1. Introduction
With the wide utilization of intelligent mobile devices in the fifth-generation (5G) era and the advances in wireless communication techniques in the forthcoming sixth-generation (6G) communication systems [
1], the internet of things (IoT) is increasingly attracting attention from both academia and industry as a versatile technology [
2,
3]. Nowadays, more and more intelligent devices get access to IoT networks, where each object with identifying, sensing, networking, and processing capabilities can communicate with other nodes. Such ubiquitous interconnections generate massive data to be stored, processed, and analyzed [
4,
5].
Recent advances in information and communication technologies are accelerating the IoT’s transition to the 6G era. As a result, new IoT infrastructures and data processing architectures are under construction. Currently, a majority of IoT applications depend on centralized cloud servers to store data and process tasks [
6], which necessitates that the third parties owning the cloud servers are quite trustworthy, otherwise the user data may be exposed to security concerns [
7]. Furthermore, centralized cloud-based applications introduce delay and privacy issues [
8,
9]. Therefore, demands for data security, processing efficiency, and low operational cost are overgrowing in IoT scenarios [
10]. Blockchain [
11], as a decentralized data storage technology, can guarantee that each transaction record is immutable through an encryption algorithm and distributed structure. As a result, blockchain has been presented as a promising technique for enhancing the security and efficiency of data storing/fetching in IoT networks, which can realize tamper resistance and data availability for IoT networks in a decentralized manner [
12,
13,
14].
In the current IoT architecture, blockchain technology has been extensively investigated [
15]. Kang et al. have proposed a blockchain-enabled internet of vehicles framework based on an upgraded delegated proof-of-stake (DPoS) consensus mechanism rather than the native proof-of-work (PoW) or proof-of-stake (PoS) mechanism to improve the security of vehicle data sharing [
16]. Fan et al. presented a blockchain-based strategy for resolving the security issue associated with time synchronization in IoT networks [
17]. W. Li et al. proposed a data security strategy based on blockchain technology for intelligent applications in 6G systems [
18]. However, these schemes focus on the security of block verification and block generation while lacking the consideration of emerging applications and services in IoT networks that require high computational workloads and low latency. Most existing blockchain-enabled frameworks cannot meet the demand for high transactional throughput in IoT scenarios with compute-intensive and real-time applications and services. Therefore, service-oriented blockchain-enabled frameworks are called for to meet the transactional throughput demands of current and next-generation IoT networks.
New IoT services and applications, such as mobile multimedia, visual sensors, smart grids, and intelligent vehicles, are computationally intensive and sensitive to latency, challenging the blockchain-enabled IoT framework design. Conventional cloud computing systems are confronted with the problems of long latency and overload problems, hindering the blockchain deployment in IoT networks. To address the above challenges, the mobile edge computing (MEC) [
19] technique is considered a potential option. Because an MEC system is deployed at the edge of IoT networks and near the access network, it is capable of bridging the divide between the limited resources in the proximity of users and the ever-increasing computational demand of IoT applications [
20]. Therefore, MEC technology can facilitate the development of low-latency, scalable, and blockchain-enabled IoT networks. Recently, several studies [
21,
22] have been proposed to enhance data security and transactional throughput for IoT by integrating blockchain technology and MEC technology. However, most existing schemes focus only on the computational offloading strategy of MEC or the working mechanism of blockchain and lack a comprehensive and specific analysis of the MEC-enabled blockchain system. Therefore, they fail to achieve a joint performance optimization, which leaves room to improve.
The efficient deployment and optimization of MEC-enabled blockchain systems in IoT networks are challenging from several perspectives: trade-off between latency and security, joint optimization, and dynamic continuous domain-based optimization. (1)
Trade-off between latency and security: the blockchain-enabled IoT network is confronted with latency-sensitive challenges and thus requires an efficient blockchain mechanism without compromising security. Additionally, due to system resource limitations [
23] in IoT networks, the collaborative design of an efficient resource allocation policy and a lightweight blockchain consensus mechanism at the edge of wireless networks is challenging. (2)
Joint optimization: the optimization problem formulation in most existing studies [
21,
22] considers the blockchain system and MEC system separately, ignoring the coupling relationship between the blockchain transactional throughput and the MEC computation rate. Due to the lack of joint optimization consideration, the performance in existing studies can be improved. (3)
Dynamic continuous domain-based optimization: the IoT services arrive in complicated stochastic patterns, and most computation tasks of IoT services have sensitive latency requirements, which also pose significant challenges to the joint optimization of MEC-enabled blockchain systems. Additionally, the parameters of MEC computation offloading policy or blockchain optimization strategy are continuous domain variables, while existing works [
21,
22] simplify them to discrete variables. Therefore, we resort to continuous domain-based DRL to realize dynamic and continuous joint control of computation resource allocation and block generation.
