A Secure Internet of Medical Things Framework for Breast Cancer Detection in Sustainable Smart Cities

Abstract

Computational intelligence (CI) and artificial intelligence (AI) have incredible roles to play in the development of smart and sustainable healthcare systems by facilitating the integration of smart technologies with conventional medical procedures. The Internet of Things (IoT) and CI healthcare systems rely heavily on data collection and machine learning since miniature devices represent the foundation and paradigm shift to sustainable healthcare. With these advancements in AI techniques, we can reduce our environmental impact, while simultaneously enhancing the quality of our services. Widespread use of these devices for innovative IoT applications, however, generates massive amounts of data, which can significantly strain processing power. There is still a need for an efficient and sustainable model in the area of disease predictions, such as lung cancer, blood cancer, and breast cancer. The fundamental purpose of this research is to prove the efficacy of a secure Internet of Medical Things (IoMT) in the detection and management of breast cancer via the use of gated recurrent units (GRUs), which are a more recent version of recurrent neural networks (RNNs). The blockchain has been employed to achieve the secure IoMT. Unlike long short-term memory units, they do not have a cell state of their own. Therefore, we have combined GRU with RNN to achieve the best results. When training a GRU-RNN classifier, it is typically necessary to collect tagged IoT data from many sources, which raises significant concerns about the confidentiality of the data. To verify the model, the experiment is performed on Wisconsin Diagnostic Breast Cancer (WDBC). The experimental result shows that the GRU-RNN has been archived 95% in terms of the accuracy metric, and the efficacy of the proposed IoMT model is superior to the existing approach in terms of accuracy, precision, and recall.

1. Introduction

The healthcare industry is currently undergoing significant changes around the globe. In addition, shifts in population, the economy, society, and technology are compelling us to rethink every aspect of the healthcare sector. As a consequence of the alterations, the methods of provision, the distribution of resources, the models of financing, the advancement of scientific knowledge, and the functions of physicians are all undergoing change. The transformation of patient medical files from traditional paper to electronic medical records has also contributed to the flourishing of the health information sector. Patients’ medical histories, clinicians’ handwritten notes, computerised physician order entries, and imaging equipment are some of the potential sources of clinical data. The datasets of medical records are fragmented and unstructured compared to other segments of industries [1]. There may be a shortage of actionable knowledge that can be used in health budgeting, policymaking, and resource allocation because of the healthcare industry’s reluctance to build tools to leverage this body of data and information. These legacy data are scattered across multiple healthcare systems, health insurers, researchers, academia, government entities, and so forth. In addition, each of these data repositories is isolated and, therefore, inherently unable to provide a platform for global data transparency.

According to the McKinsey Global Institute, if the healthcare industry in the United States were to utilise big data in an innovative and efficient manner, that would enable the market to generate more than USD 300 billion revenue in a year. Out of this, 66% of the value would be in the form of a reduction in the expenditure that the United States makes on healthcare [2]. However, according to a different study, the digital universe as a whole is expanding at a rate of 40% per year, which leads to the conclusion that the expansion of digital data used in healthcare will be more rapid than that of the rest of the digital universe [3]. The ever-increasing prevalence of the usage of auto-capturing and sensing technologies within the healthcare business paves the way for the industry to collect huge quantities of medical information. Nearly one-third of all data in the world are produced by the healthcare industry, and the proportion is steadily growing [4]. The rising tide of electronic medical information, unstructured data, and imaging records poses unprecedented challenges for data centres that face infrastructure limitations. There is a need for continuous operation, access, cleaning, and discovering knowledge. The systems and data need to be more manageable with more visibility, value, security, and measurement. Data centres need to be integral parts of healthcare facilities’ operations [5].

Among the important functions of the government is to provide the best public healthcare for every citizen. Nowadays, advancements in medical technologies have raised people’s expectations of health systems, which has increased health expenditures at all levels, viz., hospitals and para-clinical settings, as well as laboratories and diagnostics centres. Furthermore, the increase in healthcare service expenditures has changed hospitals into costly entities, as a significant portion of healthcare sector resources is allocated to hospitals. This proportion is estimated to be about two-thirds of the government health budget in many developing countries. Presently, the budget allocated to healthcare is 7.61% of the total government budget [2].

