Received August 18, 2019, accepted October 1, 2019, date of publication October 14, 2019, date of current version

October 28, 2019.

Digital Object Identifier 10.1109/ACCESS.2019.2947306

A Trust-Aware Task Offloading Framework

in Mobile Edge Computing
DEXIANG WU 1, GUOHUA SHEN 1,2,3 , ZHIQIU HUANG 1,2,3 , YAN CAO1 , AND TIANBAO DU1
1 College of Computer Science and Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China
2 Collaborative Innovation Center of Novel Software Technology and Industrialization, Nanjing 210093, China
3 Key Laboratory of Safety-Critical Software (Nanjing University of Aeronautics and Astronautics), Ministry of Industry and Information Technology, Nanjing

211106, China
Corresponding author: Guohua Shen ([email protected])
This work was supported in part by the National Natural Science Foundation of China under Grant 61772270, in part by the National Key
Research and Development Program of China under Grant 2018YFB1003902 and Grant 2016YFB1000802, and in part by the Key
Laboratory of Safety-Critical Software, Ministry of Industry and Information Technology of China under Grant 1015-XCA1816403.

ABSTRACT Task offloading in Mobile Edge Computing (MEC) is a solution to augment resource-limited
mobile devices’ capabilities by migrating tasks to the edge of the network (i.e., edge servers and idle devices).
At present, a lot of work is focused on optimizing policies to reduce latency or energy consumption for users.
However, they mostly ignore that services are not necessarily trustworthy because the resource providers are
complex, dynamic, and unreliable. The trustworthiness of a service in our paper mainly includes two aspects.
One is that resource providers will not violate users’ privacy. The other is that resource providers will perform
well to ensure the effectiveness of services. To solve this problem, we propose a trust-aware task offloading
framework. The main purpose of the framework is to select a resource provider for a user to reduce latency
or energy consumption and ensure service trustworthiness at the same time. The framework can be divided
into three modules (i.e., trust evaluation, filtering and selection). By combining trust evaluation and filtering
modules, some resource providers that are not trusted by users are filtered out to ensure that the services
provided to users are trustworthy. In the selection module, we select an appropriate provider for a user from
the qualified (i.e., left after the filtering process) resource providers based on an offloading policy. The
experimental results show that our framework not only reduces latency or energy consumption for users, but
also reduces the failure rate of tasks.

INDEX TERMS Mobile edge computing (MEC), task offloading, trust evaluation, machine learning, privacy
protection.

I. INTRODUCTION low-latency. However, cloud computing cannot meet the

In the past decade, cloud computing has developed rapidly above two requirements at the same time [4].
and brought many benefits. Computation offloading in cloud To address this challenge, Mobile Edge Computing (MEC)
computing allows users to offload computationally inten- as a complement to cloud computing has been proposed to
sive tasks to a resource-rich remote cloud [1]. With the solve the shortcomings of cloud computing. By sinking com-
rapid growth of smart mobile devices such as smartphones puting power to the edge of the network and leveraging a mul-
and wearable devices, more and more intelligent mobile titude of end-user devices to complete computation tasks [5],
applications such as face recognition, virtual reality, health it promises to provide users with low-latency services.
surveillance, and real-time translation are emerging [2]. Task offloading in MEC is a solution to augment resource-
These applications usually run resource-hungry algorithms limited mobile devices’ capabilities by migrating tasks to the
(e.g., GPU rendering and deep learning) [3], which would edge of the network. D2D (device-to-device) communication
require intensive computing and high energy consump- has become a key technology of 5G, and it is widely added
tion. In addition, these emerging applications often demand to mobile edge computing. At present, a lot of research has
been carried out on task offloading in MEC, and good results
The associate editor coordinating the review of this manuscript and have been achieved in reducing latency and energy consump-
approving it for publication was Vaibhav Rastogi. tion. However, only a little work has focused on whether

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/
VOLUME 7, 2019 150105
 D. Wu et al.: Trust-Aware Task Offloading Framework in MEC

the services provided are trustworthy. Offloading in MEC is The framework can be divided into three modules (i.e., trust
more complex than that in cloud computing. The resource evaluation, filtering, and selection). In the trust evaluation
providers in cloud computing are usually several large com- module, we introduce identity trust and behavior trust to
panies, while the resource providers in MEC can be either evaluate resource providers, and establish a novel trust evalu-
MEC servers or idle devices around the network. Therefore, ation mechanism based on machine learning. In the filtering
resource providers in MEC have new characteristics such as module, we filter out some resource providers with low trust
dynamic, complex, and unreliable, which make it difficult to level by filtering algorithm based on the results of trust eval-
ensure that services are trustworthy. uation. Based on the two modules, we are able to ensure the
The trustworthiness of a service in our paper mainly trustworthiness of services. In the selection module, we select
includes two aspects. One aspect is that resource providers the appropriate one from the qualified resource providers for
will not violate users’ privacy, which we refer to as Priv − the user to perform the task according to the offloading policy.
trustworthiness. The essence of task offloading is the out- The main contributions of the paper are as follows: (1)
sourcing of computation, so the privacy of users must also 1) To ensure the Priv − trustworthiness, we introduce
be protected in the process. When a task contains sensi- task sensitivity to measure the threat to the user’s pri-
tive information, it will pose a threat to personal privacy if vacy if the sensitive information contained in the task
such information is acquired by malicious devices and used is exposed. We introduce online social networks into
maliciously. Zhang et al. [6]–[8] pointed out that since user MEC, which maps the trust relationship established by
data in mobile edge computing is usually processed in semi- device owners in the social layer into the identity trust
trusted authorized entities, it is likely to cause data leakage relationship between devices. If a provider’s identity
or data loss. The other is that resource providers will perform trust level is lower than task sensitivity, it is untrustwor-
well to ensure the effectiveness of services, which we refer thy. To ensure the Effe − trustworthiness, we introduce
to as Effe − trustworthiness. If the number of misbehav- behavior trust to evaluate the behavior performance of
ing providers increases, there will be some bad situations, resource providers. Users have the required behavior
such as high task failure rate or actual latency far beyond levels of services. If a provider’s behavior trust level
prediction, which will disrupt offloading and lead to low is lower than the user’s required behavior level, it is
offloading effectiveness. For example, a resource provider untrustworthy.
who only wants to benefit from task offloading does not 2) Based on the relationship between task sensitivity
contribute resources or reduces a large number of resources and identity trust, as well as the relationship between
without agreement. If we do not consider the trustworthiness behavior level required by the user and behavior trust
of services, even if offloading policies achieve good results level of a provider, a trust-based filtering algorithm is
in the optimization of latency or energy consumption, it is proposed to filter out those unqualified providers.
difficult to apply task offloading in practice because of the 3) We propose a trust-aware task offloading framework to
low effectiveness or the potential threat to users’ privacy. reduce latency or energy consumption and ensure the
To ensure the trustworthiness of services (i.e., Priv − trustworthiness of services for users at the same time.
trustworthiness and Effe − trustworthiness), we conduct a The remainder of this paper is organized as follows:
trust evaluation of these two aspects. On the one hand, to pro- Section II briefly introduces the related work. Section III
tect a user’s privacy, we introduce task sensitivity to measure describes the relevant preliminary knowledge. Section IV
the threat to the user’s privacy caused by the exposure of sen- introduces our trust-aware task offloading framework in
sitive information contained in the task, and introduce identity detail. Section V introduces the experimental evaluation of
trust to evaluate the trust degree of the user to providers. our framework. Finally, we make a summary and put forward
When the task sensitivity is higher than the identity trust level the direction of future research.
of a provider, it indicates that the provider is not trusted by
the user to complete the task. On the other hand, to ensure II. RELATED WORK
the effectiveness of services, we introduce behavior trust To select trustworthy resource providers, we need to establish
to evaluate whether providers are misbehaving. Users have a trust evaluation mechanism. Meanwhile, facing a large
their required behavior levels of services. When a provider’s number of resource providers, we need to select an appro-
behavior trust level is lower than a user’s required behavior priate resource provider to complete a task according to an
level, it is not trusted by the user to complete the task. offloading policy. Therefore, in this section, we focus on trust
However, it is difficult to choose an appropriate weight for evaluation and task offloading policies in MEC.
each trust factor in establishing a trust evaluation mechanism.
Therefore, we establish a novel trust evaluation mechanism A. TRUST EVALUATION IN MOBILE EDGE COMPUTING
by formalizing trust evaluation as a classification problem in Trust evaluation has received extensive attention and research
machine learning. in areas including cloud computing, mobile cloud computing,
Overall, to reduce latency or energy consumption and and peer-to-peer computing [9]–[12]. At present, people are
ensure the trustworthiness of services at the same time, building some new trust evaluation mechanisms based on the
we propose a trust-aware task offloading framework. new characteristics of MEC.

