Secure healthcare data sharing and attack detection framework using radial basis neural network

Abstract

Secure medical data sharing and access control play a prominent role. However, it is still unclear how to provide a security architecture that can guarantee the privacy and safety of sensitive medical data. Existing methods are application-specific and fail to take into account the complex security needs of healthcare applications. Moreover, the healthcare sector needs dynamic permission enforcement, extensible context-aware access control, flexible, and on-demand authentication. Therefore, this research proposes an access control mechanism and an effective attack detection model. The proposed authenticate access control mechanism (PA2C) safeguards data integrity as well as the security and dependability of EHR data sharing are improved by the use of smart contracts, encryption, and secure key management. On the other hand, the proposed intelligent voyage optimization algorithm-based Radial basis neural network (IntVO-RBNN) effectively detects the attacks in the network. Specifically, the Intelligent Voyage Optimization algorithm effectively tunes the model hyperparameters and the deployment of hybrid features contributes to the proposed model to detect attack patterns effectively. The comparative results showed that the suggested access control strategy performed better than the current methods in terms of minimal responsiveness of 100.18 s and less information loss of 4.49% for 100 blocks. Likewise, the proposed IntVO-RBNN attack detection model performs better with 95.26% recall, 97.84% precision, and 94.02% accuracy.

Introduction

The term healthcare states to a broad system that includes various components, such as health facilities, sensors, and health. Current technologies have significantly changed healthcare practices, transforming them from traditional to technological methods, for instance using sensors and wearable devices, such as wristbands and smart watches, to monitor healthcare conditions¹. The Internet of Medical Things (IoMT) is beneficial for improving the consistency, precision, and throughput of electronic devices². In addition, the IoMT devices aid in the regular monitoring of patients as well as taking preventive actions on our own. The smart healthcare monitoring system is defined as a group, which are connected within a network via the internet^3,4. Furthermore, depending on third-party data security and privacy measures, the cloud-centric approaches reduce patient control over their medical data. Nevertheless, the centralized approaches result in increased delay during processing and communication as well as high bandwidth traffic. For healthcare applications where transparency high availability, and low latency are essential, centralized architectures are not suitable. Furthermore, because of IoMT limitations, cybercriminals have turned their attention to IoMT-based healthcare systems, where security measures are relatively laxer^5–8.

Thus, several security concerns and threats, including hardware-based attacks, network attacks, software, and system-level attacks, might jeopardize the lives of patients in smart healthcare schemes⁴. To overcome these challenges and transform healthcare systems, however, cutting-edge technologies like deep learning and blockchain present intriguing answers⁹. The majority of next-generation apps have utilized Blockchain (BC) to offer security across an extensive range of platforms, including smart cities, the Internet of Things, and others, throughout the past few years. This is due to the BC, which offers decentralized and trust-free solutions¹⁰, where data is saved over the network in a decentralized fashion using online distributed ledgers¹¹. Blockchain technology can be classified into three types namely community scheme, private, and public blockchain, which utilizes an unchangeable approach that verifies and processes the data between known and unknown nodes. Additional services are offered by several healthcare nodes and added to the workflow for network processing^12,13. As such, it has the potential to improve the systems’ efficacy, safety, and openness for exchanging medical data. BC technology helps medical organizations gain insight and improve the analysis of medical records. A novel processing approach with delay-sensitive monitoring is then required by the healthcare support system, which should be intelligent and stable^14,4.

The BC, which offers decentralized and trust-free solutions¹⁵, stores data throughout the network in a decentralized manner¹¹. The Interplanetary File System (IPFS), one of the few well-established decentralized file-sharing technologies, helps healthcare systems govern file versions and distribution. However, it is extremely challenging to manually handle and analyze the huge amounts of patient data that are produced by healthcare systems in a real-time setting¹⁶. To identify uncommon arrangements and address this fundamental issue, current methods employ a signature-based approach. One such remedy is an all-inclusive intrusion detection system (IDS)¹¹. In Machine learning-enabled IoMT-enabled healthcare approaches, immense healthcare data from various biosensors are used for processing. However, the IoMT encountered research challenges due to the current blockchain-based ICPS^17,18,13. Currently, available blockchain solutions with machine learning capabilities are unable to withstand malware attacks during execution. The primary drawback of current public technologies is their inability to identify runtime malware, which is relatively new and exhibits no patterns¹³.

To tackle the challenges of existing access control and attack detection mechanisms, this research introduces a novel PA2C scheme, which ensures data integrity and the proposed IntVO-RBNN model effectively detects the DDoS attacks. The use of a blockchain^19,20 IPFS storage system safeguards the patient EHR reports. The parameters of the IntVO-RBNN model are effectively tuned using the proposed IntVO algorithm, which improves the accuracy of attack detection. The detailed contributions of the research are mentioned as follows,

Intelligent voyage optimization algorithm enabled radial basis neural network (IntVO-RBNN)

In the proposed approach, the IntVO algorithm simulates the biological characteristics such as intelligent hunting and traveling of coot and osprey, which efficiently optimizes the hyperparameters of the model. In addition, the IntVO algorithm overcomes challenges such as local optima problems and slow convergence issues respectively. Besides, the proposed model is capable of handling inconsistent, noisy, ambiguous, and probabilistic data.

Proposed authenticate access control mechanism (PA2C)

PA2C employs multi-factor authentication to enhance security, which confirms that only authorized doctors can access the data. The system can integrate with blockchain technology to provide an immutable record of all access attempts and transactions that enhances transparency and accountability.

The subsequent sections of the manuscript are organized in the following manner, Sect. 2 covers the literature review with the problem statement. Section 3 describes the system model and Sect. 4 demonstrates the methodological explanation of the proposed access control and attack detection approaches. The results and discussion section are described in Sect. 5 and the research conclusion is provided in Sect. 6 with future works.

Literature review

The following section describes the pros and cons of existing access control mechanisms in smart healthcare applications.