To address the aforementioned issues, we present a DRL-based joint performance optimization framework for MEC-enabled blockchain systems in IoT networks, which aims to improve the scalability/throughput while guaranteeing data security and transaction processing efficiency. In particular, each IoT node can offload a portion of computation tasks to MEC servers for efficient data processing, wherein the computation tasks include the IoT application tasks and the tasks for block generation and reaching consensus. Meanwhile, blockchain technology is adopted for secure data storage and sharing inside this framework, with a consensus mechanism based on practical Byzantine fault tolerance (PBFT) and DPoS [
24] being adopted. Furthermore, the performance optimization of MEC and blockchain system is jointly formulated as a Markov decision process (MDP) problem, where state transitions mainly depend on changes in time-varying factors such as the impact of user movement on wireless transmissions, node workload, etc, which are unknown a priori. Moreover, the action is selected based on continuous space. As a result, conventional math models are ineffective in solving the MDP problem. To address this MDP problem, a novel DRL-based algorithm with continuous action space is developed. It shows superiority in tackling dynamic and complicated joint optimization problems. The primary contributions of this paper are listed as follows:
- (1)
A novel MEC-enabled blockchain framework in IoT networks is developed, considering the latency and scalability issues that arose from the throughput requirements of future wireless networks and blockchain systems. We analyze MEC computation efficiency and critical performance indicators of blockchain, i.e., decentralization, latency, throughput, and adversarial fraction, which can guide the joint optimization of the framework.
- (2)
A novel MEC and blockchain joint optimization algorithm is developed for maximizing the computational efficiency of MEC and the transaction throughput of blockchain systems, which is formulated as an MDP problem. In contrast to most existing research [
15,
16,
17] in which the modeling and optimization of MEC and blockchain systems are carried out independently, the block interval, block size, data transaction throughput, power allocation for local execution and task offloading, latency, and security constraints are jointly considered in the proposed algorithm. Therefore, we propose a more comprehensive scheme and address the blockchain deployment challenges in IoT scenarios.
- (3)
The MDP problem is solved using a deep deterministic policy gradient (DDPG)-based learning algorithm to tackle the dynamic and large-dimensional properties of IoT networks that are intractable using classic learning approaches such as
-learning [
25]. In particular, the DDPG-based algorithm enables the joint resource allocation for MEC and blockchain systems in a continuous domain so as to solve the MDP problem with better convergence.
- (4)
Extensive simulation findings demonstrate that the presented performance optimization framework has the capacity to enhance the transaction processing efficiency of MEC-enabled blockchain networks significantly. The superiority of the DDPG-based algorithm over the deep Q network (DQN)-based algorithm [
26] and other conventional schemes is verified.
The remainder of this paper is structured as follows.
Section 2 introduces the related works.
Section 3 provides the preliminaries of blockchain technology and a basic introduction of the consensus protocol based on DPoS and PBFT.
Section 4 describes the system model. In
Section 5, the DDPG-based joint performance optimization framework is proposed, wherein the joint optimization problem is formulated and solved using a DDPG-based approach.
Section 6 evaluates the proposed algorithm in detail and discusses the simulation results. At last,
Section 7 summarizes this paper and looks forward to future work.
4. System Model
In this section, the system model adopted in this work is introduced. As illustrated in
Figure 2, we propose an MEC-enabled blockchain framework for IoT networks, which comprises three parts, i.e., the IoT network including various smart devices, an MEC system including the BS and MEC servers, and a blockchain system based on the DPoS mechanism.