150106 VOLUME 7, 2019

D. Wu et al.: Trust-Aware Task Offloading Framework in MEC

According to the sources of trust factors, we classify the in the case of limited computing resources of MEC. The
existing evaluation work in MEC into three categories: indi- author proposed a layered MEC deployment architecture and
rect trust models (based on reputation), direct trust mod- solved the multi-user offloading problem by using Stackel-
els (based on direct interaction records), and hybrid models berg game theory. The author’s proposed offloading scheme
(both direct and indirect trust factors). Hussain and Almourad not only provides a significant improvement in reducing
[13], [14] studied how to compute the reputation of edge latency over local execution, but also guides the deployment
data centers, using centralized trust management, which of MEC servers. Liu et al. [20] introduced a cloud-assisted
stores the reputation of LTE deployed clouds. Through this edge computing framework with a three-tier network, and the
system, users can anonymously evaluate cloudlet services. energy consumption minimization problem is formalized as a
Zhou et al. [15] proposed a method to learn users’ prefer- mixed integer programming. Dong [21] proposed a dynamic,
ences based on their contexts, previous behaviors, and social decentralized resource allocation policy based on latency and
intimacy. He designed a trust evaluation mechanism based on energy consumption, which is used to deal with the prob-
users’ preferences to guarantee trustworthy edge computing. lem of task offloading between multiple heterogeneous edge
To enhance the reliability of trust evaluation, we will take into nodes and central clouds.
account both indirect trust factors and dircet trust factors. In addition to the above-mentioned offloading policies to
According to the technology used, we divide the exist- reduce latency or energy consumption from the perspective
ing evaluation work in MEC into two categories: emerg- of selecting the appropriate resource provider, there are other
ing technology-based models and formula-based models. interesting research on optimizing offloading policies from
Xu et al. [16] presented a novel trustless crowd-intelligence the perspective of communication and computation cooper-
ecosystem based on the common decentralization feature of ation. Li et al. [22] exploited the idea of computation repli-
mobile edge computing and blockchain technology, which cation which allows each user to offload its task to multiple
can ensure that all trustless objects in the whole system edge nodes to speed up the downloading phase via trans-
must perform trusted operations. The formula-based models mission cooperation. Their results show that computation
always assign a weight to each trust factor and integrate them replication is very useful for reducing communication latency
into a value. Yuan and Li [17] made some discussions on in tasks where the output data size is larger than the input
trust management in mobile edge computing, and elaborated data size. On the contrary, sometimes multiple users can use
the reasons and significance of trust management. He pro- collaborative properties to share part of a user’s computation
posed a trust computing mechanism based on multi-source tasks. Al-Shuwaili and Simeone [23] leveraged the inherent
feedback (MSTrust). collaborative properties of Augmented Reality (AR) appli-
However, in the traditional formula method, how to give cations and proposed a novel resource allocation approach
each trust factor an appropriate weight is a very difficult over both communication and computation resources. In this
problem. Although the value derived from the formula-based way, different users are able to share part of the computation
approach can be used as a reference for the trust level, it is tasks as well as the input and output data. Our research is
not accurate enough to be in accordance with the user’s to optimize the policy by selecting an appropriate resource
expectations because trust is a subjective metric. Besides, provider for the user to complete the task.
they are not easily understood by users directly. It is necessary Based on the above analysis, the offloading policy in
to provide users with an accurate and intuitive trust evaluation our framework aims to select the appropriate resource
method. Compared with the above solutions, trust evaluation provider for the user, taking into account latency and energy
in our framework is formalized as a classification problem in consumption.
machine learning.
Based on the above analysis, we will take hybrid trust III. PRELIMINARIES
factors into consideration and utilize a classification approach Corresponding to the two aspects of service trustworthiness,
in machine learning to establish a trust evaluation mechanism we need to introduce task sensitivity, identity trust evaluation,
for mobile edge computing. and behavior trust evaluation.
Task sensitivity and identity trust evaluation are introduced
B. TASK OFFLOADING POLICIES IN MOBILE to ensure the Priv − trustworthiness. We compare the task
EDGE COMPUTING sensitivity level with the identity trust level of a provider to
So far, task offloading has gained a lot of researchers’ atten- protect the privacy of the user. If the task sensitivity is higher
tion. An offloading policy is an important part of offloading than the identity trust level of a provider, it indicates that the
process, determining which tasks will be offloaded and where provider is unqualified to complete the task.
they will be offloaded. Behavior trust evaluation is introduced to ensure the
From the existing work, the objectives of offloading poli- Effe − trustworthiness. We compare the behavior level
cies are mainly divided into three types: reducing latency, required by the user and the behavior trust level of a provider
reducing energy consumption, and balancing latency and to ensure the effectiveness of services. If the behavior trust
energy consumption [18]. Zhang et al. [19] proposed an level of a provider is lower than the user’s required behavior
optimal offloading scheme with the goal of reducing latency level, the provider is unqualified. These providers who are

VOLUME 7, 2019 150107

 D. Wu et al.: Trust-Aware Task Offloading Framework in MEC

B. RELATED THEORIES IN SOCIAL COLLABORATIVE

MOIBLE EDGE COMPUTING
1) TASK SENSITIVITY
Our paper refers to the Health Insurance Portability and
Accountability Act (HIPAA) [24] standard to determine
which personal information is sensitive. Here are some
related concepts about task sensitivity.
• Sensitivity of personal information: It is the extent to
which users are concerned about providing private data.
It measures the damage to users when sensitive infor-
mation items are exposed. The higher the sensitivity
of personal information, the greater the harm to users
when it is exposed. SI is a finite set of personal infor-
mation items. For any personal information item s ∈
SI , the function f (s) obtains its sensitivity. We divided
the sensitivity of personal information into three levels.
In other words, f (s) ∈ {1, 2, 3}. The values from 1 to
FIGURE 1. Social collaborative mobile edge computing.
3 represent low, medium, and high, respectively. If s is
level 3, it means that when s is exposed, it will pose a
not qualified in identity or behavior will be removed. In this threat to the user’s property or life. If s is level 2, it means
section, we will explain the above concepts, which are pre- that when s is exposed, it will cause discomfort to the
requisites closely related to our framework. user. If s is level 1, it means that when s is exposed, it will
have little impact on the user. For more detailed informa-
A. DEFINITION OF SOCIAL COLLABORATIVE MOIBLE tion, Weible [25] analyzed which personal information
EDGE COMPUTING is more sensitive to users. Sensitivity is suggested by
Online social networks have become an indispensable part relevant experts and can be changed according to the
of people’s daily life and have been widely used in many user’s own requirements.
scenarios, such as friend recommendation and access control • Sensitive information set: It refers to a set of several
system based on trust. sensitive information items.
To protect the privacy of users in task offloading, we intro- • Relevant sensitivity: It refers to the harm degree to
duce online social networks into MEC. We call this new users when two personal information items si and sj are
paradigm, which combines mobile edge computing with exposed simultaneously. It can be obtained by the func-
online social networks, Social Collaborative Mobile Edge tion F(si , sj ). F(si , sj ) ∈ {1, 2, 3}. Values ranging from
Computing (SCMEC). Because mobile devices, wearable 1 to 3 represent low, medium, and high, respectively.
devices, and so on are owned and carried by people, Relevant sensitivity is given by experts after reasoning
it is hopeful and promising to construct a secure col- and analysis, and is not able to be modified by users.
laborative environment by leveraging social relationships • Task sensitivity: A task Tt consists of a piece of code
formed by people’s social interactions. For one thing, it is and some input data to perform a specific function.
able to increase people’s willingness to contribute their There are some personal information items in the input
resources; for another, based on social relationships, selecting data. The personal information items that need to be
trusted nodes for task offloading enhances privacy protection collected for a task to perform a specific function are
for users. predetermined and limited in number. We determine
As shown in Fig. 1, the diagram is divided into a device the sensitivity of a task by analyzing the sensitivity of
layer and a social layer. In the social layer, device own- these information items. The sensitivity of a task can be
ers establish trust relationships with other people through calculated by using (1), and n is the number of personal
social applications and have different trust levels for dif- information items contained in the task. 1 ≤ i ≤ n and
ferent people. Everyone can own one or more devices. In i + 1 ≤ j ≤ n. When a task contains only one sensitive
the device layer, the trust relationships between devices information item, task sensitivity is the sensitivity level
correspond to the trust relationships between their owners of the information item. When a task contains multiple
in the social layer. A device can establish a D2D connec- sensitive information items, it is necessary to consider
tion with another device within a certain range. Devices the relevant sensitivity level of each two items and
can request services from nearby MEC servers or nearby the respective sensitivity levels of the two items. Task
devices to complete tasks. It should be noted that, if there sensitivity is the maximum of these three levels in all
is no special explanation, the MEC mentioned below refers combinations. In particular, when a task does not involve
to SCMEC. any personal sensitive information, the sensitivity of the

