3252 IEEE INTERNET OF THINGS JOURNAL, VOL. 10, NO.

4, 15 FEBRUARY 2023

TAFS: A Truthful Auction for IoT Application

Offloading in Fog Computing Networks
Lijun Sun , Member, IEEE, Guoliang Xue , Fellow, IEEE, and Ruozhou Yu , Senior Member, IEEE

Abstract—Emerging as an alternative to cloud comput- second [3]. Many such devices which generate these massive
ing, fog computing is expected to provide low-latency, high- IoT data are resource-constrained and, thus, do not read-
throughput, reliable services for ever-growing Internet of Things ily process and respond locally in real time. Most of the
(IoT) applications, especially real-time applications with strict
responsiveness requirements. By offloading time-critical and current applications (e.g., online games, virtual reality, intel-
computation-intensive applications to proximal fog nodes (FNs), ligent transportation systems, and smart cities) have strict
both application response time and network congestion can be response time requirements. Today’s users tend to be delay-
markedly reduced. However, the FNs commonly suffer from lim- and jitter sensitive, such that the violation of real-time require-
ited resources compared to cloud computing nodes and, hence, ment substantially degrades their Quality of Experience (QoE)
may not serve all application users with guaranteed performance.
The dynamic and heterogeneous nature of FNs also brings dif- and drive down service providers’ revenue. Furthermore, the
ficulty and overhead to fog computing resource management. majority of applications are currently hosted on cloud plat-
These issues are addressed in the present study with the design forms, which have high latency and unpredictable Internet
of a double auction mechanism, namely, truthful auction for connections. The massive amount of data generated by IoT
the fog system (TAFS), which provides incentives for FNs to greatly exacerbates congestion on the Internet. This clear gap
satisfy as many application demands as possible with guaran-
teed performance. TAFS takes into account the latency tolerance between user demands and current computing and networking
of application users during the FN assignment and resource infrastructures has created an urgent demand for resolution.
allocation to satisfy real-time requirements. We theoretically Fortunately, the novel fog computing paradigm is a promis-
prove that TAFS satisfies several desired economic properties, ing approach to solving the above problems [4]. As a comple-
including truthfulness, individual rationality, and budget bal- ment to the cloud, fog computing moves the early processing
ance. The performance of TAFS is evaluated through simulation
experiments. and storage of data from distant clouds to fog nodes (FNs)
closer to the data source. From the users’ view, applications
Index Terms—Application offloading, double auction, edge that require intensive computation and/or real-time responses
computing, fog computing, incentive mechanism.
can be hosted on proximal FNs instead of on the user devices,
a process called computation offloading, which can greatly
reduce delay and network congestion [5].
I. I NTRODUCTION The capability of fog computing can be extended by
HE Internet of Things (IoT) has developed at an unprece-
T dented speed over recent years. There were approxi-
mately 23 billion connected IoT devices in 2018; this number
incorporating decentralized fog devices. Devices with idle
resources, connected through different networks and owned
by different owners, can be integrated into a uniform fog
is expected to reach 75 billion by 2025 [1]. These ubiqui- computing platform to provide the most needed computing
tous, interconnected things bring us more efficient, convenient, resources proximal to the end users. This process, however,
and even safer lives, but they also bring significant chal- introduces two significant challenges: 1) what are the incen-
lenges. It is estimated that in 2020, 1.7 MB of data will tives for fog devices to provide their idle resources? Sharing
be created every second for every person on Earth [2]. A the idle resources is costly for device owners in terms of
single self-driving car generates about 1 GB of data per overhead, energy, and usage, and should thus be monetarily
compensated for. An incentive mechanism should generally be
Manuscript received 14 July 2021; revised 12 October 2021; accepted truthful, fair, and efficient to provide proper incentivization for
5 January 2022. Date of publication 14 January 2022; date of current ver-
sion 6 February 2023. The work of Lijun Sun was supported in part by the sharing across ad hoc FNs, while allowing users to utilize fog
Fund Project for Visiting Scholars of Shandong Provincial Department of services with confidence and 2) how do we address the com-
Education. (Corresponding author: Lijun Sun.) plexity of resource management and optimization introduced
Lijun Sun was with the School of Computing and Augmented Intelligence,
Arizona State University, Tempe, AZ 85287 USA. She is now with the by the complex interplay between heterogeneous FNs and user
College of Computer Science and Technology, Qingdao University of demands? A well-designed resource allocation method should
Science and Technology, Qingdao 266061, Shandong, China (e-mail: lijun- maximize social welfare in terms of user demand satisfaction
[email protected]).
Guoliang Xue is with the School of Computing and Augmented and resource utilization in addition to the monetary incentive.
Intelligence, Arizona State University, Tempe, AZ 85287 USA (e-mail: This article proposes a three-layer cloud-fog computing
[email protected]). system framework designed to address the above problems
Ruozhou Yu is with the Department of Computer Science, North Carolina
State University, Raleigh, NC 27606 USA (e-mail: [email protected]). (Fig. 1). At the fog computing layer, IoT providers build and
Digital Object Identifier 10.1109/JIOT.2022.3143101 manage IoT platforms to provide functional services, such
2327-4662 
c 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 02,2024 at 11:41:02 UTC from IEEE Xplore. Restrictions apply.
 SUN et al.: TAFS: TRUTHFUL AUCTION FOR IoT APPLICATION OFFLOADING IN FOG COMPUTING NETWORKS 3253