The budget of the Ministry of Health in 2017 reached SAR 67.8 billion, amounting to 7.61% of the total kingdom budget and equivalent to an increase of SAR 8.86 billion (0.6%) from the previous year’s allocated resources [5].

The budget is allocated using traditional methods that are inefficient; therefore, allocating resources to the diseases that are most widespread in the country requires new methods and structures. One of the most common malignancies among women worldwide is breast cancer. Cancer is a disease that is caused by abnormal cells that are found in the human body and have the potential to spread to other sections of the body, in addition to the area that is currently affected. It accounts for 8.2 million deaths per year, making it the leading cause of death around the globe [6]. Over the next 2 decades, cancer cases are predicted to rise from 14 to 22 million, then continue to rise year after year. Therefore, the budget needs to spend on machine learning (ML) and IoT models for better diagnosis. At present, deep learning (DL) is a modern and motivating field of machine learning. DL is the most efficient, supervised, and time- and cost-efficient machine learning technique. It is not a limited learning technique, but it involves a variety of techniques, as well as topographies, that can be useful for a huge speculum of complex issues. This method studies illustrative and differential characteristics in a very stratified manner. DL approaches have taken a major step forward, with a substantial concert in an extensive range of applications with helpful security tools.

Medical professionals have been given a helping hand by machine learning and AI methods in making diagnoses, predicting illness progression and complications, and identifying preventive measures. This method, such as support vector machines (SVMs), artificial neural networks (ANNs), k-nearest neighbours (K-NNs), convolutional NNs (CNNs), recurrent neural networks (RNNs), and long short-term memories (LSTMs), has aided clinicians in identifying the heritable variants that point the road to illness [7,8]. In addition to its core network structural design [9,10,11,12], RNN also contains a rich collection of designs. If you compare an RNN to a network with just feedforward connections, the RNN’s distinguishing characteristic is the link it includes, which may be configured as feedback into previous levels. It recalls data from the past to model current issues. Standard BP, sometimes called back propagation via time, may be used to train and grow these networks (BPTT). The update gate and the reset gate are two types of GRU gates. An update gate’s job is to signal whether data from the previous cell are needed for upkeep. The reset gate explains how the new input is integrated into the existing data in the cell. By setting the update gate to 0 and the reset gate to 1, the GRU mimics the behaviour of a conventional RNN. The GRU model’s functional capacity is more straightforward than that of LSTMs. Skills development is rapid, and its implementation is thought to be more efficient [13].

Our contribution:

In this research, we present a secure GRU-RNN, an encrypted IoT data-training technique that uses blockchain technology to protect user privacy. Using blockchain technology, we provide a secure and consistent data-sharing concept between several data sources, whereby IoT data are encrypted and then stored on a distributed ledger. We established a trusted third-party-free GRU-RNN training algorithm and designed safe blockchain building blocks utilising the Advanced Encryption Standard (AES) cryptosystem.

The paper is organised as follows. Section 2 provides the related work, while Section 3 illustrates the material, methods, and the proposed algorithm. Section 4 provides the results, discussion, and validation of the results with the existing methods, and the paper is concluded in Section 5.

2. Literature Review

RNN, CNN, Fuzzy system, and swarm intelligence are some of the examples of computational intelligence technologies that have found application in the field of smart health. ML has been used to apply a variety of classification algorithms, such as Naive Bayes, Decision Tree, Logistic Regression, Random Forest, Support Vector Machine, and others, in a number of studies that have been done on the diagnosis of breast cancer. These studies have been carried out in a number of different ways. In order to improve the accuracy of prediction by assigning the most relevant characteristics, many feature-selection procedures have been utilized, including filtering, wrapping, and embedding, amongst others. Apart from accuracy, security is equally important; therefore, secure SVM training based on blockchain and IoT data was proposed by Shen et al. [14].

Ogundokun et al. [15] proposed a medical Internet of Things-based diagnostic system that can distinguish between individuals with cancerous and benign tumours. For malignancy/benignity classification, the authors used ANN and CNN with hyperparameter optimization, with SVM and MLP serving as comparative baselines. In that they have a direct bearing on the actions of training algorithms and, thus, on the effectiveness of machine learning models, hyperparameters play a key role in the development of machine learning algorithms and are, therefore, very important. To improve the classification performance of MLP and SVM, the authors utilized a particle swarm optimization (PSO) feature-selection strategy to pick better features from the breast cancer dataset. Similarly, a grid-based search was utilized to determine the optimal balance between the CNN and ANN hyperparameters. When utilizing CNN, the proposed method achieved an accuracy of 98.5%, and while using ANN, it achieved 99.2%.