150108 VOLUME 7, 2019

D. Wu et al.: Trust-Aware Task Offloading Framework in MEC

task is 0, that is, the task is not sensitive. If there are n TABLE 1. Elements of identity trust feature vector.
personal information items needed in a task, the number
of relevant sensitivity composed of each two personal
information items is Cn2 . Therefore, the computation of
task sensitivity has a low complexity O(n2 ).

0
 if n = 0
St = f (s1 ) if n = 1 (1)
MAX F(si , sj ), f (si ), f (sj ) if n >= 2

 

Generally speaking, if a personal information item is

exposed, it will not cause too much harm to a user. How- from 1 to 3 indicate that the trust levels are low, medium, and
ever, when multiple personal information items constitute high, respectively. After constructing the trust feature vector
an information set, it will pose a great threat to the user’s v(m, n) from the above eight factors, we need to map it to a
privacy. Therefore, the sensitivity of the information set must discrete trust level ranging from 1 to 3. We use ITC(v(m, n))
be considered. to express the mapping from a trust feature vector to a trust
level, that is, It(m, n) = ITC(v(m, n)).
2) IDENTITY TRUST EVALUATION
We associate the device layer with the social layer in SCMEC. 3) BEHAVIOR TRUST EVALUATION
We map the trust relationships established in the social layer Behavior trust evaluation is a method to evaluate providers’
to the device layer, and when a task is sensitive to the user, behavior performance, which is used to predict the effective-
we only select one of the resource providers trusted by the ness of services. Due to the new characteristics of mobile
user to offload the task. Therefore, we need to finish the trust edge computing, there will be some misbehaving providers,
evaluation in online social networks. which will disrupt offloading and lead to low offloading
At present, there has been a lot of work on trust com- effectiveness. Therefore, we need to establish a trust evalu-
puting in online social networks, which can be divided into ation mechanism for their behavior and performance in this
three categories: structure-based models, interaction-based process.
models, and hybrid models [26]. Yin et al. [27] proposed However, it is difficult to choose an appropriate weight
AUTrust trust evaluation model to evaluate social network for each factor that affects behavior evaluation. Therefore,
relationships, which is a hybrid model that comprehensively we also perform the evaluation of behavior trust by utilizing
considers three categories of trust factors. He assigned the a classification approach in machine learning.
same weight to each category of factor and then integrated it To evaluate the behavior, we first need to construct a behav-
into a global trust value. ior trust feature vector. There are mainly two main principles
We adopt the trust evaluation mechanism proposed by that need to be considered. Firstly, the feature vector should
Zhao and Pan [28] as our identity trust evaluation method, include factors that are closely related to the behavior during
which includes structural factors and interactive factors. the task offloading process. Secondly, it can be retrieved
He pointed out that most of the trust evaluations in existing from the interactive data without too much computing effort.
studies just quantify several trust related factors and integrate In mobile edge computing, we assume that j is an idle device
them into a trust value by setting a weight for each factor. and i is a device which has a task to offload. According to
For one thing, it is difficult to determine the weight value previous work and our research, the elements of the behavior
of each factor; for another, the evaluation result is a value, trust feature vector v(i, j) are listed in Tab. 2.
which is not intuitive to users. Consequently, he formalized To evaluate the behavior of j, we take 7 related factors into
trust evaluation in online social networks as a classification consideration. The first five factors are related to the perfor-
problem in machine learning. mance of all services provided by j, and the last two factors are
To evaluate the trust level in online social networks, it is related to the performance of services provided by j for i. If we
essential to construct an identity trust feature vector. In online take a time interval t as a segment, we update the behavior
social networks, m and n are two users. We suppose that m is trust level once per segment, and take the performance of j in
an evaluator and n is a object evaluated by m. The mechanism the latest s segments as the data source.
combines structural and interactive factors to construct a trust For the first five features, ANum(j) and CNum(j) respec-
feature vector v(m, n). If n is a fan of m, n is called a follower tively correspond to the total number of services provided by
of m and m is called a friend of n. The elements of the vector j and the number of tasks successfully completed by j, while
are shown in Tab. 1. For more details, please refer to Zhao’s ACRatio(j) stands for the ratio of CNum(j) and ANum(j).
paper, which will not be repeated here. Reputation(j) represents the overall impression of j given by
Corresponding to task sensitivity, we also divide identity all objects interacting with j, and Reputation(j) ∈ [0, 1]. ACTp
trust into three levels. The identity trust level of n given by m (actual completion time) and ECTp (estimated completion
can be expressed as It(m, n). It(m, n) ∈ {1, 2, 3}. The values time) respectively correspond to the actual completion time

VOLUME 7, 2019 150109

 D. Wu et al.: Trust-Aware Task Offloading Framework in MEC

TABLE 2. Elements of behavior trust feature vector.

and the estimated completion time of a successfully com-

pleted task Tp . Equation (2) is used to calculate TD_rate(j).
TD_rate(j) represents the average deviation rate of the actual
time spent by j and the estimated time. By considering the
time deviation rate, we can judge whether j is misbehaving in
the process of providing services, such as reducing a large FIGURE 2. Trust-aware task offloading framework.
number of resources and not providing resources. For the
last two factors, Num(i, j) describes the number of times j
are completed in the training phase. Based on these two
provides services for i. Score(i, j) describes the average score
models, the identity trust and behavior trust levels of resource
of j given by i according to j’s services. It is noteworthy
providers can be predicted.
that we assume that all objects are willing to contribute their
The trust evaluation module gives the predicted behavior
resources to each other.
and identity trust levels of each provider and passes the
CNum(j)