In the IoT network, smart devices, e.g., vehicles, cell phones, security surveillance, etc., collect some ambient data that must be securely stored/processed or shared with other IoT smart devices. As a result, we consider two types of data transactions among smart devices: (1) data storage/processing and (2) data sharing, both of which are stored in the blockchain system for distributed and secure data storage/retrieval. The nodes in the blockchain system are classified into three categories: (1) general nodes (GNs) that consist of all IoT devices, (2) validation nodes (VNs) that are selected out of GNs based on a specific stake distribution according to the DPoS mechanism, and (3) one primary node (PN) that is selected from VNs and authorized to produce blocks at a specific decision epoch. The blockchain system is mainly responsible for the secure storage/retrieval of transaction data from the IoT network. To achieve this goal, the blockchain system must generate blocks and reach consensus, where GNs receive/transmit transaction data from/to other nodes, VNs conduct the blockchain consensus process, and the PN is authorized to generate blocks within a specific time period. Moreover, the MEC system is responsible for sharing the computational pressure in the IoT network and blockchain system to achieve efficient data processing and blockchain consensus. In the following subsections, we will detail the network model, MEC model, and blockchain model, respectively. The notations used in this paper are summarized in
Table 2.
4.1. Network Model
In this paper, we assume that the blockchain system has
M GNs denoted by
and
K VNs. For each IoT node
, it conducts data storage/processing and data sharing within the IoT network. Meanwhile, it acts as a part of the blockchain system. Specifically,
K VNs, denoted by
,
, are selected out of
in terms of particular rules [
41]. These VNs are responsible for collecting, validating, and packaging the transactions generated by smart devices into a block. Furthermore, this new block is appended to the blockchain after the PN broadcasts the block proposal to other VNs and a consensus is reached.
As shown in
Figure 3, the MEC-enabled blockchain framework is implemented using a discrete-time model in which time is partitioned into multiple decision epochs, and the
n-th epoch,
, has
basic time slots which has an identical duration
and is indexed by
. Thus, each decision epoch
n has a dynamic duration
, where
varies for each decision epoch
n and is determined by the block interval of PN
. For each decision epoch
, the wireless channel condition, task arrival, and power allocation policy of each GN are varied. Therefore, aiming to balance data processing efficiency, security, and average energy consumption, each VN needs to determine the block size, block interval, and task offloading policy in each epoch. Assume that the PN produces a new block with block size
and block interval
in turns within the
n-th epoch. Specifically, block size
indicates the number of bits included in a new block produced by the PN at the end of epoch
n, while block interval
indicates the time required for the VN to generate a new block in each epoch
n.
In the proposed framework, we analyze a 5G macro-cell base station (BS) having
antennas that handle the uplink communications of numerous IoT nodes with a single antenna using linear zero-forcing (ZF) detection [
42], which is simple and efficient [
43]. In this study, we assume that the number of antennas at the BS exceeds the number of mobile nodes, i.e.,
. For each time slot
, we denote the channel vector of each GN
m as
, and therefore the nearest BS’s received signal can be represented by:
where
denotes the uplink power for GN
m to offload transaction tasks with the upper bound
,
denotes the complex data symbol. In addition,
is a noise vector where
denotes the variance and
is a
identity matrix. Furthermore, we adopt the following Markov block fading auto-regressive model [
44] to define the temporal relation between decision epochs and movement for each GN
m:
where
denotes the normalization of correlation function between time slots
and
in terms of Jake’s fading spectrum, and the error vector
is complex Gaussian and independent identically distributed with
. It is worth noting that
and
denote the Doppler frequency of GN
m and first-order Bessel function, respectively.
The
channel matrix between the considered 5G macro-cell BS and
M GNs is represented by
. The linear zero-forcing detection is derived by
. After applying the ZF detector, each node’s signal to interference-plus-noise ratio (SINR) is calculated by [
43]:
where
denotes the
-th item in matrix
.
For each decision epoch n, we assume that the channel condition and the power allocation (i.e., and ) is consistent throughout one decision epoch and updated at the first time slot of each epoch.
4.2. MEC Model
In this subsection, we will show how each GN
m makes use of an adaptive compute offloading policy to support blockchain-enabled IoT networks.
(bit) is the number of computing tasks during the decision epoch
n, which is assumed to be processed since decision epoch
. In addition, we consider that
is independent and identically distributed throughout different decision epochs and there is an average task-arrival rate
based on Poisson distribution. In general,
is considered as ordinary application tasks, while for the node elected as PN at decision epoch
n we consider
as block-generation tasks. Furthermore, the processing of computation tasks is considered fine-grained [
45]. Therefore, within the
n-th decision epoch,
denotes partial bits of computation tasks that are allocated to be processed locally, while
represents some other bits that are transferred to and processed by the edge server. We denote
as the queue length of the GN
m’s task buffer at decision epoch
n’s commencement, and then
can be expressed by:
where
and
.