Ashwag Albakri et al.⁴ presented a deep learning-based decentralized secure healthcare system, which offered robust performance over the existing approaches. The deployment of metaheuristic and Bayesian optimization techniques effectively tuned the model parameters. In addition, the image encryption and key generation schemes enabled additional security for the medical data. Despite its performance, the model required a large number of datasets. Aitizaz Aliet al.⁹ implemented a blockchain network-enabled hybrid deep learning approach that improved the transparency, immutability, and data integrity of healthcare information. However, system scalability remained a significant challenge, which limited the real-world implementation of the established approach.

Dhairya Jadav et al.¹¹ designed a blockchain network-enabled trustworthy healthcare management system that identified the malware activities in the network. The implemented Long Short-Term Memory identified the IoT nodes whether it was malicious or not, which enriched the security of the network. The leveraged blockchain layer stored the healthcare data into the immutable ledger. However, this system contains low robust and low-security features of the HEART. Mazin Abed Mohammed et al.¹³. introduced a Pattern-Proof Malware Validation (PoPMV) algorithm block out for blockchain in ICPS. The integration of deep learning and reinforcement learning approaches effectively received the feedback and rewards in real-time, which enhanced processing speed and identified both familiar and unfamiliar attacks. However, without federated Artificial Intelligence (AI) edge analysis, this system consumed the highest energy, time, and storage cost.

K. Raju et al.²¹ utilized a deep learning framework to develop privacy and security in the IoMT environment. The gathered databases were encrypted and the optimal key was created by the hybrid encryption algorithm. However, to prevent malicious attacks, the model required a large amount of data, which impacted the applicability of the approach. To eliminate third-party control over healthcare data Bhaskara S. Egala et al.¹⁶ introduced a distributed smart healthcare system. The established hybrid computing paradigm effectively dealt with computational, latency, and storage constraints. In the real-time environment, Random Forest (RF) was leveraged to find an optimal set of features from the patient’s data and also decision-making tasks performed by Support Vector. However, the distributed Machine Learning module RFSVM takes a long training time for large datasets.

Muhammad Izhar et al.⁸. presented a smart health monitoring and diagnosis system with the use of DLT, AI, and edge computing. This framework leveraged peer-to-peer networks and the need for centralized servers was eliminated while enabling perfect information exchange. The system incorporated a hybrid approach for early identification and forecast of security threats, improving overall effectiveness. DLT involved many levels of encryption that made the system more complicated. The Hybrid systems were made up of multiple setups so it is difficult to manage and error-prone. To ensure network security, Jitendra Kumar Samriya et al.²² introduced Blockchain technology within a cloud architecture. Data was processed and transmitted through cloud architecture to establish a Heterogeneous Autonomous Network (HAN). This network was combined with a Reinforced Neural Network called Cloud RNN, specially designed to classify the data noticed and collected by sensors. The collected data of an autonomous network was continuously monitored and distributed for malicious activity and fault detection. However, here the deep learning-based model was used, so the system lacked trustworthiness and transparency in the results.

1. P. Chinnasamy et al.³⁵ introduced a trustworthy access control framework utilizing smart contracts for attaining security in sharing the EHR between patients and healthcare providers. Further, the trustworthy access control framework serves as a promising solution for secure data sharing in mobility computing without compromising the privacy of personal health information from potential threats. However, smartphone users are restricted from modifying the smart contractual agreements or access controls in the trustworthy access control framework.

2. P. Chinnasamy and Deepalakshmi³¹. presented the hybrid cryptographic access control known as the HCAC‑HER method for secure HER retrieval in the healthcare cloud. Specifically, the HCAC‑HER algorithm utilized the Improved Key Generation algorithm to attain the effective encryption of healthcare data. In addition, the Blowfish algorithm was applied to achieve the optimal key encryption. Further, the model effectively resolved the challenges associated with sharing keys and healthcare data. However, the HCAC‑HER framework lacks blockchain technology-enabled secure storage and sharing system of healthcare data required to be resolved in the future.

Hayam Alamro et al.³². implemented the Blockchain-enabled healthcare system exploiting the Ant Lion Optimizer with the Hybrid Deep Learning (BHS-ALOHDL) technique. More precisely, the BHS-ALOHDL facilitated the healthcare sector in transmitting medical data securely and protected the data from malicious attacks. In addition, the BHS-ALOHDL technique carried out the feature selection utilizing the ALO algorithm resulting in generating the series of feature vectors. Ultimately the flower pollination algorithm (FPA) was adopted in the model for optimal tuning of the HDL model and improved the detection rate.

Adwan A. Alanazi et al.³³. utilized the blockchain with optimal DL-based secure data sharing (BCODLSDSC) technique for improving the healthcare system. Initially, the BCODL-SDSC technique enables blockchain technology for the purpose of storing and maintaining the patient’s data with multiple transactions and provides the access control to the numerous stakeholders. For improving the security of the medical data, the Fractional Order Lorenz system (FOLS) based encryption algorithm was applied in conjunction with the Tuna Swarm Optimization (TSO) algorithm enabled optimal key generation. Nevertheless, the DL model required more computational power for training, worsening the scalability concerns associated with the healthcare system.

Challenges

The key challenges addressed in the existing methods are mentioned as follows,

• The LSTM model required maximum time and energy for performing unfamiliar attack detection¹³.

• The transparency and trustworthiness of the Cloud RNN model were ineffective, which affected the performance of the model²².

• In blockchain network-enabled hybrid deep learning, system scalability remained a significant challenge, which limited the real-world implementation of the established approach¹.

• The Hybrid systems were made up of multiple setups so it is difficult to manage and error-prone⁸.

Problem statement

To compute and manage enormous volumes of data, the architecture of IoT services necessitates a scalable and dependable platform, like the cloud. However, IoT networks and cloud-centric computing lack characteristics including business intelligence (BI), high bandwidth, ubiquitous availability, and latency-aware applications⁵. Instead of using cloud service providers (CSP), IoT applications require a new strategy that allows for local and customized cloud resource management. The IoT network’s data transfers and information transactions are susceptible to several assaults, which raises security and privacy issues. These issues are brought on by a lack of confidence in IoT wireless networks^23,24. Reliable IoT networks that provide risk-free and safe services are crucial. As a result, to tackle these challenges this research introduces an effective Distributed Denial of Service (DDoS) attack detection model as well as a PA2C mechanism for securing patient information.

