Today, the Internet based on 5G and 6G has enabled the deployment of billions of IoT devices [11] . It implies that a massive amount of data will be generated from IoT devices, creating room for big data. Unfortunately, most IoT devices are controlled by central systems, which are prone to high costs of communication, storage, security, and privacy concerns. Additionally, robust algorithms are required for aggregating data on IoT platforms. To this end, federated learning provides a promising way to solve the limitations mentioned above. Federated learning is a data-driven machine learning paradigm that ensures collaborative learning between different participants without disclosing sensitive information about them. Thus, it minimizes the costs of storage, communication, and maintaining privacy. Federated learning has certain advantages for the IoT, such as scalability, improved model performance, and privacy preservation. Scalability is achieved via federated learning, where multiple IoT devices can leverage limited computation resources, including hardware, storage, etc., in a parallel manner, especially for low-bandwidth IoT devices. Due to the distributive nature of federated learning, more devices can join the network without incurring extra costs on a centralized server. Since a single IoT device may be resource-constrained, it is possible to have insufficient data to train a high-quality model. With federated learning, a single IoT device can collaborate with other devices to train high-quality models without exposing privacy. It means that raw data cannot leave devices during the model training process while model update parameters are shared between participants and the server. A review in [12] suggested systems heterogeneity and statistical heterogeneity to be challenges associated with federated learning for the Internet of medical things (IoMT). The system heterogeneity challenge occurs when each device in federated learning has distinct hardware, processing, communication, and storage capabilities. It limits the maximum efficiency when deploying federated learning and may increase laggard mitigation and fault tolerance. Statistical heterogeneity occurs when the number of data points on different devices absolutely differs, which may increase model complexity in analysis and evaluation and laggard risk. The authors in [13] presented a framework based on federated learning for healthcare IoT. The proposed framework is suitable for decentralized databases while achieving quality of service (QoS) and privacy. However, issues of scheduling and coordination in federated learning are not addressed. To this end, the authors in [14] proposed a privacy protection scheme for federated learning under edge computing (PPFLEC). The scheme preserves privacy using shared secrets and weight masks. Also, it protects devices against collusion attacks and equipment dropping while ensuring consistency and integrity using digital signatures. However, the deployment of neural network models is expensive and challenging. To resolve the challenge, a fixed-point quantizer with stochastic rounding is adopted, but how to achieve the minimum square quantization error is not resolved. The work in [15] proposed federated learning with minimum square quantization error (FedMSQE) to minimize quantization error for every participant in the federated learning. However, how to resolve the problem of malicious participants degrading the model’s quality by sharing low-quality data is not addressed. Therefore, the work in [16] addressed the problem by proposing an approach-based clustering to utilize social content data for selecting participants. Here, different groups of edge participants were established using group-specific federated learning. Aggregation is performed using models of different edge groups to achieve the robustness of the global model. However, intrusion attacks from different edge groups are not addressed. To this end, the authors in [17] proposed a method based on federated learning to detect unwanted intrusions among participants. This method guarantees the privacy and security of the local training models while sharing gradient parameters with the central global server. Afterwards, the server aggregates and shares the improved detection algorithm with the participants. However, the memory and computational costs of training unlabeled data on a cloud server are not addressed. The limitation was tackled by [18] via the federated learning-based person movement identification (FL-PMT) system. Deep reinforcement learning is used in the system to auto-label the unlabeled data, which is then used to train the model. In this case, the edge server allows parameters to travel via the cloud instead of the sensor data. In addition, bidirectional long short-term memory (BiLSTM) is employed in a number of smart healthcare system operations to categorize data. However, how to determine the contribution rate of each participant is not considered. The authors in [19] presented IoT for the healthcare system based on federated learning. Instead of using dataset size to estimate the contribution rate, the method uses the qualities of each participant’s datasets. Furthermore, the dropout-tolerable strategy is used to stop the federated learning process, particularly if the number of online participants is equal to or greater than the predetermined threshold. Furthermore, the proposed method is impervious to attacks, including model inversion and reconstruction. However, how to improve latency is not considered. The study in [20] makes use of clustered federated learning and edge computing to diagnose COVID19. The proposed method trains a multi-modal machine learning model that can diagnose COVID-19 in both X-ray and ultrasound images, enabling intelligent processing of visual input at the edge. However, the distribution and appearance of health data have unknown classes. Therefore, the authors in [21] proposed a scalable and transferable classification system, known as SCALT. The system is a one-classifier-per-class-based federated learning system that consists of a one-dimensional convolutional network used for feature extraction and an individual mini-classifier for every class. When a new class emerges, scaling is simple because only a mini-classifier needs to be trained. Only when it is moved to a new task does the feature extractor get updated. Another work in [22] combined deep neural networks, federated learning, and mutual information (MI) for effective feature selection and extraction. The proposed method can be used for anomaly detection of intrusions in the IoT network. Besides, IoT devices store information locally for model training, while models’ weights are modified and shared with a centralized server. However, none of the authors in [11] - [22] solved the problem of global model aggregates in a synchronous and asynchronous federated learning context.