4.2.1. Local Computation
This section demonstrates the number of data bits handled locally in relation to the power allocated for local computing
. By chip voltage adjustment based on the dynamic voltage and frequency scaling technology [
46], the central processing unit (CPU) frequency (Hz) scheduled for the time slot
is expressed as:
in which
denotes the effective switching capacitance of the chip, which varies according to its architecture [
46]. Additionally, we have
and
, which is the highest permitted CPU frequency of GN
m depending on system capability. Consequently, the number of locally processed bits during the time slot
is calculated by multiplying the time (s) by the computation rate of the device CPU (bit/s). Specifically, the computation rate of the device CPU (bits/s) is derived by multiplying the device CPU frequency (cycles/s) by the number of task bits that the CPU can process per cycle (bits/cycle). That is,
where
(cycles/bit) denotes the number of CPU cycles necessary for GN
m to compute one data bit and it is measured and determined with offline measurement [
47].
4.2.2. Edge Computation
To start with, we assume that the MEC server can handle different computation tasks with a minimal processing delay due to adequate computational resources such as a high-frequency multicore CPU. In addition, as a result of the small-sized computation output, it can be assumed that the feedback delay between BS and node can be ignored. Based on (
3) and with the uplink communication power
, the number of task bits offloaded by GN
m within the time slot
is calculated by:
where
W denotes the system bandwidth.
Therefore, the computation rate (bits/s) of GN
m during time slot
is calculated by:
According to the MEC computation model, each IoT node with stochastic task arrivals can utilize the nearby MEC server to process the compute-intensive tasks efficiently and adjust the ratio of computation offloading dynamically.
4.3. Blockchain System
In this subsection, we present the considered blockchain system. The blockchain system can provide data security and privacy guarantee to the IoT networks as an overlaid system. Each node inside the blockchain system is capable of collecting data transactions from the IoT networks, while only a few with a large number of stakes are elected as VNs for packaging and validating blocks. We assume that
denotes the set of stakes IoT nodes hold based on the BFT-DPoS consensus algorithm [
24] during the
t-th decision epoch. The stake of each GN is updated when consensus is reached on a new block, and the latest stake distribution is aggregated to the BS. For the start of each decision epoch
n, the BS distributes the latest stake distribution
it has recorded to all GNs.
In the following part, the details of the most significant criteria for evaluating the blockchain system performance are described.
4.3.1. Decentralization
To prevent block packaging and verification power from being monopolized, it is necessary to quantify the degree of decentralization to ensure the long-term fairness of the blockchain system [
41]. We make use of
Gini coefficient, which has been widely used to measure wealth or income inequality [
48], evaluate “contrast intensity” [
49] and capture “system inequality” [
50] in existing works. Focusing on the decentralization of VNs, we take the stake distribution of VNs into account in this study. Therefore, the Gini coefficient of stake distribution can be derived to characterize the decentralization:
is within
in which the extremes 0, 1 denote the perfect uniformity and maximal inequality among stake values, respectively. Hence, to guarantee that VNs’ stake distribution of a blockchain system is decentralized, the Gini coefficient
should satisfy the constraint:
where we have
. For simplicity, we assume that
is always satisfied in this work.
4.3.2. Latency Time to Finality and Throughput
The concept of latency time to finality (LTF) is adopted to characterize the latency of the blockchain system, which is the latency that one data transaction record becomes irreversible once it has been committed to the blockchain system [
15]. For latency-sensitive applications, it is important to guarantee that the latency is within the user’s tolerance. The LTF
including the time cost for block validation
and generation
is expressed as:
where
denotes the consensus latency which includes the time cost for packet transmission
and packet verification
that includes message authentication codes (MACs) generation, request signature, and MACs verification [
51]. Note that the latency required for one consensus process in the simulation of this paper is in the order of milliseconds. The majority of real-world mobile IoT devices (e.g., cell phones, smartwatches, etc.) generally have a small displacement within 10 ms. Therefore, we assume that the primary node and consensus nodes usually can provide stable services in one consensus process.
As introduced in Section II, each decision epoch
n has a dynamic duration
. We assume that the duration of decision epoch
n is determined by the LTF
, and thus we have
As in [
52], the blockchain transaction throughput of the proposed framework within the
n-th decision epoch is derived by:
where
denotes average transaction size.
is the block size and derived by:
where
P denotes the PN.