1 X ACTp − ECTp predicted results to the filtering module. Next, the filtering
TD_rate(j) = (2) module filters out the unqualified providers whose behavior
CNum(j) ECTp
p=1 trust level is lower than that required by the user and whose
According to actual requirements, we divide behavior trust identity trust level is lower than task sensitivity level. This
into three levels. The values from 1 to 3 indicate that the step guarantees the trustworthiness of services by removing
trust levels are poor, acceptable, and strongly recommended, these unqualified providers. Then, the filtering module hands
respectively. The first level is used to punish the resource over the qualified providers to the selection module. Last,
providers for poor performance. If a resource provider’s the selection module calculates the cost of time and energy
behavior trust level is 1, its behavior is unreliable and it overhead and selects an appropriate resource provider for
is not recommended to be selected to perform a task. If a the task. It is important to note that we trust the behav-
resource provider’s behavior trust level is 2, it means that the ior and identity of the MEC servers provided by operators
behavior of the resource provider is relatively reliable and can by default.
be recommended to perform general tasks. Level 3 indicates
that the behavior of the resource provider is highly reliable A. TRUST EVALUATION
and it can be recommended to perform very important tasks Our trust evaluation of resource providers includes two
that are not allowed to fail or are very sensitive to latency. aspects: identity trust and behavior trust, which can be
The number of behavior trust levels can be further extended if expressed as Tr = (It, Bt). Identity trust (It) comes from
needed. After constructing the feature vector, we need to map the relationships that the owners of these devices establish at
it to a discrete trust level ranging from 1 to 3. The behavior the social layer. Behavior trust (Bt) refers to the evaluation of
trust level of j given by i can be expressed as Bt(i, j). We resource providers’ behavior and performance over a period
use BTC(v(i, j)) to express the mapping from the trust feature of time.
vector to a behavior trust level, that is, Bt(i, j) = BTC(v(i, j)). We choose support vector machines (SVM) for classifi-
cation. In machine learning, support vector machines are
IV. TRUST-AWARE TASK OFFLOADING FRAMEWORK supervised learning models with associated learning algo-
To reduce latency or energy consumption and ensure the trust- rithms for classification and regression analysis. There are
worthiness of the services, we proposed a trust-aware task two reasons why we choose SVM. On the one hand, since
offloading framework (TATOF) in MEC, as shown in Fig. 2. there is no labeled dataset for this research, we have to label
Fig. 2 illustrates the overall structure of our framework. the relevant data as dataset. Compared with other classifi-
There are mainly three modules, namely, trust evaluation cation algorithms, SVM requires a relatively small sample
module, filtering module, and selection module. UIDi is size, which reduces our work on labeling data. On the other
the relevant identification information of device i. Identity hand, SVM can not only achieve good results in processing
model and behavior model in the trust evaluation module linearly separable sample data, but also be good at dealing

150110 VOLUME 7, 2019

D. Wu et al.: Trust-Aware Task Offloading Framework in MEC

with linearly inseparable sample data. We cannot guarantee

that our labeled training data must be linearly separable.
The trust evaluation module is mainly composed of data
collection, feature extraction, model training, and trust level
prediction. In the data collection part, the behavior data is
obtained through a simulation tool. We obtain L records in the
format (UIDBi , UIDBj ) generated by the simulation tool, and
label these records with different levels. After that, the feature
extraction sub-module will use UIDBi and UIDBj as parame-
ters to produce L pairs (vector(UIDBi , UIDBj ), Bt(i, j)). Then, FIGURE 3. Device identity trust graph.
these data will be sent to the model training sub-module and
used as training data. After the model is trained and tested, the
trust level prediction sub-module will predict the trust level of
any user pair on the basis of the trained model. The identity
data is captured from a provided API of a social network
application, and the processing steps are similar to behavior
trust. We will not repeat it here.
In general, the SVM-based trust evaluation process can be
summarized as follows.
• Collect a large number of trust feature vectors with the
corresponding trust levels, and put them as training data.
• Choose suitable parameters and kernel function for
FIGURE 4. D2D connection graph.
SVM and train the classifier with training data.
• Test and validate our classifier. If the accuracy meets our
requirements, then proceed to the next step, or go back
to do some adjustment.
• Use the classifier to make inference, that is, when a
feature vector v(i, j) is given, the classifier can infer
which trust level i holds on j.

B. FILTERING
The filtering module is mainly used to filter out unqualified
nodes. If the task sensitivity level is higher than a provider’s FIGURE 5. D2D connection associated with behavior trust graph.
identity trust level, the provider is unqualified. If the behavior
trust level of a provider is lower than the user’s required
behavior level, the provider is unqualified. device j, and the weight of the directed edge represents the
For the convenience of description, a computation task level of identity trust. In particular, when a device and another
model is constructed here. For a computation task, we use device belong to the same user, the weight of the directed edge
a five-tuple hIi , Li , Oi , Si , Bi i to describe it. Ii is the input between the two devices is 3.
data size of the task, which includes program code, input Next, we introduce the D2D connection graph and the
parameters, and so on. Li is the amount of computing resource D2D connection associated with behavior trust graph. Based
(i.e., number of CPU cycles) required by the task. Oi is the on the technology of D2D communication, we can acquire
output data size of the task. Si represents the sensitivity level the devices in the current D2D communication range in real
of the task, which is obtained by analyzing the sensitivity of time. These devices within the communication range can
the sensitive information set contained in the task. Bi repre- form a connectable graph, which is a D2D connection graph,
sents the behavior trust level required by the user generating as shown in Fig. 4. Each node represents a device. If there
the task. is a connection between two devices, it means that the two
To establish the relationship between the candidate devices are in the range of communication with each other.
resource providers and the devices generating tasks, we need If there is a small blue ellipse in a node, it indicates that a task
to construct three related graphs, which are described in detail is generated on the device node. The small ellipse has a task
as follows. sensitivity attribute Si which indicates that the sensitivity level
Specifically, we introduce the device identity trust graph of the task is r. On the basis of Fig. 4, we introduce behavior
Gtru to model the social tie among the devices, as shown trust between devices, and form a D2D connection associated
in Fig. 3. Each node represents a device. If there is a directed with behavior trust graph, as shown in Fig. 5. If there is a
edge from node i to node j, it indicates that the owner of directed edge from node i to node j, the weight of the edge
device i has an identity trust relationship with the owner of represents the behavior trust level of j given by i.

VOLUME 7, 2019 150111

 D. Wu et al.: Trust-Aware Task Offloading Framework in MEC

After constructing the above diagrams, we can perform Algorithm 1 Trust-Based Filtering Algorithm 1
filtering in the nodes filtering sub-module to remove those Input: a task Ti , task’s id: task_id
unqualified nodes. Output: d2dTrust[ ]
For a task Ti = hIi , Li , Oi , Si , Bi i on device i, the filtering
1: get device number device_i via task_id
module will remove the nodes that can not meet the require-
2: get the position location_i of devicei
ments of Si and Bi . Si determines the required identity trust
3: get the array d2dConnection [ ] of devices within the
level, while Bi determines the required behavior trust level.
D2D communication range of devicei via location_i
According to whether the task contains sensitive information,
4: si =getTaskSensitivity(Ti )
there are two cases.
5: bi =getBehaviorLevel(Ti )
In one case, when the task does not contain any sensitive
6: if si =0 then
information, we think that the sensitivity level of the task is
7: for each d2d_j in d2dConnection [ ] do
0. We no longer need to consider identity trust. For a task
8: Bt=getBehaviorTrust(device_i, d2d_j)
Ti on device i, the qualified node j in the array d2dTrust[ ]
9: if Bt > bi then
must satisfy the following requirement: In Fig. 5, there is an
10: d2dTrust [counter]=d2d_j
edge from i to j and the weight of the edge is greater than or
11: counter++
equal to Bi (i.e., Bi <= Bt(i, j) ). In this case, the qualified
12: end if
nodes in the array d2dTrust[ ] are within the range of its D2D
13: end for
communication and meet the required behavior trust level at
14: else
the same time.
15: get the owner deviceIOwner of device i by device_i
In the other case, when the task contains sensitive infor-
16: get the array social_person [ ] of human who have
mation, the sensitivity level of the task is Si . We need to
social relationship with deviceIOwner
combine Fig. 3 and Fig. 5 for filtering. For a task Ti on device
17: trustPerson[ ]=returnTrustPerson(si , social_person[ ])
i, the qualified node j in the array d2dTrust[ ] must satisfy the
following requirements: In Fig. 5, there is an edge from i to
18: for each d2d_j in d2dConnection [ ] do
j and the weight of the edge is greater than or equal to Bi
19: Bt=getBehaviorTrust(device_i, d2d_j)
(i.e., Bi <= Bt(i, j) ); in Fig. 3, there is an edge from i to j
20: if Bt > bi then
and the weight of the edge is greater than or equal to the Si of
21: deviceJOwner=getOwnerOfDevice(d2d_j)
the task in Fig. 5 (i.e., Si <= It(i, j) ).
22: if deviceIOwner==deviceJOwner then
We give the corresponding trust-based filtering algorithm 1
23: d2dTrust [counter]=d2d_j
as follows.
24: counter++
25: end if
C. SELECTION
26: for each trust_id in trustPerson [ ] do
Qualified nodes after filtering are considered to be trusted by 27: if deviceJOwner==trust_id then
the user both in identity and behavior. For one thing, they 28: d2dTrust [counter]=d2d_j
will not illegally obtain our sensitive information; for another, 29: counter++
they behave well during the task offloading process. 30: end if
Faced with a large number of candidate nodes, how to 31: end for
select a corresponding node to complete the task is gener- 32: end if
ally determined by an offloading policy. In an offloading 33: end for
policy, there may be only one execution mode or several 34: end if
optional execution modes. In this module, we adopt a hybrid 35: return d2dTrust [ ]
offloading policy, which mainly provides four alternative
modes of task execution, as shown in Fig. 6. This policy
mainly refers to the work of Chen et al. [29] and we expand • D2D-Assisted edge server offloaded execution: A
latency on this basis. The four modes it provides are described device with poor cellular connection first transmits its
as follows. computation task to a neighboring device through D2D
• Local execution: Users can choose to perform tasks
communication. The neighboring device has a strong
on their mobile devices without the overhead of task cellular connection, which can help it to offload the
offloading. task to an edge server for execution. After the task
• D2D offloaded execution: Devices in proximity at the
is completed, the output can be obtained through the
network edge can share their computing resources neighboring device.
through D2D communication to help other devices com- It is noteworthy that we think for users from users’ point of
plete their tasks, and these devices can benefit from the view to minimize the consumption of all users in the system.
mode. Edge servers (also known as MEC servers) generally have
• Direct edge server offloaded execution: A device can constant power supply. Therefore, we do not add the energy
offload its tasks directly to an edge server. consumption of MEC servers to the cost of execution modes.