Research on offloading in fog computing has garnered con-

siderable research attention. Aazam et al. [5] reviewed the
recent offloading schemes in the fields of fog computing and
IoT, and discussed the middleware technologies that enable
offloading in a Cloud–IoT case. They also highlighted chal-
lenges with the incentive for the entity performing offloading
tasks.
Shah-Mansouri and Wong [8] considered the competition
between IoT users, where each user seeks to maximize its own
QoE while minimizing storage energy and delay via offload-
ing. A QoE maximization problem was formulated for each
user to determine the computation offloading decision.
To ensure fairness for all users, Du et al. [9] minimized
the maximum weighted cost of delay and energy consumption
among them, while ensuring that maximum delay constraints
were met. Chang et al. [10] and Liu et al. [11] studied
optimization problems for finding the optimal offloading prob-
Fig. 1. Three-layer cloud-fog computing system framework. ability and transmit power. The objective of [10] was to
minimize energy consumption, subject to the delay and power
constraints while [11] centered on multiobjective optimization,
as virtualization, QoS management, resource optimization, which involves minimizing the energy consumption, delay,
and security [6]. Three algorithms match FNs and offloading and payment cost. Zhou et al. [12] designed a server recruit-
applications to minimize the delay. ment and task offloading strategy under information asym-
Our main contributions can be summarized as follows. metry and uncertainty for autonomous driving scenarios,
1) We propose a double-auction-based incentive mech- guarantying servers with more contributions receive higher
anism, truthful auction for fog systems (TAFS), to rewards.
incentivize FNs to maximize the resource utilization Yousefpour et al. [13] introduced a general framework to
when serving application demands. analyze the service delay in an IoT–fog–cloud application
2) The proposed incentive mechanism satisfies truthfulness, scenario. They proposed a delay-minimizing offloading pol-
individual rationality, and budget balance. icy for FNs to minimize service delay across IoT nodes.
3) At its core, our incentive mechanism relies on solving a An FN cooperation strategy called “offload forwarding” was
matching and resource allocation problem between FNs proposed in [14], in which all FNs jointly determine the
and user applications (UAs), which is NP-hard. We pro- optimal size of the offloaded workload to improve users’ QoE.
pose fast heuristic algorithms for solving this problem Chang et al. [15] proposed a dynamic optimization scheme
under real-time conditions. for an edge computing system with multiple users, where the
4) We present simulation results that illustrate the effective- radio and computational resources, and offloading decisions,
ness and performance of the proposed mechanism and can be dynamically allocated with the variation of computa-
algorithms. tion demands, radio channels and the computation resources.
The remainder of this article is organized as follows. Other scholars [16] combined fog computing and the Internet
Section II gives a review of related work. Section III introduces of Vehicles (IoV), to design an offloading scheme for real-
the system model and problem model. Section IV discusses time traffic management in fog-based IoV systems, ensuring
the challenges of applying a double auction scheme, and then low latency services for end users.
proposes the TAFS auction scheme. Section V presents the Despite these and other valuable contributions to the
economic properties of the TAFS scheme. Section VI pro- literature, none of the aforementioned works has studied
vides performance evaluation results. Section VII concludes incentive mechanisms for FNs supporting delay-sensitive
this article. workload.