4.3.3. Adversarial Fraction
To ensure the blockchain system’s security performance, it is vital to prevent transactions from being unilaterally tampered with or reversed. The adversarial fraction of hashing power that an adversary can control without endangering system security is one kind of fundamental performance measurement of a blockchain system [
52]. For the PBFT-based consensus protocols, unambiguous finality can be reached under the assumption that less than a
fraction of the nodes are adversarial [
24]. Therefore, the following constraint needs to be satisfied:
in which
f denotes the number of adversarial VNs. For simplicity, in this work, when there are
K VNs involved in the consensus process, we assume that
. In other words, it is assumed that the above constraint is always met in this work.
5. DDPG-Based Performance Optimization Framework
In the following part, we investigate the optimal block generation and computation offloading policies to maximize the transaction throughput of the proposed framework under the constraints of latency and system resources. The joint optimization problem is formulated as an MDP. Moreover, in order to deal with the dynamic and large-dimensional properties of the above-mentioned systems, we develop a DRL-based scheme. Specifically, elements included in the action space (i.e., power allocation and block interval) are continuous variables, which motivates us to employ the DDPG algorithm so as to achieve better performance than the DQN-based approach with a discrete action space [
53].
The architecture of the DDPG-based framework is shown in
Figure 4 and the DDPG agent is implemented in each GN. To deploy the framework, we first define the problem formulation and construct the state space, action space, and reward function. Then, we design a DDPG-based algorithm to solve it as follows.
5.1. Problem Formulation
Due to the difficulty of obtaining the state transition probabilities and reward values in advance which are related to user mobility, node workload, etc., the optimization problem is constructed as an MDP. As mentioned above, to learn the resource-aware dynamic block generation and computation offloading strategies, we propose maximizing the blockchain transactional throughput and node computation rate while guaranteeing the security and decentralization of data storing/processing with the constraints of computation resources and latency. In other words, each GN
m needs to solve the following optimization problem in each decision epoch:
where
is the upper bound of
,
denotes the weight factor for integrating the two objective components, and
is a mapping factor which ensures that the blockchain transactional throughput and the MEC total computation rate are of the same order of magnitude.
The constraints and specify latency and power consumption limits, respectively. Note that, for satisfying the latency requirement of IoT applications, it is assumed that each block should be published and validated within consecutive time slots.
A decentralized dynamic performance optimization strategy will be learned separately at each node, which determines the block interval and power allocation for both local computing and edge computing, depending on the local observation of the environment. Note that the DDPG-based online learning process is totally model-free, which means this algorithm does not require each node to have prior knowledge of the blockchain and MEC systems.
5.2. State Space
It is worth noting that collecting a full observation of the system for all nodes and then distributing them to each node requires high system overheads. Therefore, it is assumed that each node’s state is decided by the observation of the system from its own perspective to avoid high system overheads and make the framework more scalable.
For each decision epoch
n, the workload of each GN’s task buffer
is updated in terms of (
4). Meanwhile, GN receives one message from the BS conveying the stake distribution
and the latest SINR of GN to BS
. In addition,
for the forthcoming uplink communication is calculated according to (
2). As a result, the state space is defined as:
in which the projected power ratio after ZF detection
is calculated by:
In addition, ZF detection projects the received signal
into a space orthogonal to the one spanned by channel vectors of other nodes so that GN
m’s offloaded symbols can be decoded without inter-stream interference [
54].
5.3. Action Space
According to the state
, each GN
m will independently select an action
which includes block interval
, the allocated power for local computing
and the allocated power for computation offloading
. Consequently, the action space in decision epoch
n can be defined as:
where we have
,
and
. Note that the output action of the DDPG algorithm directly maps the states to the optimal power allocation and block generation policy in a continuous action space, which is different from other conventional DRL algorithms where the output is the probability distribution across a discrete action space. Therefore, the dimensional disaster can be avoided in the DDPG algorithm [
53].
5.4. Reward Function
Considering that each node agent’s behavior is incentive driven, the reward function is important to the convergence of DDPG algorithm. According to the objective of our joint performance optimization problem defined in (
16), we construct the reward function
which GN
m receives after decision epoch
n as:
Furthermore, note that the value function of node agent
m starting from a random initial state
under the policy
can be expressed by
where
is the discounting factor in the Bellman equation. The value function
can be used to quantify the performance of the policy
via an infinite horizon and discounted MDP [
55] at node agent
m. The following average transactional throughput
will be maximized by implementing the optimal block generation and computation offloading policy
.
