only the blockchain recording the final transaction information. But the TPS of Hyperledger fabric system can still reach 250. Therefore, it is feasible to storage the reputation values and the results of model training.

Figure 7. Average latency of blocks.

Figure 8. Transactions per second of blocks.

6.2. Performance comparison

Table 6 shows that we use the COVID-Chestxray dataset in the case of heterogeneous α = 1 for experiments. Figure 9 shows that 80% of the local hospital servers are involved in the training process. When the accuracy of the model does not improve significantly, we think the model reaches convergence. From the figure, we can see that our algorithm model converges faster. Other methods can only randomly select data providers to train the model. Ours uses a reputation incentive mechanism to control the quality of data through blockchain, so our method is more accurate. Figure 10 shows that AUC of our approach and the three baseline approaches. Compared to the traditional FL, the communication cost generated by FD depends only on the model output dimension. In our method, the communication cost of uploading from a local hospital server to block nodes is determined by the size of the soft labels.

Figure 9. Accuracy trends on the COVID-Chestxray dataset.

Figure 10. AUC scores on the COVID-Chestxray dataset.

Table 6. Different methods on COVID-Chestxray.

	COVID-Chestxray (α = 1)
Methods	Accuracy (%)	Recall (%)	Precision (%)	AUC (%)
FedAvg	82.73 ± 0.13	76.23 ± 0.61	80.11 ± 0.93	86.93 ± 0.23
FedDF	83.94 ± 0.31	81.13 ± 0.87	85.34 ± 0.74	90.22 ± 0.38
FedGen	84.41 ± 0.21	83.82 ± 0.71	88.27 ± 0.82	91.87 ± 0.21
FedRMD	87.33 ± 0.35	85.26 ± 0.76	91.19 ± 0.65	93.78 ± 0.19

Table 7 shows that we use COVID-19 radiography dataset in the case of heterogeneous α = 0.2 for experiments. From tables clearly shows that our proposed FedRMD is better than the other three algorithms in medical data is heterogeneous. Obviously, our method is 12% higher than FedAvg. FedGen also gets a good training result. FedGen has 6% higher training accuracy than FedDF. In FedRMD, the local hospital model needs to classify the resulting images, but also needs to participate in other FL tasks hospitals, which is more accurate than others that rely entirely on generators to transmit information. Figure 11 shows that 40% of the local hospital servers are involved in the training process. Although our accuracy is similar to FedGen sometimes, our final training results were 3% higher than FedGen. Figure 12 shows that AUC of our approach and the three baseline approaches.

Figure 11. Accuracy trends on the COVID-19 radiography dataset.

Figure 12. AUC scores on the COVID-19 radiography dataset.

Table 7. Different methods on COVID-19 radiography.

	COVID-19 radiography (α = 0.2)
Methods	Accuracy (%)	Recall (%)	Precision (%)	AUC (%)
FedAvg	54.31 ± 0.32	51.45 ± 0.82	62.23 ± 0.63	67.20 ± 0.33
FedDF	57.84 ± 0.19	56.23 ± 0.79	68.53 ± 0.58	70.94 ± 0.38
FedGen	62.36 ± 0.35	63.86 ± 0.52	70.88 ± 0.72	77.88 ± 0.41
FedRMD	66.53 ± 0.27	68.19 ± 0.66	73.16 ± 0.53	82.74 ± 0.39

6.3. Security analysis

In traditional FL, cryptographic encryption technology encrypts the local model parameters and gradient updates so that attackers cannot obtain any information. But, it cannot control the quality of the data source data. So, we propose a reputation incentive based on consortium blockchain to reduce the attack of malicious nodes, and increase the weight of involved in the FL tasks encourage the hospital to actively participate. The transaction is only passed into the block after the consortium blockchain has verified it. So blockchain guarantees that the task publisher will have access to high-quality data. Even if malicious attackers can participate in interactive training at the beginning. After many iterations, their reputation values become low. Their reputation values are below the reputation threshold. Reputation incentive successfully defended against the poisoning attacks.

7. Conclusion and future work

In this paper, we propose a new FL algorithm called FedRMD for data security sharing in the remote medical system. Combining knowledge distillation and blockchain technology solves two key problems in FL. First, the model accuracy decreases due to data heterogeneity, and the communication cost is high. Second, the quality of data uploaded by participants cannot be controlled during FL training, which is affected by poisoning attacks. To comprehensively solve the above problems, we use Hyperledger Fabric to design a reputation-based consensus to help FL task publishers get excellent global models. We use the knowledge distillation method to transfer soft targets rather than model parameters during the model training process, significantly reducing communication overhead. The task publisher uses the reputation mechanism to select the data provider, and constantly adjusts the reputation value of the participants in each round of training, so as to control the uploading of toxic data by malicious nodes and successfully resist the influence of poisoning attacks on the model training. In addition, the combination of FL and blockchain solves the system paralysis caused by the failure of the central server and improves the robustness of the remote medical system. Extensive experiments under the heterogeneous COVID-19 data show that our method performs well compared to the FedAvg, FedDF, and FedGen.

In future work, we will study how to improve the consensus mechanism of blockchain to improve the sharing efficiency, how to dynamically adjust the size of the reputation threshold, improve the flexibility of the reputation mechanism, and enable more medical institutions to share data while improving the accuracy of the model. In addition, at present, we only provide remote medical data sharing in the Hyperledger Fabric, so we will consider how to solve the problem of cross-chain data sharing and provide a more convenient data-sharing environment for the remote medical system.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This work was supported in part by the Natural Science Foundation of Hainan Province [621RC508], State Key Laboratory of Information Security [2022-MS-04], the Science Project of Hainan University [KYQD(ZR)-21075].