II. R ELATED W ORK B. Incentives

A. Offloading and Resource Allocation in Fog Computing Sun et al. [17], Liu et al. [18], and Hasana et al. [19]
Offloading is very attractive for emerging IoT and cloud explored the incentive mechanism in edge computing.
computing applications. There have been many previous stud- Sun et al. [17] proposed two MEC dynamic pricing double
ies on computational offloading in the context of mobile-edge auction schemes for the Industrial IoT (IIoT) to encourage
computing (MEC) [7]. In these studies, the offloading task was mobile devices and edge servers to participate in offloading. To
restricted to a single hop. Fog computation offloading, con- model the cloud service operator and edge server owners par-
versely, is hierarchical: several nodes form a network, while ticipating in computation offloading, the interaction between
the MEC relies on a separate node that does not represent a them was designated as a Stackelberg game in [18] with the
network. goal of maximizing their respective utilities by obtaining an

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 02,2024 at 11:41:02 UTC from IEEE Xplore. Restrictions apply.
 3254 IEEE INTERNET OF THINGS JOURNAL, VOL. 10, NO. 4, 15 FEBRUARY 2023

TABLE I
E XISTING I NCENTIVE S CHEMES

optimal payment. Hasana et al. [19] presented a highly local- Fig. 2. Local fog computing system offloading model.
ized and mobile ad hoc cloud computing model which runs
a reward scheme for rating and penalizing IoT worker nodes
based on task performance, task offloading, partial status, and offloading service? In addressing these questions, we use the
task outcome preservation. local fog computing system as an open market and model the
In [20], a dynamic resource allocation method based on problem of UA offloading in the system as a double auction.
pricing was developed in a fog computing environment. This is a two-stage process consisting of an assignment stage
According to the various QoS requirements of user tasks, and a pricing stage. Below, we will also explain our design
on-demand billing, daily billing, and auction billing meth- goals.
ods were designed in which resources are allowed to be
overbooked. However, the authors made no mentioning of
motivating participants. A. Local Fog Computing System
Recent work in [21]–[25] centered on incentive mecha- We argue that a trusted platform is an ideal platform for the
nisms for different objects, but not FNs within fog computing free trading of computational resources and demands between
environments. Considering a proof-of-work-based blockchain fog providers and application users. A similar approach has
network, Jiao et al. [21] focused on the transaction between been taken in the cloud computing market [26] to reduce cap-
cloud/fog computing service providers and miners with the ital expenditures by reusing an existing infrastructure for users,
goal of motivating miners. Chang et al. [22] formulated a boosting resource utilization for Thing-owners, and enabling
two-stage Stackelberg game between the miners and the edge more flexible pricing options (e.g., pay-as-you-go) for com-
service provider (ESP) in order to find the optimal incentive puting or network providers. Inspired by the Fog-of-Things
for the ESP and miners to choose autofit mining strate- (FoT) paradigm presented in [6], in this study, the IoT provider
gies. Others [23] encouraged an operator provider (OP) to (shown in Fig. 1) in FoT is a third party for the incentive mech-
invest in computing and network resources at the edge of the anism. Fig. 2 shows a local fog computing system that consists
network, in order to accommodate requests from application of a fog network, smart-device UAs, and an IoT provider.
service providers (ASPs). The OP negotiates with ASPs to The UA (illustrated by a rectangle in the figure) is gener-
determine the serving QoS of applications as well as how ated by the IoT terminal in reference to terminal devices, such
much participants are paid. In a novel mobile task offload- as smart mobile phones, sensor nodes, and cameras. These
ing D2D fogging framework, an energy-saving incentive was devices collect a large amount of data to be stored or ana-
proposed [24] to prevent overexploiting and free-riding behav- lyzed in a timely manner. However, IoT terminal resources are
iors. A joint optimization framework for FNs, data service restricted and many applications require real-time responses,
operators (DSOs), and data service subscribers (DSSs) was so the user must make an application request for storage
developed in [25] where DSOs rent a computation resource or computing resources from the IoT provider. Here, we
block (CRB) from FNs to serve DSSs. Researchers [25] have assume that all applications are independent and indivisible
also explored the pricing for DSOs - both for profitability and as such that any application can only be offloaded to at most
to motivate DSSs to purchase large quantities of CRBs. one FN.
The studies referenced above regarding offloading and An FN (illustrated by a circle in the figure) is arbitrar-
resource allocation in the fog computing context did not con- ily distributed in the fog computing layer. All FNs in the
sider communication delay along the transmission path in the local area together form the fog network. FNs usually have
fog network. The difference between these related works and richer resources than IoT terminals, so they can provide
the present study is highlighted in Table I. paid service for offloading applications by using their sur-
plus resources. A single FN can handle multiple offloaded
applications simultaneously.
III. S YSTEM M ODEL AND P ROBLEM M ODEL In a real system, the computing resources of FNs generally
Who coordinates the interactions among the FNs, between include CPU resources and memory resources, which can be
FNs and users, or between the fog layer and the cloud managed by the IoT provider in a manner similar to cloud
layer? Who is responsible for the maintenance of the whole computing, i.e., via virtualization. For example: 1) the CPU