150112 VOLUME 7, 2019

D. Wu et al.: Trust-Aware Task Offloading Framework in MEC

by Eijd2 = ρjc Li . Among them, Pdit and Pdir represent

the energy consumption of device i in transmitting and
receiving data via D2D per unit time, respectively, and
Dij represents the rate when data is transmitted from i to
j through D2D. cj and ρjc are similar to those mentioned
in local execution. Accordingly, we can obtain the cost
of the D2D offloaded execution mode  in terms of both

time and energy overhead as θijd = λti Tijd1 + Tijd2 +
 
λei Eijd1 + Eijd2 .
• Direct edge server offloaded execution: In this mode,
the cost of energy for device i is mainly derived from
the transmission of the task, and the cost of time is
mainly derived from transmission and execution of the
task. The time consumption is given by Tic1 + Tic2 and
the energy consumption is given by Eic . In this case,
the time consumption of transmitting input and output
FIGURE 6. An illustration of four different task execution modes. data between the device and an edge server is given by
Tic1 = Ii /Dti + Oi /Dri and energy consumption of trans-
According to the task five-tuple hIi , Li , Oi , Si , Bi i, the latency mitting input and output data between the device and
and energy consumption in the above four different execution an edge server is given by Eic = Pcit Ii /Dti + Pcir Oi /Dri .
modes can be given as follows, which are realized by the cost The time for executing the offloaded task in the edge
prediction sub-module. server is given by Tic2 = Li /Km . Among them, Dti and
• Local execution: In this mode, the time spent mainly
Dri represent the speed of uploading and downloading
comes from the task’s local computation time. The exe- data between i and the edge server, respectively, and Km
cution time for computation is given by Til = Li /ci and represents the CPU operating frequency of the virtual
the energy consumption is given by Eil = ρic Li . Among machine m assigned to the task by the server. Accord-
them, ci = (1 − δi ) Zi represents the CPU computing ingly, we can obtain the cost of the Direct edge server
power provided by device i per unit time and Zi repre- offloaded execution mode in terms
of time and energy
sents the CPU working frequency. Besides, δi represents overhead as θim c = λt T c1 + T c2 + λe E c .
i i i i i
the current load ratio (device i may need to run some • D2D-Assisted edge server offloaded execution: In this
background loads) and ρic is the energy cost per CPU mode, the cost of energy for devices is mainly derived
cycle for computation. Accordingly, we can obtain the from the transmission of a task, and the cost of time is
cost of the local execution mode in terms of both time mainly derived from the transmission and execution of
and energy overhead as θil = λti Til +λei Eil . λei , λti ∈ [0, 1] the task. The time consumption is given by Tijdc1 + Tjdc2
denote the weights of time and energy for device i’s and the energy consumption is given by Eijdc1 + Ejdc2 .
selection and λei + λti = 1. The user can adjust the two In this case, the time consumption of transmitting input
weight parameters according to different situations. For and output data is given by Tijdc1 = Ii /Dij + Oi /Dji +
example, when a user is running some applications that Ii /Dtj + Oi /Drj . The time consumption of task execution
are sensitive to latency, the user can set λei = 0 and is given by Tjdc2 = Li /Km . Km is similar to hat men-
λti = 1 in the selection process. In contrast, when a tioned in the above execution mode. The energy con-
device is at a low battery state and the user cares about sumption of transmitting input and output data  between
energy consumption, the user can set λei = 1 and λti = 0. two devices is given by Eij = Pit + Pjr Ii /Dij +
dc1 d d
• D2D offloaded execution: In this mode, the cost of time  
and energy is mainly derived from the transmission of Pdir + Pdjt Oi /Dji . Besides, the energy consumption of
a task and the execution of the task. The time spent transmitting input and output data between device j and
is given by Tijd1 + Tijd2 and the energy consumption is an edge server is given by Ejdc2 = Pcjt Ii /Dtj + Pcjr Oi /Drj .
given by Eijd1 + Eijd2 . In this case, the time spent in Accordingly, we can obtain the cost of the D2D-Assisted
transmitting input and output data between two devices edge server offloaded execution modein terms of both 
is given by Tijd1 = Ii /Dij + Oi /Dji and energy con- time and energy overhead as θijdc = λti Tijdc1 + Tjdc2 +
 
sumption of transmitting input and output data  between λei Eijdc1 + Ejdc2 .
two devices is given by Eij = Pit + Pjr Ii /Dij +
d1 d d
  Based on the above analysis, we can predict the latency and
Pdjt + Pdir Oi /Dji . The time for executing the offloaded energy consumption under different task execution modes,
task by device j is given by Tijd2 = Li /cj and the energy and select an appropriate mode for a user. However, when
for executing the offloaded task by device j is given multiple users in the system have tasks waiting for offloading,