The IoT paradigm allows a significant number of physical devices to be connected to the Internet, where massive amounts of data are generated. The data generated from IoT devices is useful for training high-quality machine learning models, more specifically deep learning models. It means that patterns can be easily inferred and appropriate intelligent decisions can be derived from data via learning models. Unfortunately, the IoT paradigm is centralized, which makes it vulnerable to privacy leakage, issues of scalability, security threats, etc. Federated learning, among many solutions, reduces privacy leakages through collaborative training without disclosing sensitive information, but it is still prone to security threats, trust issues, and single points of failure or attack. By integrating blockchain technology, federated learning for IoT systems can be holistically se-cure, and problems of trust and single points of attack can be resolved [23] . A survey in [24] discussed the current challenges of blockchain-based federated learning for IoT applications, which include privacy leakage; lightweight blockchain for federated learning and IoT that should balance between privacy and security, storage and communication cost, and scalability and power consumption; lazy clients; stragglers; statistical heterogeneity; system heterogeneity; unsupervised federated learning; and artificial intelligence (AI)-enabled smart contracts. Additionally, an efficient real-time privacy policy is required to secure participants’ data in healthcare IoT applications. Therefore, the authors in [25] presented a blockchain and federated learning-empowered secure architecture for privacy preservation in smart healthcare. Also, blockchain-based cloud platforms are employed for privacy and security. However, personalized healthcare demands cannot be met by a one-size-fits-all model due to the diversity of health problems among patients. A blockchain-enabled personalized federated learning system that enables users to train personalized models without directly uploading sensitive data was proposed in [26] as a solution to this issue. However, only a subset of IoT devices with limited training data are selected to perform federated learning tasks due to their low budget. The problem was overcome by the authors in [27] , who presented a blockchain-based federated learning market to decentralize federated learning via blockchain, maximize the amount of training data to maximize budget, and provide data for devices with limited resources. They proposed the quality-oriented task allocation algorithm (QTA) to assign suitable devices to complete federated learning tasks while optimizing training quality under a set budget and the trust-based collaborative learning approach (TCL) for data sharing among trusted devices. In order to fend off attacks from malevolent devices, an encrypted model training scheme (EMT) employs countervailable differential privacy technology. Furthermore, a fair reward distribution is guaranteed using the proposed contribution-driven DPoS consensus mechanism.