System model for IoT-blockchain-based smart healthcare system

In recent decades, the healthcare industry is undergoing a transformation driven by advancements in technology, targeting to deal with the difficulties of data interoperability, scalability, and security. Specifically, the smart healthcare system incorporating devices, and sensors endlessly monitor, and exchange healthcare information over insecure public channels. Further, the endless connectivity of devices causes smart healthcare systems susceptible to different security issues such as denial-of-service, impersonation, eavesdropping, and other malicious attacks. Hence, the research proposes the Optimal deep learning-based attack detection model and access control mechanism for secure EHR data sharing in blockchain-based smart healthcare systems to address the challenges associated with efficient data sharing and ensure data protection amongst healthcare service providers. The system model for the IoT-blockchain-based smart healthcare system is delineated in Fig. 1, which depicts the communication occurring in various communicating entities, such as medical IoT devices, edge servers, blockchain storage, and deep learning models for attack detection. Initially, the medical information is collected from K number of patients, which can be mathematically specified as follows,

Fig. 1.

System Model for IoT-blockchain-based Smart Healthcare system.

1

Where P indicates the patient information collected through IoT devices and specifies the patient. The medical information gathered from several departments is stored distinctly, and the corresponding authorized person only has the admittance to access the data for genuine purposes. The medical information stored in the data acquisition phase can be described as follows,

2

Where D denotes the healthcare data, and represents the healthcare data of the patient. Additionally, if the professional requires information about other departments, access should be provided by the relevant department owner. The data access that authorized users are permitted is represented as

3

Here, l indicates the data, and u specifies the corresponding doctor. If the criteria are met, the data access is prohibited because r is illegal access. On the other hand, access will be granted to the appropriate authorized user if the condition is met and access to the data is provided to the corresponding doctor. If the doctor demands for the data access is granted. However, if the doctor requests access to the data , the request is rejected because the doctor must acknowledge it. These conditions are carried out in accordance with the denial takes precedence concept, and the user’s access control is provided by

4

To prevent unauthorized access, the sensitivity label is added to offer additional security to the healthcare data. The sensitivity label is mathematically denoted as,

5

Where indicates the hospital department to which the data belongs, and specifies the name of the health department. The system model comprises of IoT device layer, edge server, blockchain layer, and deep learning model. Initially, the IoT devices collect information from patients and share it with the healthcare service provider through an edge server. The blockchain layer stores the information of the patients and the Deep learning model effectively detects the presence of attacks in the network.

Proposed methodology for IoT-blockchain-based smart healthcare system

The research aims to design an access control and attack detection mechanism based on an optimized deep learning approach. The schematic depiction of the proposed workflow is illustrated in Fig. 2. First, through the verifier, the IoT devices gather patient data and register it with the edge server. Following registration, the Blockchain-enabled security architecture stops unwanted access and confirms the data’s legitimacy. The Proof of Authority (PoA) consensus mechanism and the IPFS, which are components of the blockchain-enabled security architecture, protect data from security breaches and enforce security regulations. Additionally, the optimized Radial basis neural network is used to extract the most important features, including entropy, average flow, growth of flow and port, and land, to detect the attack. The proposed radial basis network checks the transactions as malicious and normal. Moreover, the hybrid optimization algorithm fine-tunes the network parameters thus enabling accurate attack detection.

Fig. 2.

Schematic representation of the proposed model.

Access control scheme in blockchain systems

An access control mechanism in the blockchain system is designed to prevent unauthorized access to the medical information of patients, which enhances data security and integrity. Initially, the health ministry or the hospital establishes the corresponding departments, and the PA2C protocol is intended to verify access to the legitimate user. The health ministry then assigns a specific person to each department. If the patient and the doctor communicate, the data is saved and accessible through the doctor whenever necessary. Table 1 and 2 lists the steps involved in the internal access control method.

Table 1.

List of notations.

Notation	Definition
	Number of patients
	patient
	Patient information
	Healthcare data
	Total number of healthcare data
	Data access for authorized users
	Data access of the corresponding doctor
	Corresponding doctor
	illegal data access
	Identity of the doctor
	Encryption function
	Sensitivity label
	Hospital department
	Name of the health department
	user’s identity
	Public key
	hash function
	prime order
	patient identify
	Security degree
	random number
	private key
	acknowledgment
	security keys
	Address location
	Address location of requested data
	Key of the requested data
	Data owner identity and password
	security key
	Address Index
	authentic key
	entropy
	total number of packets
	Probability of each separate source IP
	packet per flow
	flow features
	single flow growth
	number of flow pairs
	time interval
	variation in port growth
	port number
	feature vector
	Dimension of data
	Number of Hidden layers
	Center of each RBF
	RBF function
	width
	layer weights
	bias
	RBF units
	network’s outputs
	Solutions
	Fitness function
	First two best solutions
	Global average of solutions
	random number in range (0,1)
	conditional factor
	guiding factor
	fitness acquired for best solutions
	best solutions in the previous iteration
	progressive factor
	maximum velocity
	random position of the solution
	Velocity of the random solution
	random number

Table 2.

PA2C protocol.

Setup and establishment of EHR systems

In this phase, initially, the healthcare ministry classifies data regarding several departments. For each department, a legitimate doctor is allocated as a legalized user. Moreover, Information exchanged between the patient and an authorized physician is recorded in the ledger. Only the designated physician from the relevant department has access to the material.

User request and data retrieval in the EHR system

Initially, to access the data the doctor sends the identity , and the system verifies the authenticity of the doctor. Once the verification is successful, the authorized user is permitted to login into the portal. Following the process of verification, the user requests the data, and the system promptly responds with a user inquiry and acknowledges it. After questions are answered, policies are created for both the owner and the user to prevent future legal requirements. Additionally, the users’ level of sensitivity is evaluated at the same time. Users are only granted access to the data if their degree of sensitivity is surpassed. The user receives the data after it has been validated. Figure 3 depicts a schematic perspective of the EHR systems’ setup and establishment phase.