In Muhammad et al. [16], the authors suggested an IoT-based diagnosis system that can accurately categorize tumours as either cancerous or benign. The method suggested uses a support vector machine to distinguish between potentially malignant and benign. The authors employed a recursive feature-selection approach to refine their feature choices from the breast cancer dataset in order to boost the classification performance of the system. The predictive model is trained and tested using a classifier, with training and testing splits. The experimental findings show that the best subset of features is chosen using the recursive feature-selection process, and that the SVM classifier provides the best classification performance when used with these features. High levels of accuracy in classification (99%), specificity (99%), and sensitivity (98%), as well as a Matthews’ correlation value of 99%, were attained with the SVM kernel linear.

Siddhant et al. [17] emphasized the need for machine learning models for early detection of breast cancer. The authors made an effort to come up with a method that might help the patient evaluate whether or not she is at risk for developing breast cancer at an early stage herself so that the breast cancer cells can be eliminated with the help of appropriate treatment. A CNN method was proposed by the authors, and it has an accuracy of approximately 86% on the validation dataset.

Chougrad et al. [18] proposed the use of a network method to enhance breast cancer patient survival in the early stages of the disease. The authors were successful in detecting breast cancer in its early stages by utilizing a deep convolutional neural network (DCNN). The DCNN model with the highest accuracy, 98.94%, served as the foundation for the Breast Cancer Screening Framework.

Ragab et al. [19] presented a novel, ensemble, deep learning-supported decision support system for analysis of ultrasound images of breast cancer. The authors developed an optimal, multilevel, thresholding-based, picture-segmentation technique to identify the tumour-affected regions, in addition to developing a powerful ML classifier for breast cancer detection. Further, the author developed a feature-extraction ensemble including three deep-learning models. According to the authors, the level of accuracy is 97.52 percent.

Shravya et al. [20] presented a comparative analysis of the implementation of models utilizing Logistic Regression, SVM, and K-NN on the dataset retrieved from the UCI repository. The effectiveness of each algorithm was measured and compared based on the outcomes of its accuracy, precision, and sensitivity. Spyder, the Scientific Python Development Environment, executes these Python-coded approaches. The experiments demonstrated that SVM provides the most accurate predictive analysis at 92.7%.

Mansour et al. [21] presented a computer-assisted approach for breast cancer screening that employs a Gaussian aggregation model and AlexNet-DNN for feature extraction, which is an adaptive learning-based model. Additionally, the authors performed principal component analysis (PCA) and linear discriminant analysis (LDA) for better understanding. The proposed approach obtained an accuracy of 96.70%.

Sahar et al. [22] proposed transfer-learning AlexNet to classify and detect breast cancer. Both deep learning and transfer learning are methods that can be tailored to the particular characteristics of a given dataset. On three different datasets, the suggested model applied a modified version of the AlexNet algorithm. Using a modified version of AlexNet helped this proposed model, which is equipped with transfer learning, obtain the best possible outcomes. On the A2 dataset, the best possible score was accomplished.

Prachumrasee et al. [23] developed a prototype that is an IoT for prediction of breast cancer. The authors performed the analysis in Thailand. The prototype was discovered to be capable of monitoring fixation times, and it even has an inbuilt notification feature for when those times go above the golden period.

Zheng et al. [24] proposed an efficient mathematical to present the Deep Learning-aided Efficient Adaboost Algorithm (DLA-EABA) for spotting breast cancer. Traditional computer vision techniques are actively supplemented by the DCNN to produce transfer learning for tumour classification. The classification and error estimation are implemented in a fully connected layer and a softmax layer. In order to determine the best strategy, the proposed study integrated these machine learning strategies with feature-selection and extraction techniques, testing their results with classification and segmentation tools. Compared to other current systems, the test findings revealed a high degree of accuracy of 97.2%, sensitivity of 98.3%, and specificity of 96.5%.