However, the heterogeneity problem is not resolved. The work in [28] addressed the het-erogeneity problem in federated learning by designing systems, model tiers, and data to reduce heterogeneity and propagate high-quality models for each federated client. Be-sides, blockchain-enabled federated learning lowers latency and consumption while pre-serving privacy. Two technological challenges, namely how to provide correct global model aggregation via a centralized federated learning server and how to develop techniques that encourage federated learning clients to donate their computer resources and time, were not solved. The authors of [29] addressed the issues by putting forth a decentralized solution based on federated learning and blockchain. They also proposed a trust-decentralized loop federated learning consensus protocol to manage resources within IoMT. During the aggregation process, suitable features are chosen using a hybrid weighted-leader exponential distribution optimization algorithm, which suggests that multiple features exhibit different degrees of variation across each feature. These chosen features are then sent to the training phase through the proposed pyramid squeeze attention generative adversarial networks in order to categorize the data as positive and negative. To achieve more efficiency in blockchain-based federated learning, the authors in [30] integrate blockchain and federated learning in a fog-IoT network. The proposed approach utilizes the distributive structure of the fog-IoT network to produce an adaptive network for IoT devices while preserving privacy. However, the problems of latency, da-ta sufficiency, traceability, and privacy concerns due to the increasing number of heterogeneous data are not solved. The authors in [31] proposed a federated learning-enabled blockchain-based framework, known as PPFchain, to guarantee the security and privacy preservation of IoT devices. In PPFchain, cryptographic primitives and a federated learning model are employed to ensure privacy in off-chain fog nodes, while blockchain is used to achieve low-cost and high performance in the network. Another work in [32] presented a technique based on blockchain and deep learning to preserve the privacy of electronic health records. In this technique, a convolutional neural network (CNN) is employed to distinguish between normal and abnormal users, while blockchain-based federated learning processes abnormal users and removes them from the database along with their accessibility of the health records. But because unscrupulous individuals can readily tamper with model updates, federated learning is susceptible to this kind of attack. The authors of [33] created a chameleon hash scheme with a changeable trapdoor (CHCT) for secure federated learning in industrial IoT environments in order to solve this limitation. The use of trapdoors was subject to a number of restrictions under the planned plan. Furthermore, a redactable medical blockchain (RMB) is an implementation of the CHCT system. However, a blockchain-based federated learning approach may not be the most plausible solution, as it is easy for models to be manipulated and intercepted during transmission between users and servers. Therefore, the authors in [34] proposed a system that ensures the security of the transmitted model between every component in the federated learning. Besides, the model is encrypted with Rivest-Shamir-Adleman (RSA), data encryption standard (DES), and advanced encryption standard (AES) algorithms, while the checksum is determined using a hash function, which is stored along with a private key in the blockchain. However, there are needs to improve existing solutions for safeguarding and effective administration of sensitive data generated by IoT devices. The authors in [35] designed an optimized data management and secured federated learning (ODMSMFL) system while integrating blockchain. The proposed system is capable of addressing unique requirements of IoMT, such as decentralized data ad-ministration, federated learning, and security. Besides, data management is enhanced using blockchain, which facilitates retrieval, adequate storage, and exchange of data without disclosing privacy.