Fig. 3.

Schematic perspective of the EHR systems’ setup and data retrieval phase.

System entitles

The elements of the system include the registration phase, which detects requests, and the blockchain, which is where the encryption and storing process occurs. The system entities ensure secure, transparent, and efficient handling of requests and data storage.

a) Registration phase.

The registration phase is the fundamental process in the blockchain system, in which the user needs to interact with the system must be registered, which provides the necessary information and credentials to create a unique identity. Once registration is completed, the system can identify requests from the users that ensure the authenticated user only accesses the data stored in the blockchain storage system. Data providers and users are ensured to be recorded on the blockchain during the registration process. Consider that if a user z wants to access their patient’s data in the EHR, the patient must give the EHR their user identity to do so. The steps involved in the registration phase are depicted in Fig. 4, which are described as follows.

Fig. 4.

Schematic illustration of user registration and owner registration phase..

Initially, the patient, who is the data provider, estimates the public key and sends it to the blockchain together with the user’s identity. The public key can be calculated as,

6

Where signifies the encryption function, denotes the hash function, and e indicates the prime order. The user identity typically consists of user ID and password, which is sent to the blockchain for verification, once verified, the blockchain securely stores user’s identity. The security factor is then generated by the blockchain storage system in order to determine the private key of the security factor, and the private key of the user can be estimated as follows,

7

The user’s private key is constructed as follows. The data user now receives the private key that the blockchain created. If , the data owner has been properly verified and registered with Trusted Certificate Authority (TCA).

Owner registration phase

During this phase, the owner (patient) must submit it identify to the blockchain. The following steps are involved in this process.

First, the user identification is sent to the blockchain by the data owner, who also evaluates the public key .

8

User ID and password are sent to the blockchain for verification, and the user’s identity is stored as Following that, the blockchain storage system generates the security factor and the private key can be approximated using the following formula,

9

At this point, the data user receives the private key that the blockchain generated. If the private key matches the expected criteria , the data owner is successfully validated. Upon successful validation, the data owner is registered in the blockchain. The registration phase guarantees that the identity of the data provider and the user is securely verified, and their private key is safely generated and transmitted, enabling secure interactions within the blockchain system. Once the registration is successful, users can begin accessing and storing data for which they have signed a smart contract.

b) Contract sign and acknowledgment.

The contract sign and acknowledgment phase serve as a robust and secure method, which involves the creation of a unique digital signature that is shown in Fig. 5. Here, to sign a contract the following procedures are done.

Fig. 5.

Contract sign and acknowledgement phase.

Initially, to create a smart contract the data owner shares its identity and password to the blockchain, and the blockchain verifies the incoming patient is entered into the blockchain database. If j is available in blockchain, it determines a security degree, which can be calculated using the user attributes and credentials. The security degree can be formulated as follows,

10

where denotes a random number, indicates the private key. Similarly, the calculation for is expressed as follows,

11

When the security key , the acknowledgment is shared to the IPFS. Once acknowledged by the blockchain, the security keys F are released,

12

where signifies the security keys which have a range between 1 to N, the security keys are forwarded to IPFS. Furthermore, the address location and corresponding keys are generated and stored in the blockchain. In addition, the patient sends a request for data storage to the blockchain, and the smart contract communicates the data request to the IPFS. The storage system replies to the address location of requested data and shares the key to the data owner. At the same time, the data is managed as chunks and encrypted as follows,

13

Each chunk is encrypted separately and communicated to IPFS through blockchain. The blockchain knows the data location and address keys. Moreover, the address index is shared with the data owner. As a result, the contract sign and acknowledgment phase guarantees that the data owner’s identity is safely confirmed and that blockchain and IPFS are used to store and manage their data securely. The security and dependability of EHR data sharing are improved by the use of smart contracts, encryption, and secure key management.

c) Authentication for access and decryption.

Whenever the new user (doctor) wants to access the data to suggest any disease-related instruction, he needs to be acknowledged through blockchain. Authentication for data access and data decryption verifies the identity of the user, which ensures data confidentiality as well as enhances data security from unau , department , specialization , and hospital ID , to the blockchain. Furthermore, the blockchain checks the user details in its directory. Based on the user request for the private key of the blockchain, a key is generated, which is represented as follows,

14

where specifies the random security integer, denotes the private key, and indicates the public key. To verify the genuineness of the doctor, the authentic key provided during registration and the key generated from the blockchain are compared as follows,

15

If , the doctor is genuine, and the data request from the doctor is informed to the IPFS via blockchain. The blockchain shows the address location and shares it with the doctor as well as releases the key . With the help of security keys, the doctor can decrypt the data from the IPFS storage, the decryption can be represented as follows,

16

Once the doctor gets permission to access the data or dashboard, they will connect the patient for any health assistance. Because a central authority is no longer necessary, the blockchain lowers the possibility of a single point of failure and increases the robustness of the system. The schematic illustration of the Authentication for access and decryption phase is shown in Fig. 6.

Fig. 6.

Authentication for data access and decryption.

Attack detection using radial basis neural network

The hospital portal stores the patient data that has been gathered, and it uses the radial basis network to evaluate the data for attack detection. The network transmission parameters are employed to carry out the attack detection, which facilitates the identification of unidentified network users. Without compromising data quality, the detection would guarantee the safe transfer of medical data over the system.