Siddiqui et al. [25] proposed a cloud-based IoMT model to accurately predict breast cancer stages. Breast cancer detection and staging is achieved with the help of the proposed approach. The experimental results showed an accuracy of 98.86% during training and 97.81% during validation. The proposed intelligent prediction of breast cancer stages enabled by deep learning (IPBCS-DL) model shows more accuracy than existing state-of-the-art methods, suggesting it has the potential to reduce the death rate associated with breast cancer.

Aldhyani et al. [26] presented a deep learning model that has been suggested as a potential solution for identifying cancer cells that are rather tiny. The authors used the BreCaHAD dataset to perform histological annotation and diagnosis. The authors of the study used data augmentation with 19 different parameters, such as scale, rotation, and gamma, to prevent overfitting. The hybrid-dilation deep-learning model that was suggested combines two different approaches. Through the utilization of the dilated residual expanding kernel model, the suggested dilated unit analyses the picture and transmits the processed features to the Alexnet. Additionally, it is able to distinguish minute objects and thin borders. The new technique is superior to the old one according to the accuracy, which came in at 96.15%. A comparative analysis of the existing work is shown in

Table 1. Comparative analysis of existing work.

Work	Method	Dataset	Accuracy
Chougrad et al. (2018) [18]	DCNN	BCDR	98.94
Shravya et al. (2019) [20]	SVM	UCI	92.7
Muhammad et al. (2019) [1]	SVM	WDBC	99
Mansour et al. (2020) [21]	AlexNet-FC7	Custom DB	96.7
Zheng et al. (2020) [23]	DLA-EABA	Cancer Imaging	98.3
Siddhant et al. (2021) [17]	CNN	Custom DB	87.84
Siddiqui et al. (2021) [24]	IPBCS-DL	IRMA	98.86
Aldhyani et al. (2022) [25]	Alexnet	BreCaHAD	96.15
Ragab et al. (2022) [19]	VGG-16 + VGG-19	Breast Ultrasound	97.52
Ogundokun et al. (2022) [1]	CNN + ANN	WDBC	99.2
Sahar Arooj et al. (2022) [22]	CNN-AlexNet	Image DB	99.4

.

3. Material and Methods

In many areas of smart cities, where massive amounts of data are collected from a broad variety of IoT devices, DL methods have been extensively used. To effectively classify data, DL models, such as RNN, are put to use in real-world situations, such as illness diagnosis and anomaly detection. When compared to older RNNs, such as LSTM units, the newer GRU does not use a cell state. Data privacy is a major problem since training a GRU-RNN classifier often requires a collection of labelled IoT data from different sources. Existing systems generally assume incorrectly that training data can be successfully acquired from many data suppliers. Utilizing gated recurrent units (GRUs), which are a more up-to-date form of recurrent neural network, the primary objective of this study is to demonstrate that a safe Internet of Medical Things (IoMT) is effective in the diagnosis and treatment of breast cancer. This will be accomplished through the application of gated recurrent units (RNNs). The distributed ledger technology (blockchain) was used so that the IoMT could be made secure. They are not like long-term memory units in that they do not have their own distinct cell states. As a consequence, we decided to mix GRU and RNN in order to get the best possible results. In this research, we presented safe GRU-RNN, a privacy-preserving GRU-RNN training method that uses blockchain-based encrypted IoT data to bridge the gap between ideal assumptions and reality constraints. Using blockchain technology, we provide a trustworthy platform for data exchange between various data sources, whereby IoT data are encrypted before being stored on a distributed ledger. With the help of the AES cryptosystem, we established a trustworthy, secure GRU-RNN training algorithm. The purpose of a GRU is to address the vanishing gradient issue that occurs with a regular RNN. Similarly, the GRU may be seen as an improvement in the LSTMs. Update gates and reset gates are used in GRU. With these gates, a GRU may impact the future by storing data over many time periods. That is to say, the value is saved in memory for a certain period of time, and then, at a crucial moment, it is retrieved and combined with the existing state to affect an update at some point in the future. This protected classifier model is intended for use in the context of individual healthcare, where the GRU-RNN classifier may be put to good use in the detection of illness and the provision of precise diagnoses.