The IoT paradigm allows interactions among systems, processes, software, and technology over the Internet. It provides room for achieving the system’s efficiency but increases factors such as time of operation, energy consumption, delay, and workload, which may lead to conflict among them. To mitigate this problem, the factors can be formulated as a single or multi-objective optimization problem. In [36] , the authors provided a multi-objective-based evolutionary algorithm where rapid mutation operators and multi-objective differential evolution (MODE) are employed to address the problem of stagnation of the local optimum. They considered the objectives of IoT services, such as load, energy consumption, delay, and service cost, as the basis for the multi-objective optimization problem. Pareto front in the local space provides sufficient diversity and accelerates the rate of convergence. However, because of the resource constraints of IoT devices, task completion increases along with delay. Therefore, the authors in [37] presented an improved multi-objective Aquila optimizer (IMOAO) with Pareto front to offload tasks from devices to fog nodes while minimizing delay. Opposition based learning (OBL) is used to improve the IMOAO algorithm and achieve sufficient diversity. How-ever, improper deployment of IoT devices on either a federated learning system or fog computing can lead to resource and bandwidth waste, a rise in energy consumption, and a poor QoS level. Thus, in [38] , the authors proposed a mechanism to reduce bandwidth wastage, energy consumption, and single points of failure. Also, a multi-objective is formulated to minimize both energy consumption and delay between each component of the IoT network. Besides, a combinatorial optimization problem is solved using the multi-objective cuckoo search algorithm (MOCSA). However, an improper estimation model for control and monitoring of end-to-end communication and sensing is caused by the energy constraints of IoT devices. Also, it is challenging to achieve QoS requirements in IoT networks. Therefore, in [39] , multi-objective optimization for QoS routing method is proposed to distinguish traffics while deriving better data communication. Also, a mechanism based on energy-efficient priority-based multi-objective QoS routing (PMQoSR) is designed to ensure QoS and energy in the IoT networks. The whale lion fireworks optimization method with fitness function routing (WLFA) mechanisms is an optimization technique with three hybrid algorithms that the proposed system uses to control the routing performance based on QoS criteria. The WLFA uses priority label and time delay patterns while transferring data to the destination in order to minimize localization errors, prevent congestion, and choose the shortest way via the network. However, IoT device selection, which is an NP-hard problem, needs to be solved. The authors in [40] formulated the IoT device selection problem as a multi-objective problem and incorporated OBL in the general framework of a multi-objective evolutionary algorithm-based on decomposition (MOEA/D). Also, convergence and diversity are enhanced using the many-objective algorithm, known as the non-dominated sorting-based genetic algorithm incorporated with OBL (NSGA-III/OBL). However, there is a challenge with IoT service placement in fog. Therefore, in [41] , the authors provided a conceptual framework based on fog-cloud control to optimize IoT service placement. An automated planning model is formulated to manage service requests because of the heterogeneity of IoT resources and applications. Besides, automated evolutionary-based particle swarm optimization was employed to solve the IoT service placement problem while maximizing resources and improving the QoS of fog. However, achieving optimal security is challenging in the IoT network. To this end, the authors in [42] improved a meta-heuristic-based clustering protocol to achieve optimal communication. Here, an optimized deep neural network (ODNN) is used to detect malicious IoT devices based on their energy characteristics. Energy characteristics are determined using optimal cluster head selection, optimal routing, and neighborhood-based spider monkey optimization. However, due to the dynamic nature of the IoT environment, intrusion detection systems are paramount. In [43] , the authors presented a multi-objective evolutionary CNN to detect intrusion. In the proposed approach, CNN is a classifier to detect intrusions, while MOEA/D is used to simplify the parameter tuning process of CNN. Specifically, MOEA/D simultaneously optimizes the two competing goals of the CNN model, such as detection performance and model complexity. It is achieved through a novel encoding scheme that converts CNN’s topological architecture into a chromosome. It allows MOEA/D to produce a variety of intrusions with different CNN model detection performances and complexities. Similar work in [44] designed an IoT feature extraction CNN (IoTFECNN) to detect anomalies in the IoT network. Besides, a binary multi-objective enhanced Capuchin search algorithm (CSA) called BMECapSA is proposed to efficiently select features. However, there is a lack of collaboration between application containers and resource allocation problems in the IoT-based cloud. The work in [45] presented a container model named Band-area application container to express in reality the variety of things. Also, an artificial fish swarm algorithm is proposed for optimizing container-enabled task scheduling. Another work in [46] presented multi-objective combinatorial convex optimization to minimize execution cost while blockchain-enabled cost-efficient scheduling algorithm framework to address deadlines and security challenges in IoMT. However, it is more challenging to satisfy the requirements of industrial IoT systems. Therefore, the authors in [47] presented a multi-agent deep reinforcement learning-based offloading method to satisfy the various requirements of different tasks in cloud-edge device computing.

The proposed EINFO optimization algorithm consists of a population of x vectors in N dimensional search space and a k number of decision variables, which corresponds to the number of clients in the federated learning. The random population is defined as follows:

$x=\min \left(x\right)+\left(\max \left(x\right)-\min \left(x\right)\right)\times r,$(15)

where $\min \left(x\right)$ and $\text{max}\left(x\right)$ are the lower and upper bounds of the population, $r\in \left[0,1\right]$ is a random number, and there are two control parameters: weighted mean factor $\delta =\left(2\beta r-\beta \right)$ where $\beta =c\exp \left(-d\frac{y}{ma{x}_y}\right)$, r is a random value,

$ma{x}_y$ is the maximum number of generations, and scaling factor $\sigma$. The control parameters are employed to amplify the obtained vector through an updating rule that depends on the size of the search space. Moreover, the control parameters can be tuned dynamically on the basis of population generation.