Feature extraction

For effective attack detection performance, the significant features including the growth of flow, average flow, entropy, port, and Land are extracted, which is explained as follows,

a) Entropy: The entropy features are significant for measuring the randomness of specific attributes in the header of the network packets. The estimated entropy values offer information associated with the detection or identification of attacks²⁵. By examining the series of continuous packets, packet creation rate, source port number, and destination IP address the entropy is measured is mathematically described as follows,

17

Where Q specifies the entropy, n indicates the total number of packets, denotes the probability of each separate source IP address. The entropy of source port numbers is calculated based on the predefined threshold, a significant drop in the entropy of source port numbers can indicate a potential DDoS attack. On the other hand, the packet creation rate’s entropy is defined as the frequency at which packets are generated and sent over the network. In normal conditions, the packet creation rate is normal, while during the DDoS attack, the packet creation rate spikes dramatically. The entropy of the destination IP address is utilized to determine whether the traffic is being directed towards a single target or multiple targets. The value drops with the maximum data flow, indicating the risk of an attack.

b) Average Flow: Since it can be difficult to identify the attacker’s initial source, the average packet flow is assessed to identify the DDoS attack using the median computation, which is represented as,

18

Where g indicates the packet per flow. The average byte/flow is used to calculate the payload size, an increase in value indicates the risk of an attack. A smaller number indicates the fewest false positives. The average duration/flow property assesses the flow time that occurs in the flow table. The flow features improve the attack detection accuracy and can detect anomalies in network traffic patterns that are indicative of DDoS attacks, allowing for timely and accurate detection and mitigation.

c) Growth of flow and port: The growth parameter tracks the increase in data flow over time, for effective attack detection the single flow growth feature is extracted, in which for DDoS attacks the data growth is very high. The single flow growth calculation can be performed as follows,

19

where denotes the number of flow pairs, and specifies the time interval, and a total number of flows is denoted as . The network ports are responsible for sending and receiving the data, unusual activity on a port can indicate an attempt to exploit a vulnerability. The variation in port growth can be calculated as follows,

20

where represents the port number. In addition, the pair flow has the identical source port and destination IP, and the fair flow’s elevation expresses the risk of a DDoS attack, which is calculated as

21

where signifies the fair flow percentage

d) Land: Land is a type of Local Area Network Denial attack in which the attackers repeatedly send TCP packets to the nodes, causing the target system or portal to crash. Therefore, taking land into account as one of the aspects would allow for the tracking of clear messages to the IoT nodes. Collectively, these attributes contribute to the deep learning model to detect patterns that indicate an attack, and the obtained parameters are integrated to form a feature vector, and provided into the proposed IntVO-RBNN model which can be described as,

22

Radial basis neural network architecture for attack detection

The research intended to design an IntVO-RBNN model for effective attack detection, which is made up of several simple processing units, or neurons, that are connected. The IntVo-RBNN model gains problem-solving skills by suitably modifying the linkages’ strength based on incoming data. Additionally, it is easily adaptable to different settings through learning. In addition, it can handle probabilistic, ambiguous, noisy, and inconsistent data²⁶. The traditional attack detection models have faced several challenges regarding model complexity, anonymization services, and IP spoofing. To tackle these challenges, this research integrates the benefits of the IntVO algorithm into the RBNN model that effectively tunes the model’s hyperparameters.

The radial basis function (RBF) concept is one of the most widely used and very effective, which learns the nonlinear mapping between an input space²⁷. In this research, the -dimensional input is sent straight to a hidden layer in an RBF. Assume that the hidden layer has neurons. The Euclidean distance between the input and a -a-dimensional prototype vector determines the activation function applied by each of the neurons in the hidden layer. If the distance between the centerof each RBF and the input vector is equal to zero then the contribution of this point is 1, however as the distance rises, the contribution goes to 0. The distance can be mathematically formulated as follows,

23

There may be multiple predictor variables in the input layer, each of which is connected to a separate neuron, and propagates input to the hidden layers. Numerous RBF units with Gaussian kernels and bias are part of the hidden layer. A center and a width determine the k-th Gaussian function. The IntVo-RBNN classifier does the nonlinear transformation in the hidden layer as follows,

24

Every hidden neuron has a parameter called a prototype vector. Each hidden neuron’s output is then sent to the output layer after being weighted. The weighted hidden layer neurons’ sums make up the network’s outputs, which is mathematically denoted as follows,

25

Where denotes the layer weights, the form of the RBF functions at the hidden units, the number and location of the centers, and the method for determining the network weights all have a significant impact on the IntVo-RBNN model’s performance. The output layer of the proposed model effectively detects the presence of attack with better accuracy. The IntVo-RBNN model is a three-layer network, which is depicted in Fig. 7. The layer weights and biases are optimally tuned using the IntVo algorithm.

Fig. 7.

Architecture of the IntVO-RBNN model.

Intelligent voyage optimization algorithm

Motivation

The IntVO algorithm is designed to optimize the hyperparameters of the RBNN model, which simulates the traveling and intelligent hunting behaviors of osprey¹⁹ and coot²⁸. Similar to other metaheuristic algorithms, the fundamental coot algorithm has a low diversity tendency, a poor convergence speed, and an inadequate balance between exploitation and exploration. The local optimal solution may trap the approach. To expand the fundamental COOT method’s local and global search trends, this research integrates the intelligent hunting behaviors of osprey.

Inspiration

The collective activities of coots and osprey serve as the foundation for the IntVO algorithm. The coots’ collective activities include both regular and erratic motions across the water’s surface. Each individual in the group makes an effort to get closer to the goal of attaining the best target. As a result, they revise their current stances in light of the group leader’s positions. For instance, the random movement to one side and the other, chain movement, positional adjustment based on the group leaders, and leader movement play a significant role in the proposed algorithm²⁹. In addition, the key source of inspiration for IntVO is the method used by ospreys to locate, capture, and move their prey to a convenient location for consumption when hunting fish from the ocean. Mathematical models are used to simulate the IntVO implementation processes in two phases. The clever natural actions of ospreys, such as catching fish and transporting them to a suitable location for consumption, can serve as the basis for the development of new optimization algorithms. As a result, the IntVO algorithm is designed using mathematical modeling of these intelligent osprey behaviors.

Solution initialization

In the IntVO algorithm, the solutions are randomly generated based on layer parameters such as weights and biases. Here, the initialized solution for parameter tuning is mathematically indicated as follows,

26

where M indicates the solutions and

Fitness evaluation

The evaluation of fitness function is the process of assessing the quality of potential solutions to an issue using a fitness function. In the proposed IntVO algorithm, the fitness function aids the algorithm in deciding which solutions to retain and which to discard. The mathematical formulation can be estimated as follows,

27

Solution update

Once all the solutions are evaluated for their strength, the first two best solutions are declared, which remain as the leader for the search and hunt processes associated with the proposed algorithm. Based on the conditional factor, the algorithm enables two phases such as exploitation and exploration, which are detailed in the subsequent section.