3.1. Dataset

3.1.1. Wisconsin Breast Cancer Dataset

The Wisconsin (diagnostic) Breast Cancer Dataset was obtained via Kaggle. When cells in the breast grow and divide out of control, this may lead to breast cancer. It is probable that there might be a large variety of breast cancer instances. What are the many kinds of cells that may grow from breast tissue? Malignant breast cancer is the most severe kind. The Health Wisconsin Diagnostic Breast Cancer (WDBC) Dataset is what medical professionals turn to in order to determine if a tumour is cancerous or benign.

The following website allows learning more about the breast cancer research that has been conducted in Wisconsin: https://www.kaggle.com/datasets/uciml/breast-cancer-wisconsin-data/discussion/62297 (accessed on 22 November 2022). There are a total of 569 instances with 32 distinct features. ID/diagnosis (with 30 input characteristics and their corresponding real-world values) is a dataset containing two classes, M = malignant and B = benign.

3.1.2. HDD Dataset

Datasets pertaining to cardiac disease were extracted from the UCI collection. This data collection has 303 instances, each with their own set of multivariate characteristics (which may include the integer, category, or real value), as well as 14 distinct variables to investigate. The link of the dataset is http://archive.ics.uci.edu/ml/datasets.php?format=&task=cla&att=&area=life&numAtt=10to100&numIns=&type=mvar&sort=typeUp&view=list (accessed on 22 November 2022).

3.2. Gated Recurrent Unit-Based Recurrent Neural Network (GRU-RNN)

The GRU-RNN is a variant of RNN that focuses on solving the problem of a vanishing gradient. It consists of only two gates, viz., the update and reset gates. The functionality of the update gate is similar to the input gate and forget gate of LSTMs, as shown in Figure 1.

This gate decides which input passes through the network and which input is thrown away.

The reset gate decides on the amount of information to be erased. There are fewer tensor operations taking place inside the GRU-RNN, and they work faster than the LSTMs. The gate equations of the GRU-RNN are expressed below.

$$h_t=g\left(W{x}_t+U{h}_{t-1}+b\right)$$

(1)

$$U_t=\sigma \left(W^u{x}_t+V^u{h}_{t-1}\right)$$

(2)

where $x_t$ denotes input data; it is multiplied by its own weight $W^u$. The variable $h_{t-1}$ holds the data of the previous iteration t − 1. The previous data are multiplied by their own weight $V^u$. After the findings have been combined together, a sigmoid activation function is applied to them.

$$R_t=\sigma \left(W^r{x}_t+V^r{h}_{t-1}\right)$$

(3)

A sigmoid activation function is applied to the product of the weights multiplied by themselves, and then the product is added to the original weights. The following equations can be used to express the final hidden state.

$${\tilde{h}}_t=\tan \mathrm{h}(\left(W_h{x}_t+r_t\odot U_h{h}_{t-1}+b_h\right)$$

(4)

To minimize this loss function in terms of W, we chose the cross-entropy of the scalar output y as the metric of choice. To learn the parameters of the proposed model, we employed the cross-entropy of the scalar outcome y as the loss function and searched for its minimization in terms of $W_{\left\{z,r,h\right\}}$, $U_{\left\{z,r,h\right\}},b_{\left\{z,r,h\right\}}$, $w,$ and $b$. The loss funtion can be defined as follows.

Table 2. The main parameter values of GRU method.

Parameters	Values
maxEpochs	100
miniBatchSize	20
InitialLearnRate	1 × 10−4
Optimizer function	rmsprop
NumOfUnitsInFirstlayer	120
NumOfUnitsInSecondlayer	75

shows the parameters of the GRU method.

$$L=-{\displaystyle \sum }_{j=1}^n(c^j\log y^j+\left(1-c^j\right)+\log \left(1-y^j\right))$$

(5)

3.3. Block Chain

The blockchain system is defined as an open and scattered ledger that takes the inputs in the form of block lists. These blocks are constructed for the purpose of storing transactions with the help of cryptocurrency systems. Blockchain systems ensure a reliable form of transaction since the actual data are not passed. The data that are passed are in encrypted form. In this work, all input data were ciphered using the AES algorithm. The ciphertext was then converted into a block. Several blocks were combined with hash keys to form a complete blockchain [27].