In this stage, a new vector is created using the updating rules. The updating rules increase population diversity during the search procedure. We modified the mean-based rule of [54] using the inverse square law. The mean-based rule is derived from the weighted mean of the random vectors, where it begins with the initial population and moves to the next solution using the weighted mean information. Moreover, the updating rule accelerates the convergence rate, which helps reach the optimum solution and improves the algorithm’s performance. The mean-based rule is designed on the basis of the best $x_{bs}$, better $x_{bt}$, and worst $x_{ws}$ solutions. The mean-based rule is defined as follows:

$MR=rW{M}_1+\left(1-r\right)W{M}_2,$(16)

where $r\in \left[0,0.5\right]$ is a random value.

$W{M}_1=\delta \frac{w_1{\left(x_{a1}-x_{a2}\right)}^2+w_2{\left(x_{a1}-x_{a3}\right)}^2+w_3{\left(x_{a2}-x_{a3}\right)}^2}{\left(4\pi {\left(x_{ws}-x\right)}^2\right)\left(w_1+w_2+w_3\right)+ϵ}+ϵr,$(17)

where $ϵ$ is a constant with small values and $f\left(x\right)$ is the fitness function such that $a_1\ne a_2\ne a_3$

$w_1=\cos \left(\left(f\left(x_{a1}\right)-f\left(x_{a2}\right)\right)+\pi \right)\exp \left(-\frac{f\left(x_{a1}\right)-f\left(x_{a2}\right)}{\omega }\right)$(18)

$w_2=\cos \left(\left(f\left(x_{a1}\right)-f\left(x_{a3}\right)\right)+\pi \right)\exp \left(-\frac{f\left(x_{a1}\right)-f\left(x_{a3}\right)}{\omega }\right)$(19)

$w_3=\cos \left(\left(f\left(x_{a2}\right)-f\left(x_{a3}\right)\right)+\pi \right)\exp \left(-\frac{f\left(x_{a2}\right)-f\left(x_{a3}\right)}{\omega }\right)$(20)

$\omega =\text{max}\left(f\left(x_{a1}\right),f\left(x_{a2}\right),f\left(x_{a3}\right)\right).$(21)

$W{M}_2=\delta \frac{w_1{\left(x_{bs}-x_{bt}\right)}^2+w_2{\left(x_{bs}-x_{ws}\right)}^2+w_3{\left(x_{bt}-x_{ws}\right)}^2}{\left(4\pi {\left(x_{ws}-x\right)}^2\right)\left(w_1+w_2+w_3\right)+ϵ}+ϵr,$(22)

where

$w_1=\cos \left(\left(f\left(x_{bs}\right)-f\left(x_{bt}\right)\right)+\pi \right)\exp \left(-\frac{f\left(x_{bs}\right)-f\left(x_{bt}\right)}{\omega }\right)$(23)

$w_2=\cos \left(\left(f\left(x_{bs}\right)-f\left(x_{ws}\right)\right)+\pi \right)\exp \left(-\frac{f\left(x_{bs}\right)-f\left(x_{ws}\right)}{\omega }\right)$(24)

$w_3=\cos \left(\left(f\left(x_{bt}\right)-f\left(x_{ws}\right)\right)+\pi \right)\exp \left(-\frac{f\left(x_{bt}\right)-f\left(x_{ws}\right)}{\omega }\right)$(25)

$\omega =\max \left(f\left(x_{bs}\right),f\left(x_{bt}\right),f\left(x_{ws}\right)\right).$(26)

This study adds convergence acceleration to the updating rules of the proposed EINFO optimization algorithm to enhance global search capability via the best solutions. The convergence acceleration is defined as follows:

$CA=r\frac{x_{bs}-x_{a1}}{\left(f\left(x_{bs}\right)-f\left(x_{a1}\right)\right)+ϵ}$(27)

Therefore, the new vector $x_{n\in w}$ is defined as:

$x_{n\in w}=g\left(x\right)+\sigma MR+CA.$(28)