VOLUME 7, 2019 150113

 D. Wu et al.: Trust-Aware Task Offloading Framework in MEC

V. PERFORMANCE EVALUATION
To verify the effectiveness of our proposed framework,
we have carried out some experiments. We mainly used
LIBSVM tool [30] for model training in trust evaluation
module and EdgeCloudSim [31], [32] for simulating mobile
edge computing environment. LIBSVM tool is an integrated
toolkit developed for supported vector machine and has
been widely used. EdgeCloudSim is a simulation tool that
supports the modeling of both computational and networking
resources to handle the edge computing scenarios and it
has good extensibility. It includes task generation module,
FIGURE 7. An illustration of bipartite graph of task execution modes. mobility module, network environment simulation module,
offloading policy module and so on. User devices include
how to choose an appropriate execution mode for each user? smart bracelet, wearable glasses, wearable medical devices,
The node selection sub-module can also help us to solve this mobile phones, tablets, computers, etc., covering a wide
problem. range. D2D connections between devices will change with
To face the multi-user scenario in MEC, the multi-user movement of devices, which depends on mobility mod-
offloading problem is reduced to a matching problem of ule in the EdgeCloudSim. Meanwhile, the cellular data
bipartite graph. To minimize the overhead of the whole sys- rate decreases with the increase of the number of devices
tem, we use the minimum-weight bipartite perfect matching in the hotspot, because we simulate the network environ-
solution over the bipartite graph to select a suitable node for ment by this tool. We modified and extended this tool
each user device to complete a task. to test the effectiveness of our framework. In particular,
The nodes left by the common action of the three graphs we searched for answers to the following three research
of the filtering module are qualified. If users request the questions:
behavior trust level of resource providers to be acceptable, • RQ1: How does the classification of machine learning
a bipartite graph consisting of the candidate nodes (edge perform in trust evaluation (i.e., identity trust and behav-
servers or qualified devices) and the tasks nodes waiting to ior trust)?
select execution modes is shown in Fig. 7. The graph is • RQ2: After introducing behavior trust evaluation, can
divided into the side of task generation (referred to as the the misbehaving nodes be filtered out for users to
initial side) and the side of optional execution modes (referred improve the performance of task offloading?
to as the execution side). A node i of the initial side indicates • RQ3: What is the performance of our framework in
that device i has a task, and a node j of the execution side reducing latency or energy consumption for users?
represents an optional object for task offloading. The weight
of one edge represents the time and energy consumption A. EXPERIMENTS ON RQ1
when a task chooses this execution mode. The construction 1) EXPERIMENTAL DESIGN OF PREDICTION PERFORMANCE
method of the bipartite graph is given as follows. By using the Firstly, we need to collect relevant data, and then label them
minimum-weight bipartite perfect matching method, we can with different levels. Secondly, after the dataset is ready,
choose an appropriate node for each task. For our prob- to give full play to its performance, SVM needs to select an
lem, the number of edges is proportional to the number of appropriate kernel, the kernel’s parameters, and soft margin
nodes N, and hence our bipartite matching method has a parameter C. Since there may be linearly inseparable in trust
low complexity of O(N 2 ), which can scale well for practical evaluation, we choose the RBF kernel. The best combination
implementation. of C and γ is often selected by a grid search with expo-
• Local execution: There is an edge between a task node i nentially growing sequences of C and the kernel’s parameter
and its corresponding local node i, and the weight of this γ . Finally, there are several measures commonly used in
edge is θil . machine learning. For each category, a confusion matrix is
• D2D offloaded execution: There is an edge between a shown in Tab. 3, where m1 represents the number of items
task node i and a node j if node j is in i’s d2dTrust[ ] after correctly predicted for the category, m2 represents the number
filtering for device i, and the weight of this edge is θijd . of items that should fall into this category but are predicted
• Direct edge server offloaded execution: There is an edge to fall outside that category, m3 represents the number of
between a task node i and its subscribed server node, and items that do not belong to this category but are predicted
the weight of this edge is θimc .
to fall into this category, and m4 represents the number of
• D2D-Assisted edge server offloaded execution: There is items that do not belong to this category and are correctly
an edge between a task node i and a bundled node (j and predicted. In our experiments, three measures are taken into
the edge server serving for j) if node j is in i’s d2dTrust[ ] consideration: precision, recall and F-measure. The precision
after filtering for device i, and the weight of this P is P = m1 / (m1 + m3 ). The recall R is R = m1 / (m1 + m2 ),
edge is θijdc . and the F-measure is defined as F = 2PR/ (P + R).

150114 VOLUME 7, 2019

D. Wu et al.: Trust-Aware Task Offloading Framework in MEC

TABLE 3. Confusion matrix. TABLE 4. Measurement of identity trust classifier.

Weibo, one of the most popular social networks in China,

TABLE 5. Measurement of behavior trust classifier.
was selected as the data source for user identity trust evalu-
ation. With the API of Weibo which is open for developers
to build applications, a lot of information can be obtained,
including user configuration, relationships, and so on.
In the data collection part of identity trust evaluation,
the data we employ is originated from Weibo dataset of KDD-
Cup2012 track one [33]. The dataset represents a sampled
snapshot of Weibo, and provides standardized format and
abundant information in multiple domains. There are 7 txt
files in all. Our experiment makes use of the information
in three of them which include: user_profile.txt, user_sns.txt
and user_action.txt. Because we only consider the identity
trust levels that one user has on its followers, Dis(m, n) in the
identity trust feature vector is set to be 1 here. To ensure the
quality of training data, we randomly selected 1920 records
and invited three experts in the field of trust evaluation to label
these records at the level of 1 to 3. After labeling the records,
we used 1728 records as training data and the remaining
192 records as test data. In the experiment of identity trust,
we obtained the best accuracy under the combination of FIGURE 8. Changes in accuracy of identity trust.
C = 512 and γ = 8. We compared our identity trust
evaluation method with the AUTrust method of Yin et al. Tab. 4 shows the results of three measures mentioned above
(refer to [27]). AUTrust is a trust evaluation method of social for each level in identity trust evaluation, and Tab. 5 shows the
networks by assigning weights to trust factors. The trust results of three measures mentioned above for each level in
factors in this method are basically the same as those in our behavior trust evaluation. They verify the usability and good
identity trust evaluation method. The author divides these performance of the trust classifiers we trained.
trust factors into three categories and assigns equal weight To verify the accuracy of the classifiers we trained,
to each category. we added some data and have performed experiments on
In the data collection part of behavior trust evaluation, different test data sizes.
we modified and extended EdgeCloudSim to collect data The comparison result of identity trust evaluation is shown
that needs to be used. To ensure the quality of training data, in Fig. 8. ITC is our identity trust classifier. To compare
we invited three experts who have studied mobile edge com- our identity classifier and AUTrust, the experts analyzed the
puting to label these data. We divided behavior trust into global trust value and divided the range of [0, 1] into three
three levels, and labeled 1920 records, of which 1728 records intervals [0, 0.3), [0.3, 0.6), and [0.6, 1] according to their
were used as training data, and the remaining 192 records experience. These three intervals correspond to level 1, 2 and
were used as test data. In the experiment of behavior trust, 3 of identity trust respectively. As shown in the figure, our
we obtained the best accuracy with the combination of classifier can achieve better accuracy. Because in the AUTrust
C = 32 and γ = 8. We compared our behavior trust eval- method, each category of trust factor is given the same weight,
uation method with the MSTrust method of Yuan et al. (refer which is inaccurate and unrealistic.
to [17]). MSTrust is a trust evaluation method for service The comparison result of behavior trust evaluation is
behavior in edge computing and the author gives a formula shown in Fig. 9. BTC is our identity trust classifier. To com-
for aggregating global trust. This method mainly considers pare our behavior classifier and MSTrust, the experts ana-
three trust factors: success rate, direct feedback and overall lyzed the global trust value and divided the range of [0, 1] into
reputation. three intervals [0, 0.35), [0.35, 0.6), and [0.6, 1] according to
their experience. These three intervals correspond to level 1,
2) THE RESULT 2 and 3 of behavior trust respectively. As shown in the figure,
RQ1: How does the classification of machine learning per- our classifier can achieve better accuracy. Because in the
form in trust evaluation (i.e., identity trust and behavior MSTrust method, only three trust factors, namely success
trust)? rate, direct feedback and overall reputation, are considered.

VOLUME 7, 2019 150115

 D. Wu et al.: Trust-Aware Task Offloading Framework in MEC

FIGURE 9. Changes in accuracy of behavior trust.

FIGURE 10. Average task failure rate.