Using WDBC data, a blockchain transaction can be generated. Large quantities of information, including data, control, and results, are produced as a result of the gateway’s series of transactions. Selecting a node in the blockchain that is reliable and efficient is needed for a fast and stable consensus method. Mining nodes are selected based on a variety of metrics, including as computational capability, storage space, social standing, mining rewards, output, and network throughput. Proof-of-work (PoW) is an established consensus mechanism in most of the blockchain application.

Before a block is formed in the Merkle tree, the mining node compiles a list of all of the transactions that have occurred and iteratively hashes the information that it has collected. The hashing process is complete when the hash of the transactions reaches or falls below a predetermined goal value known as a threshold. This value is denoted by the symbol $T_h$ and is described mathematically as Equation (6). H is for the SHA-512 hashing algorithm, and $BC$ is for the current block.

$$H(n||H\left(B\right))\le T_h$$

(6)

Equation (7) present the probability of finding nonce evidence of $H$.

$$P\left(H\le T_h\right)=\frac{T_h}{2^{512}}$$

(7)

The miner that successfully computed the target hash sends the proof to all nodes in the blockchain network, along with the transactions and other data necessary for other miners to recompute and, thus, connect/add the new block to the network.

In an IoT environment, an IoT analyst takes care of the data processes. However, for applications such as disease prediction, each patient’s information must be kept private and secure. Hence, this work proposes a privacy-preserving secure scheme of training for disease prediction. This ensures that the third-party analyst does not have access to the actual data. The analyst has access only to the encrypted data stored in the blockchain. The training for disease prediction is done using this ciphertext model of encrypted data. This model was tested for structured and unstructured data to validate its performance.

4. Results and Discussion

• Recall:

The recall (or sensitivity) of a classification system is a crucial metric that may be derived from a confusion matrix by dividing the sum of correct positive predictions by the sum of all positive predictions (see Equation (8)). Sensitivity is synonymous with the true positive rate (TPR) or recall (REC) number of correct model predictions for the positive class.

$$Recall \%=\frac{TP}{TP+FN}\times 100$$

(8)

• Precision:

An essential metric known as precision (PR) is calculated from the confusion matrix by dividing the number of accurate positive predictions by the number of positive predictions, as illustrated in Equation (9). A positive prediction value (PPV) is another name for it.

$$Precision \%=\frac{TP}{TP+FP}\times 100$$

(9)

• F1-Score:

The F1-score (sometimes called the F-score or F-measure) gives more weight to accuracy than recall. It is the mathematical mean or average of how well a system can identify and retrieve information. With the right balance between accuracy and recall, an F1-score may be maximised. In contrast, the F1-score suffers if one metric is improved at the expense of another. Consider the case when P = 1 and R = 0. The resulting F1-score is 0.

$$F1-score \%=\frac{2\ast \left(Recall\ast Precision\right)}{\left(Recall+Precision\right)}\times 100$$

(10)

In this study, we put the methodology to work using two datasets taken from the real world: the Wisconsin Dataset Breast Cancer (WDBC) and the Heart Disease Dataset (HDD), both of which may be downloaded from the UCI machine learning repository. The WDBC’s features, which characterise the properties of the cell nuclei contained in a digital picture of a fine needle aspirate of a breast mass, are calculated from the image. Cases are classified as either benign or malignant based on the evidence. There are 13 numerical characteristics that classify each instance of HDD according to the various cardiac conditions that might affect humans. Training data accounted for 70% of the dataset, whereas testing data accounted for 30%. In the future, we can apply this technique to brain tumour disease, as well [28,29,30,31,32,33,34,35,36,37,38,39].

4.1. WDBC Dataset

4.1.1. GRU-RNN

Figure 2 shows the training progress of GRU-RNN for the WDBC dataset.

4.1.2. Secure GRU-RNN

In a similar manner, the results of the secure GRU-RNN are given below. Figure 3 shows the training progress of the secure GRU-RNN for the WDBC dataset.

• HDD Dataset
• GRU-RNN

The results of the GRU-RNN are provided below. Figure 4 shows the training progress of the GRU-RNN for the HDD dataset.

4.1.3. Secure GRU-RNN

The results of the secure GRU-RNN are provided below. Figure 5 shows the training progress of the secure GRU-RNN for the HDD dataset.

4.2. Performance Comparison

4.2.1. WDBC Dataset

Table 3. Parameter comparison of the WDBC dataset.