Stage 1: Exploitation : When the conditional factor is less than or equal to one, the solution searches for the best position inside the search boundary. In this phase, the solution updation can be performed based on the leading solution. In addition, the use of guiding factor is used to update the position of the solutions through the leaders’ average location

28

where specifies the first two best solutions, denotes the global average of solutions for the previous 3 iterations, denotes the random number, indicates the guiding factor, which can be estimated as follows,

29

where signifies the fitness acquired for best solutions till the current iteration, represents the best solutions in the previous iteration, denotes the progressive factor that drives the solutions to global points, and can be estimated based on the velocity corresponding to the first two best solutions.

30

The above equation denotes the maximum velocity.

Stage 2: Exploration :The exploration phase represents that the condition factor is greater than one, therefore the solution searches out of the search boundary to obtain optimal results. Based on the attacking behaviors of osprey, the position of the solution in the search space altered dramatically, enhancing the algorithms’ ability to explore and find the perfect location while avoiding local optimality. Here, the mathematical formulation for solution updation can be performed using the velocity of the random solution and the random position of the solution respectively.

31

where signifies the random position of the solution, denotes the velocity of the random solution, and indicates the random number .

Termination

When the algorithm stops its iteration and declares the global best solution for effective parameter tuning. The flow diagram of the proposed IntVO algorithm is depicted in Fig. 8.

Fig. 8.

Flowchart of the proposed IntVO algorithm.

Results and discussion

The experimental results of the suggested attack detection model and access control method are shown in the section that follows. Furthermore, the performance of the suggested approach is contrasted with some current approaches, and the outcomes are examined.

Experimental setup

The research execution for access control and attack detection is performed in MATLAB software on a Windows 11 operating system. The system equipped with 16 GB of RAM and 128 GB ROM serves as a valuable tool for experimenting. The initial parameter settings include the batch size of 64, learning rate of 0.01, activation function: “ReLU”, loss function “Categorical cross entropy” and default optimizer Adam. This setup ensures that the research can be executed efficiently, ensuring accurate and reliable results in access control and attack detection. The proposed model utilizes the training and testing data split ratio of 80:20, in which 80% of data is used for training and 20% of the data is utilized for testing.

Dataset description

The dataset utilized for the proposed research is explained in this section.

APA-DDoS dataset

The APA-DDoS dataset³⁰ comprises 151,201 entries, in which every record comprises different characteristics of network connections to assist in identifying the DDoS attacks. The dataset involves 23 characteristics that offer a complete overview of network traffic. These elements include essential network characteristics such as source and destination TCP ports, source and destination IP addresses, along with IP protocol numbers. Further, the dataset offers information about the length of the individual frame and the presence of several TCP flags (SYN, reset, push, acknowledgement) and IP flags.

Performance metrics

In this research, the attack detection efficiency of the IntVO-RBNN model is evaluated in terms of accuracy, False Positive Rate (FPR), recall, and precision. The FPR is used to analyze the false alarm indication, and a negative number is regarded as positive. Precision is the measurement of accurately identified DDoS attacks using the IntVO-RBNN model while taking positive detection into account. Precision refers to the assessment of correctly identified DDoS attacks using the proposed IntVO-RBNN. On the other hand, for access control, the performance of the PA2C scheme is analyzed in terms of responsiveness, genuine index, privacy, and information loss.

Comparative analysis for access control scheme

The performance of the PA2C scheme is compared with the conventional methods such as RFSVM¹⁶, ECC-IBC⁹, BSHS-EODL⁴, BHS-ALOHDL³², BCODL-SDSC³³ and PoPMV¹³ with varying block sizes (50, 100), and number of users ranges from 20 to 100. The detailed discussion of the results obtained with the comparison of different access control schemes is explained in this section.

Comparative analysis for access control scheme with 50 blocks

Figure 9 delineates the comparative evaluation outcomes of the PA2C scheme using 50 data blocks with performance metrics such as responsiveness, privacy, GUD, and information loss respectively. In comparison to the traditional methods, the PA2C scheme achieves superior responsiveness of 100.18 s, which is reduced over the PoPMV by 0.48 s, BHS-ALOHDL by 53.18 s, BCODL-SDSC by 150.23 s, and BSHS-EODL by 5.28 s. Furthermore, the PA2C scheme attains GUD of 95%, achieving the relative improvement of 8.77% over PoPMV, 15.35% over BHS-ALOHDL, 25.69% over BCODL-SDSC, 13.16% over BSHS-EODL, 17.54% over ECC-IBC, and 33.83% over RFSVM. In addition, for 100 users the privacy of the proposed scheme is 95.51%, which surpasses the conventional PoPMV by 1.93%, BHS-ALOHDL by 3.37, BCODL-SDSC by 9.36%, BSHS-EODL by 2.57%, ECC-IBC by 4.16%, and RFSVM by 14.56%. Similarly, the information loss of the PA2C scheme is 4.49%, which is quite minimal to the conventional approaches achieving loss differences of 1.85, 3.22, 8.94, 2.46, 3.98, and 13.91 over PoPMV, BHS-ALOHDL, BCODL-SDSC, BSHS-EODL, ECC-IBC, RFSVM respectively. Moreover, the PA2C scheme obtained superior results compared to the other baseline techniques utilized for the comparison. Specifically, the PA2C scheme exploits the multi-factor authentication boosting the security level and ensuring that only the authorized doctors to access the data. In addition, blockchain technology offers the immutability of the records of all access attempts and transactions that add an additional layer of security thereby improving transparency and confidentiality.

Fig. 9.

Comparative Analysis of access control scheme with 50 blocks.