5.5. DDPG-Based Algorithm Design
To solve this problem, we provide a model-free and DRL-based approach for finding the optimal block generation and computation offloading strategies jointly. Moreover, owing to the continuous action space of our MDP model, which is intractable by using traditional learning methods, we suggest a DDPG-based approach to address this problem. The DDPG algorithm is a widely used model-free and off-policy algorithm for continuous action spaces. This subsection first introduces the basic mechanism of the DDPG algorithm and then describes the approach to solve the considered problem.
5.5.1. DDPG Background
As shown in
Figure 5, DDPG is a DRL framework that includes two main networks: (1) the actor network and (2) the critic network. Specifically, the actor network is trained for generating the current policy, whereas the critic network is trained for evaluating the advantages and disadvantages of the current policy.
Specifically, the critic network uses neural networks to simulate real
Q-table to circumvent the curse of dimensionality. Given the current state
, the action
and the deterministic policy
, we can write the action-value function as:
where
represents the expectation distribution for
and
.
Similar to the DQN algorithm [
26], the critic function
is updated by minimization of the loss function that can be written by:
where
and
represent the distribution of state
and action
, respectively.
is given by:
where
represents the deterministic policy at epoch
.
The actor function
can map a state
s to a deterministic action
a in a continuous space. Based on the critic function, the policy’s updating gradient of the actor is calculated using the chain rule:
Therefore, based on (
24) and (
26), the actor and critic networks’ parameters are softly updated in terms of
and
where
is the soft update rate.
5.5.2. The Proposed Algorithm
The detailed DDPG-based optimization algorithm is demonstrated in Algorithm 1 which is implemented in Tensorflow [
56]. The algorithm terminates after a preset maximal number of steps
every episode. For each training episode, the beginning state
is initialized randomly. For each decision epoch
n, each GN will accumulate and preserve a transition
into its own experience memory
. Meanwhile, a random sample of
Z transitions
from
will be utilized to update the node’s own actor and critic networks. After the predefined maximum episodes
, each GN will autonomously learn the resource-aware adaptive block generation and computation offloading policy.
Furthermore, at the testing phase, each node agent will directly load the model learned during the previous training phase, and then interact with the environment, beginning with an empty data buffer
. Similarly, its current state is determined by local observations of the environment and the corresponding action is selected in terms of the output of the actor network.
Algorithm 1 DDPG-based Optimization Framework for MEC-enabled Blockchain IoT Systems. |
- 1:
for each GN do - 2:
Initialization: replay memory , critic network , actor network and corresponding target networks and with weights and ; - 3:
end for - 4:
for each episode do - 5:
Initialization: state for each GN ; - 6:
for each decision epoch do - 7:
for each GN do - 8:
Select action based on the exploration noise to decide block interval and power allocation; - 9:
Observe reward and next state ; - 10:
Store transition data into replay memory ; - 11:
Sample a mini-batch of Z transition tuples from memory at random; - 12:
Update critic network by minimizing the loss L:
- 13:
Update actor policy based on the sampled policy gradient:
- 14:
- 15:
end for - 16:
end for - 17:
end for
|
7. Conclusions
In this paper, we studied an MEC-enabled blockchain system for future wireless IoT networks and investigated the joint performance optimization problem of blockchain transaction throughput and MEC computational efficiency. The joint problem of MEC computation offloading and block generation policy was formulated, where the power allocation and block interval were optimized. As a result of the time-varying properties of wireless channels in this system, we modeled the joint optimization problem as an MDP. A DDPG-based algorithm was proposed to solve the MDP problem, which can cope with the problem under continuous action space and learn the optimal strategy without prior knowledge of the environment.
The simulation results have demonstrated the proposed scheme outperforms DQN-based and some other greedy (non-joint optimization) schemes under different task arrival rates, thresholds of LTF, and weight factors. The joint optimization scheme can achieve better performance than other schemes with a high convergence rate. Meanwhile, we evaluated its performance in terms of power consumption and block interval, and the simulation results have shown that the joint optimization scheme slightly compromises the average block interval to obtain the lowest power consumption. Additionally, we discussed the impact of the number of mobile users on the convergence performance and the joint optimization scheme always has an advantage over other schemes.
In conclusion, the proposed scheme paves the road for the efficient deployment of blockchain technology in IoT networks, which can be applied in latency-sensitive and security-sensitive IoT applications (e.g., internet of vehicles, virtual reality, smart healthcare, etc.).