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 02,2024 at 11:41:02 UTC from IEEE Xplore. Restrictions apply.
 SUN et al.: TAFS: TRUTHFUL AUCTION FOR IoT APPLICATION OFFLOADING IN FOG COMPUTING NETWORKS 3255

can be virtualized into multiple CPUs running in parallel. Each

offloading application can use one or more virtual CPUs, and
the virtualized CPUs in the FN can be isolated from each other
and 2) virtualized memory can be uniformly managed, serving
well for the offloading application through the virtualization
encapsulation technology.
As a trustworthy third-party platform, the IoT provider man-
ages resource optimization and QoS [6]. The IoT provider
has two key modules: 1) fog network maintenance module
and 2) double auction module. 1) The fog network mainte-
nance module regularly and dynamically updates the location Fig. 3. Illustration of FB, FN, and UA.
information of the multihop fog network nodes within the
local area, estimates or measures the communication delay
(and cost) between the network nodes, accounted for when
computing the shortest path, and informs the double auction this may be related to, for example, the occupied computing
module. 2) The double auction module collects UA requests capacity.
and available resource information from FNs, makes central At the beginning of the auction, buyer ui ∈ U submits the
offloading decisions, and feeds back to the user and FN. Each resource demand di (computing resource in the FN is virtual-
time the double auction module conducts an auction, it obtains ized into CRB [25], as the unit quantity of measurement for
the most recent delay measurement data from the fog network counting), maximum tolerance delay ti , and the bid vector of
j
maintenance module to compute the shortest path from the Bi = (B1i , B2i , . . . , Bm
i ) to the auctioneer, where Bi is the bid
application to each FN. These two modules are independent of the designated seller fj . Bi may or may not be the same
and parallel. This assumption benefits resource-consuming as Vi . Let B =(B1 ; B2 ; · · · ; Bn ) represent the bid matrix con-
applications, because within the range of delay tolerance, more sisting of the bid vectors of all buyers. Seller fj ∈ F submits
FNs are reachable across multiple hops for offloading deci- the amount of available resource rj (also counted as the CRB
sions, than when only a single-hop edge server is available as number) and an ask price Aj . Similarly, the ask Aj may or
in MEC. In the given system, dynamics can happen due to the may not the same as cj . Let A = (A1 , A2 , . . . , Am ) represent
joining, departure, failure, or renewal of FNs. We assume that the set of prices submitted by all sellers. Note that Aj is the
FNs are static within each decision (auction) period, which we unit price, i.e., the ask for each CRB. The auction referred to
call “semistatic.” in this article is a sealed-bid auction. We assume that bids and
asks remain the same in a single auction. After receiving B
and A, the auctioneer performs the auction in the following
B. Problem Model two stages.
Consider a local fog system consisting of n UAs and m FNs. 1) The assignment stage, where FNs are assigned to the
We use U = {u1 , u2 , . . . , un } to denote the set of UAs and requesting UAs; this involves deciding the set of can-
F = {f1 , f2 , . . . , fm } to refer to the set of FNs. We formulate didate winning buyers and the set of candidate winning