Comparative analysis for access control scheme with 100 blocks

Figure 10 shows the comparative evaluation results of the PA2C scheme using 100 data blocks in terms of metrics such as responsiveness, GUD, privacy, and information loss. The PA2C scheme provides outstanding responsiveness of 185.47 s, which shows performance differences of 0.26 s, 95.23 s, 279 s, 4.78 s, and 185 s when compared to the conventional approaches including PoPMV, BHS-ALOHDL, BCODL-SDSC, BSHS-EODL, ECC-IBC respectively. Additionally, the PA2C scheme achieves a GUD of 95%, which is improved over the existing methods PoPMV by 7.41%, BHS-ALOHDL by 14.21%, BCODL-SDSC by 24.26%, BSHS-EODL by 11.57%, ECC-IBC by 16.85%, and RFSVM by 31.68%. Furthermore, the suggested PA2C scheme’s privacy for 100 users is 96.50%, outperforming the PoPMV by 0.95%, BHS-ALOHDL by 3.26%, BCODL-SDSC by 4.49%, BSHS-EODL by 2.75%, ECC-IBC by 3.78%, and RFSVM by 5.20%. Similarly, the PA2C scheme’s information loss is 3.50%, which is significantly less than that of the traditional methods PoPMV, BHS-ALOHDL, BCODL-SDSC, BSHS-EODL, ECC-IBC, RFSVM by 0.92, 3.15, 4.33, 2.65, 3.65, and 5.02 respectively. From the overall analysis, the proposed PA2C scheme achieved superior results compared to other existing techniques. Specifically, the incorporation of multi-factor authentication in the PA2C scheme boosts the security level and guarantees that only legitimate doctors can access the medical data. Furthermore, the application of blockchain technology enhances the immutability of the records and transactions improving the overall security level thus boosting the transparency and confidentiality of medical data.

Fig. 10.

Comparative Analysis for access control scheme with 100 blocks.

Comparative analysis for attack detection

The attack detection performance of the proposed IntVO-RBNN is compared with the prevailing methods such as BPHS-STRIDE², SA-BiLSTM³⁴, LSTM¹, HVC⁸, and RBNN²⁶ with varying number of users. The detailed discussion of the results obtained with the comparison of models for attack detection is explained in this section.

Comparative analysis for 50 users

Figure 11 shows the results of the analysis utilizing 50 users in the blockchain in terms of FPR, sensitivity, accuracy, recall, and precision. With an accuracy of 94.01%, the suggested approach outperformed the traditional BPHS-STRIDE, SA-BiLSTM, LSTM, HVC, and RBNN techniques for 90% of training data by 17.03%, 13.20%, 11.93%, 11.47%, and 2.93%. With 90% of training data, the suggested method’s FPR of 4.20% represents a performance enhancement of12.07, 10.29, 8.83, 8.60, and 1.90 over the traditional BPHS-STRIDE, SA-BiLSTM, LSTM, HVC, and RBNN approaches. With a precision of 97.58%, the suggested approach outperformed the traditional BPHS-STRIDE, and RBNN approaches by 18.02% and 1.08% for 70% of training data. With a recall of 95.26%, the suggested approach outperformed the traditional BPHS-STRIDE, SA-BiLSTM, LSTM, HVC, and RBNN techniques by 12.11%, 8.33%, 4.87%, 4.38%, and 2.38% for 90% of the training data. Moreover, the proposed model attained impressive results compared to the other baseline techniques utilized for comparison. More specifically, the radial basis neural network excels in its potential to tackle all types of attacks with high detection accuracy. Additionally, the IntVO algorithm incorporated in the proposed approach addressed the local optima problems and slow convergence issues enhancing the performance of the proposed model.

Fig. 11.

Comparative analysis for 50 users.

Comparative analysis for 100 users

The comparative evaluation of the IntVO-RBNN with 100 users in terms of varying training percentages is delineated in Fig. 12. The IntVO-RBNN method outperforms the conventional BPHS-STRIDE, SA-BiLSTM, LSTM, HVC, and RBNN approaches for 90% of training data by 14.87%, 11.20%, 10.22%, 5.43%, and 23.97%, respectively, with an accuracy of 92.15%. The IntVO-RBNN method’s FPR of 5.56% with 90% of training data shows a performance improvement of 11.19, 9.21, 8.43, 6.48 and 2.54 compared to the conventional BPHS-STRIDE, SA-BiLSTM, LSTM, HVC, and RBNN methods respectively. The proposed method outperforms the conventional methods BPHS-STRIDE by 8.17% and RBNN by 1.37% for 90% of training data, respectively, with a precision of 96.90%. Similarly, the proposed method outperformed the conventional BPHS-STRIDE, SA-BiLSTM, LSTM, HVC, and RBNN approaches by 9.21%, 5.87%, 3.90%, 2.98%, and 2.37% for 90% of the training data, respectively, with a recall of 95.26%. Besides, the proposed model achieved remarkable results compared to the other existing techniques. In the proposed approach, the radial basis neural network effectively tackled all types of attacks with high detection accuracy. Furthermore, the IntVO algorithm effectively fine-tunes the hyperparameters of the classifier boosting the performance of the proposed model.

Fig. 12.

Comparative analysis for 100 users

Comparative discussion

The comparative discussion of the proposed access control mechanism and attack detection model is shown in Tables 2 and 4 respectively. Most of the existing techniques including the PoPMV, BHS-ALOHDL, BCODL-SDSC, BSHS-EODL, ECC-IBC, and RFSVM are found with specific challenges that limited their performance. Existing techniques are encountered with limitations including the inability to take into account the complex security requirements of healthcare applications, challenges associated with sharing keys, and healthcare data limiting the applicability for secure healthcare data sharing. Further, the existing attack detection models required more training time for larger datasets and provided misdetection of the attack instances. However, the proposed system addressed the above limitations in the existing techniques via the application of multi-factor authentication in the PA2C scheme boosting the security level, and ensuring that only legitimate users can access the medical data. Besides, the adoption of blockchain technology augments the immutability of the records and transactions which improves the overall security level. Furthermore, the radial basis neural network effectively identified all types of attacks with maximum detection accuracy. Additionally, the IntVO algorithm adaptively fine-tunes the hyperparameters of the model boosting the performance of the attack detection. Moreover, the obtained results indicate that the proposed system outperformed other existing techniques in terms of the metrics that are depicted in Tables 3 and 4.