Table 3 shows the performance of the proposed EINFO. The table shows that the best cost and execution time of the algorithm are used to evaluate performance. After around 30 iterations, the proposed EINFO exhibits the smallest optimal cost and surpasses the current approach in terms of average objective function values. The proposed EINFO’s performance shows that the algorithm is capable of doing appropriate exploration and search. Additionally, as Figure 2 illustrates, there is a suitable balance between exploration and exploitation, preventing early convergence. By utilizing the inverse square law of Equation (14), an improvement over the work in [54] , the updating rule enhances both local and global search. However, there is a tradeoff between best cost and execution, as presented in Table 3 .

The convergence analysis presented in Figure 2 indicates that the proposed EINFO avoids premature convergence because it can search both locally and globally in solution space, finding solutions with a high density in the vicinity of the global optimal and a low density in the vicinity of the global optimal. It indicates that the proposed EINFO can effectively discover the optimal solutions by examining interesting regions in the search space that have the best cost.

Table 4 shows the performance of the proposed system in terms of accuracy, sensitivity, and specificity metrics. Specificity has a true negative rate and reveals more about the model than the accuracy metric, especially if the number of true positive and true negative instances is imbalanced. It is also similar to specificity [56] . This study compares the proposed model with multi-layer perception (MLP) [57] , radial basis function neural network (RBFNN) [58] , and extreme learning machine (ELM) [59] .

In the table, it is observed that the proposed system model outperforms existing models in terms of accuracy, specificity, and sensitivity with higher values. The performance is achieved because of the optimized gradient parameter of the model using the proposed EINFO optimization algorithm. Moreover, the binary classification of the model is enhanced.

In the proposed system, there is a tendency that blockchain can be split into different states, which creates rooms for compromise, known as forks. It implies that there are disagreements among delegates. This type of scenario allows two or more delegates to solve a nonce of blockchain simultaneously [60] . The probability that a fork will happen follows a Poisson, which is as follows:

$P_f=\alpha \left(\exp -\lambda \left(L-1\right)\gamma \right)+\left(1-\alpha \right)\exp \left(-\lambda \left(L-1\right)\right)\gamma ,$(29)

where $\gamma$ is block propagation delay and $L=\left[10,20,30\right]$ is the delegate; $\lambda =\left[0.01,0.1\right]$ is the expected time to generate blockchain, and $\alpha$ is the degree of honesty. The block propagation delay of the proposed blockchain-based federated learning is defined as:

$\gamma =\frac{b_h+t_{r}bs}{\aleph }$,(30)

where $b_h=200$ Kbits is the block header size, $t_r=5$ Kbits is the transaction size, $b_s=1000$ is the block height, and $\aleph =\left[1,5,20\right]$ Mbits is the capacity of the transmission link. As shown in Figure 3 , we assume different values of $\lambda ,\aleph$and L. It is observed that the probability of fork creation slowly approaches zero for each value of $\lambda$ and $\aleph$. It implies that a new block can be propagated if the value of $\alpha \ge 0$. Furthermore, only delegates with high reputation scores can mine and propagate new blocks. As the number of delegates increases, the probability of having a fork approaches zero, irrespective of the client’s capacity for transmission

Using Equation (11), it is observed in Figure 4 that the credibility status approaches one as the number of clients increases along with $\alpha$. It implies that delegates have a high level of reputation. Furthermore, credibility status is directly proportional to reputation scores. Hence, the proposed DPoS consensus protocol is efficient in resolving the problem of in-discriminate propagation of new blocks.

In Figure 5 , the evaluation of the proposed system model in terms of communication latency is shown for different numbers of clients. It is observed from the figure that $P_f$ and $\gamma$ are considered in computing the communication delay.

We consider different communication delays (in percent) to analyze the proposed system. Besides, when the $N=10$ and communication delay is 10% and above, the communication response time is minimal because fewer clients are used for both validators and delegates. It implies that communication delays among clients are minimized. Also, as N continues to increase, the proposed system achieves

better response times for different values of communication delay. It shows the efficacy of the proposed system in terms of communication delay.