Our behavior trust evaluation method includes not only the
above trust factors, but also the total number of providing
services (i.e., ANum(j) ) and the time deviation rate (i.e., that of ATATOF_MSTrust because our behavior trust eval-
TD_rate(j) ). These two trust factors can help us better uation mechanism takes more behavior factors into account
evaluate the behavior of resource providers. For example, and can better identify some improper behaviors. Task fail-
one provider provided and successfully completed 10 times ure occurs in two cases. In one case, when misbehaving
service while the other provided and successfully completed devices incorrectly complete some tasks, these tasks will
90 times service. Although both providers have a 100% suc- fail. In the other case, because this paper does not deal
cess rate, we have a higher trust level for the provider that with the mobility management of task migration, we also
completed service 90 times. The time deviation rate can help assume that a task will fail when two devices are out of
us identify whether the provider reduces a large amount of communication range with each other. Task migration means
resources in the service process without agreement. that the virtual machine shoule be migrated if a user roam
from one network region to another. In fact, we will do
B. EXPERIMENTS ON RQ2 further research on task migration in the future. When a
1) EXPERIMENTAL DESIGN OF THE IMPACT AFTER device is beyond the original communication range due
INTRODUCING TRUST EVALUATION to mobility, we will choose an appropriate approach to
We invited three experts in edge computing to inject several transmit output data to the user to ensure the continuity
misbehaving devices that reduce a lot of resources or inten- of service.
tionally fail to perform tasks correctly during service into the
simulation tool. When tasks are assigned to these devices, C. EXPERIMENTS ON RQ3
they may fail because of their improper behavior. By intro- 1) EXPERIMENTAL DESIGN OF REDUCING LATENCY
ducing our behavior trust evaluation, we can filter out these AND ENERGY CONSUMPTION
misbehaving devices and reduce the failure rate of tasks in the The main purpose of this experiment is to verify the perfor-
system. We compared the task failure rate in three cases: our mance of the proposed task offloading framework in reducing
TATOF using BTC trust classifier (i.e., TATOF_BTC), only latency and energy consumption as the number of devices
selection module (i.e., Selection_only), and another ATATOF increases. After adding task sensitivity computation, trust
that replaces our TATOF’s behavior trust evaluation mecha- evaluation and filtering, these processes will introduce some
nism with MSTrust (i.e., ATATOF_MSTrust). time overhead. In this case, can our framework reduce latency
or energy consumption for users during task offloading? To
2) THE RESULT better observe the effect of our framework, we set up three
RQ2: After introducing behavior trust evaluation, can the offloading scenarios ( i.e., local computation, only offloaded
misbehaving nodes be filtered out for users to improve the to the edge server (only_edge), and our proposed TATOF).
performance of task offloading? The average reduction rates of time and energy of another
The average task failure rate in the system is shown two scenarios are compared by using the local computation
in Fig. 10. The Selection_only indicates that it has only as a benchmark.
selection module but no trust evaluation and filtering mod- The levels of identity and behavior trust between devices
ules. In this case, task failure rate is high because of some is invariable in this experiment. Tab. 6 shows the simulation
misbehaved nodes. When behavior trust evaluation is intro- settings, most of which are consistent with the actual mea-
duced, the nodes with poor performance are filtered out, sured values in practice [34], [35]. The number of computing
so the task failure rate of the system is reduced. At the same resources required by a task is expressed in terms of the MIPS
time, our TATOF_BTC has a lower task failure rate than and 1Mips = 4MHz.

150116 VOLUME 7, 2019

D. Wu et al.: Trust-Aware Task Offloading Framework in MEC

TABLE 6. Simulation setting.

FIGURE 11. Time reduction rate.

For better observation, when we focus on the performance
of reducing latency, we assume that users’ weight of energy
consumption in task offloading is 0. Conversely, when we
focus on the performance of reducing energy consumption,
we assume that users’ weight of latency is 0.

2) THE RESULT
RQ3: What is the performance of our framework in reducing
latency or energy consumption for users?
In Web services, users need to provide some personal
information required by services as input when enjoying
services. Similar to Web services, a task generated by an
application needs to collect some personal information as
FIGURE 12. Energy reduction rate.
input to complete a specific function. The information items
that need to be collected are predetermined, predefined, and
limited in number. For the computation of task sensitivity, The average energy reduction rate of the two scenarios
if there are n personal information items needed in a task, is shown in Fig. 12. As the number of devices increases,
the number of relevant sensitivity composed of each two the number of alternative devices for D2D connections
personal information items is Cn2 . Therefore, combined with increases, so the energy reduction rate increases a little under
the analysis of the task sensitivity formula, the computation TATOF. As the number of devices continues to grow, the cel-
of task sensitivity has a low complexity of O(n2 ). For the lular data rate between devices in a hotspot and edge servers
trust evaluation, once the model is completed in the training decreases. As a result, transmission time increases and energy
phase, the trust classifier is obtained. In practice, the time reduction rate in the system decreases. Under TATOF, with
cost of predicting trust level based on the classifier is low. the help of D2D and D2D-Assist execution modes, the energy
For the filtering process, if the number of devices in the reduction rate is superior to that of only_edge scenario.
communication range of device i is N , the process has a In conclusion, although our TATOF takes into account trust
low complexity of O(N 2 ) because the identity trust level and evaluation mechanism and filters out unqualified resource
behavior trust level of these N devices need to be considered providers to ensure service trustworthiness, it can still reduce
at the same time. Therefore, our framework does not cause latency or energy consumption for users.
too much time overhead.
As the number of devices in the system increases, the aver- VI. CONCLUSION AND FUTURE WORK
age time reduction rate of the two scenarios is shown We introduce online social networks into mobile edge com-
in Fig. 11. When the number of devices in the system puting and propose a trust-aware task offloading framework.
increases gradually, the cellular data rate between a device To ensure the trustworthiness of services, we introduce iden-
in a hotspot and an edge server decreases, and the trans- tity trust and behavior trust into trust evaluation in mobile
mission time increases. At the same time, as the number of edge computing. To address the challenge that it is difficult to
devices increases, edge servers have to handle more tasks, assign an appropriate weight to each trust factor, we formalize
so the servers’ response time to tasks increases. Due to the trust evaluation of mobile edge computing into a classifica-
above two reasons, the time reduction rate of only_edge is tion problem. Then, we filter out unqualified nodes that do not
gradually reduced, and our TATOF, with the help of D2D meet the requirements of tasks. Finally, the selection module
and D2D-Assist execution modes, can still maintain a better selects one of the qualified providers for a user to execute the
performance than only_edge. task based on the offloading policy. Based on our framework,