Models	Precision (%)	Recall (%)	F-Score (%)
SVM [1]	90.47	97.24	-
secure SVM [1]	90.35	96.19	-
GRU-RNN	98.48	97.74	98.11
secure GRU-RNN	94.03	96.92	95.45

presents the comparison table of the accuracy, precision, recall, error rate, and F-score of the existing SVM and secure SVM, along with the proposed GRU-RNN and secure GRU-RNN. The performance comparison graph is shown in Figure 6. From the achieved results, it is observed that the results of the proposed method are better than those of the existing method.

The performance comparison graph of precision and recall for the WDBC dataset is shown in Figure 7, in which the existing SVM and secure SVM have precisions of 90.47% and 90.35%, respectively. The GRU-RNN and secure GRU-RNN obtained precisions of 98.48% and 94.03%, respectively. It was observed that there was an improvement of 8.13% without secure training and 3.91% with secure training for the precision measure obtained, which is superior to others.

The existing SVM and secure SVM had recalls of 97.24% and 96.19%, respectively. The GRU-RNN and secure GRU-RNN obtained recalls of 97.74% and 96.92%, respectively. It was observed that there was an improvement of 0.5% without secure training and 0.7% with secure training for the recall measure obtained, which is superior to others. The F-scores obtained were 98.11% and 95.45% for the GRU-RNN and secure GRU-RNN, respectively.

4.2.2. HDD Dataset

Table 4. Parameter comparison of the HDD dataset.

Models	Precision (%)	Recall (%)	F-Score (%)
SVM [18]	93.35	90.87	-
secure SVM [18]	93.89	89.78	-
GRU-RNN	96.61	95	95.79
secure GRU-RNN	96.29	94.54	95.41

provides a comparison table of the accuracy, precision, recall, error rate, and F-score of the existing SVM and secure SVM, along with the proposed GRU-RNN and secure GRU-RNN. The performance comparison graph is shown in Figure 7. From the achieved results, it is observed that the results of the proposed method are better than others.

The performance comparison graph of the precision and recall for the HDD dataset is shown in Figure 7, with precisions of 93.35% and 93.89% for the existing SVM and secure SVM, respectively. The GRU-RNN and secure GRU-RNN obtained precisions of 96.61% and 96.29%, respectively. It was observed that there was an improvement of 3.37% without secure training and 2.49% with secure training for the precision measure obtained for the proposed method, which is better than other methods.

The existing SVM and secure SVM have recalls of 90.87% and 89.78%, respectively. The GRU-RNN and secure GRU-RNN obtained recalls of 95% and 94.54%, respectively. It was observed that there was an improvement in recall of the proposed method compared to the existing method. It was observed that there was an improvement of 4.35% without secure training and 5.03% with secure training for the recall measure obtained for the proposed method compared to the existing method. The F-scores obtained were 95% and 94.5455% for the GRU-RNN and secure GRU-RNN, respectively.

5. Conclusions

In this research, we presented secure GRU-RNN, an encrypted IoT data-training technique that uses blockchain technology to protect user privacy. Here, IoT data were encrypted before being recorded on a distributed ledger, and blockchain technology was used to create a secure and trustworthy platform for data sharing among multiple data providers. We designed secure blockchain building blocks using the AES cryptosystem, and constructed a secure GRU-RNN training algorithm, which does not require a trusted third party. Finally, a comparison of the precision, recall, and F-scores of the existing SVM and secure SVM were provided along with the proposed GRU-RNN and secure GRU-RNN. The results obtained for the WDBC dataset (structured) showed that the performance of the secure GRU-RNN training scheme performed 3.91% better than the existing scheme in terms of precision and 0.7% better than the existing scheme in terms of recall. The results obtained for the HDD dataset (unstructured) showed that the performance of the secure GRU-RNN training scheme performed 2.49% better than the existing scheme in terms of precision and 5.03% better than the existing scheme in terms of recall. From the achieved results, it was observed that the results of the proposed secure GRU-RNN training scheme performed better than those of the existing secure training scheme for both structured and unstructured datasets. The limitation of the proposed research is using a standard medical dataset. In the future, we will try to solve the limitation of this proposed research by using a real dataset from Saudi Arabian hospitals. Furthermore, advanced deep learning, such as transform, will be suggested to improve the accuracy of the proposed system.