Table 3.

Comparative discussion for access control.

Metrics/ Methods		PA2C	PoPMV	BHS-ALOHDL	BCODL-SDSC	BSHS-EODL	ECC-IBC	RFSVM
50 blocks	Responsiveness (sec)	100.18	100.66	153.36	251.41	105.46	201.26	301.56
	GUD (%)	95.00	86.67	80.42	70.60	82.50	78.33	62.86
	Privacy (%)	95.51	93.66	92.29	86.57	93.05	91.53	81.60
	Information loss (%)	4.49	6.34	7.71	13.43	6.95	8.47	18.40
100 blocks	Responsiveness (sec)	185.47	185.73	280.69	464.24	190.25	371.13	557.36
	GUD (%)	95.00	87.96	81.50	71.95	84.01	78.99	64.90
	Privacy (%)	96.50	95.59	93.35	92.17	93.85	92.85	91.49
	Information loss (%)	3.50	4.42	6.65	7.83	6.15	7.15	8.52

Table 4.

Comparative discussion for attack detection model.

Metrics/ Methods		BPHS-STRIDE	SA-BiLSTM	LSTM	HVC	RBNN	IntVO-RBNN
50 users	Accuracy (%)	78.00	81.60	82.80	83.23	91.25	94.02
	FPR (%)	16.27	14.45	13.08	12.80	6.11	4.20
	Precision (%)	89.95	90.00	91.16	91.25	96.94	97.84
	Recall (%)	83.72	87.33	90.62	91.10	93.00	95.27
100 users	Accuracy (%)	78.45	81.83	82.73	87.15	88.50	92.16
	FPR (%)	16.76	14.78	14.00	12.05	8.11	5.57
	Precision (%)	88.99	89.40	89.45	89.48	95.58	96.91
	Recall (%)	86.48	89.66	91.54	92.41	93.00	95.26

Confusion matrix

The confusion matrix is used to assess the performance of the proposed model and compares the predicted class to the actual class for a set of instances. The confusion matrix involves the values of True Positives (TP), True Negatives (TN), False Positives (FP), and False Negatives (FN) allowing for the evaluation of the classifier performance to indicate that an attack instance is correctly detected as anomalous. Figure 13 displays the confusion matrix of the proposed model for attack detection. From the confusion matrix, the proposed model correctly detected the 1226 non-attack instances as normal, and 1225 DDoS attack instances as DDoS attack instances. In addition, the proposed model is found with small number of misdetections. Moreover, the confusion matrix demonstrates the effectiveness of the proposed IntVO-RBNN model in detecting the attacks.

Fig. 13.

Confusion matrix.

Computation complexity analysis

The computational complexity is conducted to analyze the computation efficiency of the proposed PA2C scheme and other existing techniques. The results of the computation complexity are depicted in Fig. 14. For iteration 100, the proposed approach reduced the computation time to 20.58ms and surpassed other baseline techniques. In contrast, the existing techniques such as PoPMV, BHS-ALOHDL, BCODL-SDSC, BSHS-EODL, ECC-IBC, and RFSVM take higher computation time of 20.77ms,20.78ms,20.80 ms,20.81ms, 20.82ms, and 20.83ms, respectively. Specifically, the proposed PA2C scheme requires low computation time with the application of the PoA consensus mechanism that offers high speed of validating the transactions and minimizes the overall computation complexity.

Fig. 14.

Computation complexity Analysis.

Security analysis

The security analysis carried out for the proposed system with the application of different attacks including the Man-in-the-middle (MITM) attacks, and replay attacks is depicted in Fig. 15. While analyzing the proposed system with 100 blocks, the proposed approach takes the responsiveness of 212.42 s for without attack, 294.58 s for MITM attack, and 301.61 s for replay attack. In terms of GUD, the proposed approach achieves better performance gaining the GUD of 61.06% for no attack,39.75% for MITM attack, and 25.96% for replay attack. Similarly, for 100 blocks, the proposed model achieved the privacy of 96.50% for without attack,67.55% for MITM attacks, and 77.20% for replay attacks. Subsequently, the proposed approach is observed with an information loss of 3.50% for without any attack,4.55% for MITM attack, and 4.20% for replay attack. Moreover, the proposed PA2C approach achieved better performance in the criteria of without attacks during medical data transmission. The incorporation of the multifactor authentication mechanism offers secure data transmission even with the existence of the replay attack and MITM attacks revealing the effectiveness of the proposed approach in achieving security in data transmission.

Fig. 15.

Security analysis with 100 blocks.

.

Conclusion

In conclusion, the research introduces a deep learning-based DDoS attack detection model and access control mechanism for secure EHR data sharing in blockchain networks. The PA2C scheme designed for access control eliminates the threats by allowing only authorized and legal hospital personnel to view the data. On the other hand, for effective attack detection, the proposed IntVO-RBNN model is designed, in which, the IntVO algorithm adjusts the weights of the RBNN classifier to improve the attack detection efficacy. Furthermore, the added feature extraction step assists in reducing the computation overhead and improves the smart contract-based attack detection. Additionally, the deployment of blockchain and IPFS storage systems addresses the single-point failure and enhances system resilience. When compared with the existing techniques, the experimental outcomes exhibited that the PA2C scheme offers superior performance in terms of minimum responsiveness of 100.18 s, and less information loss of 4.49% for 100 blocks. Similarly, the DDoS attack detection model exhibits improved performance achieving a high accuracy of 94.02%, precision of 97.84%, and recall of 95.26%. In the future, additional encryption mechanisms will be included in the PA2C scheme, to enhance the security of the shared EHR records.

Author contributions

A.B conducted the overall work, wrote the manuscript including drafting and implementation. P.B carried out implementation and final review. P.S.R and S. A carried out the analysis and drafted the literature survey.

Funding

Open access funding provided by Manipal University Jaipur.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Competing interests

The authors declare no competing interests.