VOLUME 7, 2019 150117

 D. Wu et al.: Trust-Aware Task Offloading Framework in MEC

we are able to reduce latency or energy consumption and [19] K. Zhang, Y. Mao, S. Leng, S. Maharjan, and Y. Zhang, ‘‘Optimal delay
ensure the trustworthiness of the services for users. constrained offloading for vehicular edge computing networks,’’ in Proc.
ICC, Paris, France, May 2017, pp. 1–6.
In future work, we plan to add more related factors into a [20] F. Liu, Z. Huang, and L. Wang, ‘‘Energy-efficient collaborative task com-
trust feature vector and conduct experiments on larger dataset. putation offloading in cloud-assisted edge computing for IoT sensors,’’
Meanwhile, we will conduct further research on mobility Sensors, vol. 19, no. 5, p. 1105, Mar. 2019.
[21] C. Dong and W. Wen, ‘‘Joint optimization for task offloading in edge
management of task migration. When a device is out of the computing: An evolutionary game approach,’’ Sensors, vol. 19, no. 3,
original communication range due to movement, we want to p. 740, Feb. 2019.
choose an appropriate way to return the output data of the task [22] K. Li, M. Tao, and Z. Chen, ‘‘Exploiting computation replication for
mobile edge computing: A fundamental computation-communication
to the device. tradeoff study,’’ 2019, arXiv:1903.10837. [Online]. Available: https://
arxiv.org/abs/1903.10837
REFERENCES [23] A. Al-Shuwaili and O. Simeone, ‘‘Energy-efficient resource alloca-
tion for mobile edge computing-based augmented reality applica-
[1] H. T. Dinh, C. Lee, D. Niyato, and P. Wang, ‘‘A survey of mobile cloud tions,’’ IEEE Wireless Commun. Lett., vol. 6, no. 3, pp. 398–401,
computing: Architecture, applications, and approaches,’’ Wireless Com- Jun. 2017.
mun. Mobile Comput., vol. 13, no. 18, pp. 1587–1611, Dec. 2013. [24] Standards for Privacy of Individually Identifiable Health Information,
[2] E. Cuervo, A. Balasubramanian, D.-K. Cho, A. Wolman, S. Saroiu, Standard 45 CFR Parts 160 and 164, HIPAA, Dec. 2000.
R. Chandra, and P. Bahl, ‘‘MAUI: Making smartphones last longer [25] J. R. Weible, ‘‘Privacy and data: An empirical study of the influence
with code offload,’’ in Proc. MobiSys, San Francisco, CA, USA, 2010, of types of data and situational context upon privacy perceptions,’’
pp. 49–62. Ph.D. dissertation, Dept. Elect. Eng., Mississippi State Univ., Starkville,
[3] Z. M. Fadlullah, F. Tang, B. Mao, N. Kato, O. Akashi, T. Inoue, MS, USA, 1993.
and K. Mizutani, ‘‘State-of-the-art deep learning: Evolving machine [26] W. Sherchan, S. Nepal, and C. Paris, ‘‘A survey of trust in social
intelligence toward tomorrow’s intelligent network traffic control sys- networks,’’ ACM Comput. Surv., vol. 45, no. 4, pp. 47:1–47:33,
tems,’’ IEEE Commun. Surveys Tuts., vol. 19, no. 4, pp. 2432–2455, Aug. 2013.
4th Quart., 2017. [27] G. Yin, F. Jiang, S. Cheng, X. Li, and X. He, ‘‘AUTrust: A practical trust
[4] Y. Mao, C. You, J. Zhang, K. Huang, and K. B. Letaief, ‘‘A survey on measurement for adjacent users in social networks,’’ in Proc. CGC, Hunan,
mobile edge computing: The communication perspective,’’ IEEE Commun. China, Nov. 2012, pp. 360–367.
Surveys Tuts., vol. 19, no. 4, pp. 2322–2358, 4th Quart., 2017. [28] K. Zhao and L. Pan, ‘‘A machine learning based trust evaluation framework
[5] F. Bonomi, R. A. Milito, J. Zhu, and S. Addepalli, ‘‘Fog computing and for online social networks,’’ in Proc. IEEE TrustCom, Beijing, China,
its role in the Internet of Things,’’ in Proc. MCC SIGCOMM, Helsinki, Sep. 2014, pp. 69–74.
Finland, 2012, pp. 13–16. [29] X. Chen, Z. Zhou, W. Wu, D. Wu, and J. Zhang, ‘‘Socially-motivated coop-
[6] J. Zhang, B. Chen, and Y. Zhao, ‘‘Data security and privacy-preserving in erative mobile edge computing,’’ IEEE Netw., vol. 32, no. 6, pp. 177–183,
edge computing paradigm: Survey and open issues,’’ IEEE Access, vol. 6, Nov./Dec. 2018.
pp. 18209–18237, 2018. [30] C.-C. Chang and C.-J. Lin, ‘‘LIBSVM: A library for support vector
[7] M. Haus, M. Waqas, A. Y. Ding, Y. Li, S. Tarkoma, and J. Ott, ‘‘Security machines,’’ ACM Trans. Intell. Syst. Technol., vol. 2, no. 3, pp. 27:1–27:27,
and privacy in device-to-device (D2D) communication: A review,’’ IEEE 2011.
Commun. Surveys Tuts., vol. 19, no. 2, pp. 1054–1079, 2nd Quart., 2017. [31] C. Sonmez, A. Ozgovde, and C. Ersoy, ‘‘EdgeCloudSim: An envi-
[8] R. Roman, J. López, and M. Mambo, ‘‘Mobile edge computing, fog et al.: ronment for performance evaluation of edge computing systems,’’
A survey and analysis of security threats and challenges,’’ Future Gener. Trans. Emerg. Telecommun. Technol., vol. 29, no. 11, Aug. 2018,
Comp. Syst., vol. 78, pp. 680–698, Jan. 2018. Art. no. e3493.
[9] G. Shang-Fu and Z. Jian-Lei, ‘‘A survey of reputation and trust mechanism [32] Simulation Tools: EdgeCloudSim. Accessed: Oct. 16, 2019. [Online].
in peer-to-peer network,’’ in Proc. Int. Conf. Ind. Control Electron. Eng., Available: http://github.com/CagataySonmez/EdgeCloudSim
Aug. 2012, pp. 116–119. [33] Kddcup. Accessed: Oct. 16, 2019. [Online]. Available: http://www.
[10] Z. Yan, P. Zhang, and A. V. Vasilakos, ‘‘A survey on trust management kddcup2012.org/c/kddcup2012-track1
for Internet of Things,’’ J. Netw. Comput. Appl., vol. 42, pp. 120–134, [34] X. Chen and J. Zhang, ‘‘When D2D meets cloud: Hybrid mobile
Jun. 2014. task offloadings in fog computing,’’ in Proc. IEEE ICC, Paris, France,
[11] W. Li, J. Cao, K. Hu, J. Xu, and R. Buyya, ‘‘A trust-based agent learning May 2017, pp. 1–6.
model for service composition in mobile cloud computing environments,’’ [35] X. Chen, L. Pu, L. Gao, W. Wu, and D. Wu, ‘‘Exploiting massive D2D
IEEE Access, vol. 7, pp. 34207–34226, 2019. collaboration for energy-efficient mobile edge computing,’’ IEEE Wireless
[12] X. Lin, R. Lu, X. Liang, and X. Shen, ‘‘STAP: A social-tier-assisted packet Commun., vol. 24, no. 4, pp. 64–71, Aug. 2017.
forwarding protocol for achieving receiver-location privacy preservation
in VANETs,’’ in Proc. IEEE Comput. Commun. Soc., Shanghai, China,
Apr. 2011, pp. 2147–2155.
[13] M. Hussain and B. M. Almourad, ‘‘Trust in mobile cloud computing with
LTE-based deployment,’’ in Proc. IEEE 14th Int. Conf. Commun. Assoc.
Workshops, Bali, Indonesia, Dec. 2014, pp. 643–648.
[14] W. Ahmad, S. Wang, A. Ullah, Sheharyar, and Z. Mahmood, ‘‘Reputation-
aware trust and privacy-preservation for mobile cloud computing,’’ IEEE
Access, vol. 6, pp. 46363–46381, 2018.
[15] P. Zhou, K. Wang, J. Xu, and D. Wu, ‘‘Differentially-private and trustwor-
thy online social multimedia big data retrieval in edge computing,’’ IEEE
Trans. Multimedia, vol. 21, no. 3, pp. 539–554, Mar. 2019.
[16] J. Xu, S. Wang, B. K. Bhargava, and F. Yang, ‘‘A blockchain-enabled DEXIANG WU received the B.S. degree in com-
trustless crowd-intelligence ecosystem on mobile edge computing,’’ puter science from Jiangsu Normal University.
IEEE Trans. Ind. Informat., vol. 15, no. 6, pp. 3538–3547, Jun. 2019. She is currently pursuing the M.S. degree with
doi: 10.1109/TII.2019.2896965. the Computer Science Department, Nanjing Uni-
[17] J. Yuan and X. Li, ‘‘A multi-source feedback based trust calculation versity of Aeronautics and Astronautics, Nanjing.
mechanism for edge computing,’’ in Proc. IEEE INFOCOM, Honolulu, Her current research interests include mobile edge
HI, USA, Apr. 2018, pp. 819–824. computing, cloud computing, access control, pri-
[18] M.-H. Chen, B. Liang, and M. Dong, ‘‘Joint offloading decision and vacy preservation, and service composition.
resource allocation for multi-user multi-task mobile cloud,’’ in Proc. IEEE
ICC, Kuala Lumpur, Malaysia, May 2016, pp. 1–6.

150118 VOLUME 7, 2019

D. Wu et al.: Trust-Aware Task Offloading Framework in MEC

GUOHUA SHEN received the M.S. and Ph.D. YAN CAO is currently pursuing the Ph.D. degree
degrees in computer science from the Nanjing Uni- with the Computer Science Department, Nanjing
versity of Aeronautics and Astronautics, China. University of Aeronautics and Astronautics, Nan-
He is currently an Associate Professor with the jing. Her research interests include formal method,
College of Computer Science and Engineering, access control, cyber physical systems, informa-
Nanjing University of Aeronautics and Astronau- tion security and privacy preservation, and mobile
tics. His research interests include requirement edge computing.
traceability, fog computing, description logic,
semantic web, and web services and ontology.

ZHIQIU HUANG received the Ph.D. degree in

computer science from the Nanjing University TIANBAO DU is currently pursuing the M.S.
of Aeronautics and Astronautics. He is currently degree with the Computer Science Department,
a Professor with the College of Computer Sci- Nanjing University of Aeronautics and Astro-
ence and Engineering, Nanjing University of nautics, Nanjing. His research interests include
Aeronautics and Astronautics. He is also the requirement traceability, machine learning, soft-
Director of software safety in computer science ware engineering, and edge computing.
with the Nanjing University of Aeronautics and
Astronautics. His research interests include formal
method, cloud computing, edge computing, web
security, and privacy preservation.

VOLUME 7, 2019 150119