the fog computing offloading decision as a single-round–multi- sellers.
item double auction [27]. In this auction, as shown in Fig. 2, 2) The pricing stage, where the set of winning buyers,
UAs are buyers, FNs are sellers, and the IoT Provider is an the set of winning sellers, and the clearing prices are
auctioneer. Throughout this article, we may use the terms UA determined.
and buyer, FN and seller, and IoT provider and auctioneer Under the assumptions above, there may be a many-to-one
interchangeably. assignment stage wherein multiple UAs are assigned to the
Each auction is independent, and the auction frequency can same FN. However, the pricing stage must have a one-to-one
be dynamically adjusted according to the interauction interval trading relationship where one buyer corresponds to one seller.
and the number of waiting users. Presetting two thresholds for Definitions related to the problem are given below, and their
them, the new auction can be started when any one of the two usage in our mechanism is further elaborated in Section IV.
threshold is reached. The problem is illustrated in Fig. 3.
Buyers bid for offloading services. For a given buyer, a Definition 1 (Candidate FN): In the assignment stage, given
different offloading path and FN for offloading results in a UA set U and FN set F, the auctioneer determines the can-
didate UA set Uc ⊆ U and the candidate FN set Fc ⊆ F. A
j
different acquisition value. Let Vi be buyer ui ’s true valuation
of offloading service from fj , which also represents the true many-to-one mapping  : {i : ui ∈ Uc } → {j : fj ∈ Fc } is
price that ui is willing to pay for the offloading service. Let the result of this stage. −1 (j) = {i | (i) = j} is the set of
Vi = (Vi1 , Vi2 , . . . , Vim ) be the true valuation vector of buyer ui . indices of UAs to which fj is assigned.
Sellers offer offloading services at the cost of resources, e.g., Definition 2 (Fog Block): Given Uc , Fc , and (·), since one
computing power, and receive monetary payment in return. candidate FN may be assigned to several candidate UAs in
Note that Vi is the private information to each buyer, and is not the assignment stage, any fj ∈ Fc can be seen as consisting
known to other buyers and/or the auctioneer in general. Let cj of several fog blocks (FBs). Let Fc = {fjk | (i) = j, ui ∈
be seller fj ’s true unit cost of providing the offloading service; Uc , k = 1, 2, . . . , |Uc |} be the set of all FBs assigned to the

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 02,2024 at 11:41:02 UTC from IEEE Xplore. Restrictions apply.
 3256 IEEE INTERNET OF THINGS JOURNAL, VOL. 10, NO. 4, 15 FEBRUARY 2023

TABLE II (i)
N OTATIONS its ask, i.e., Pbi ≤ Bi for all ui ∈ Uw , and Psj,k ≥
Aj · dφ −1 (k) for all fj ∈ Fw .
k

3) Budget Balance: As an auctioneer, the IoT provider

needs to maintain the service. Therefore, the payment
received from buyers should  be no less than
 the total
price paid to the sellers, i.e., ui ∈Uw Pbi ≥ fjk ∈Fw Psj,k .
4) Delay Bound: Each UA’s delay requirement must be
satisfied.
5) Utilization Maximization: User demands are met by
maximally utilizing the FNs’ resources.
Designing a scheme that fully satisfies the above five goals
is extremely challenging. Our ultimate goal is to motivate
the participation of FNs and IoT users in offloading appli-
cations while preventing any party from rigging its ask or
bid to manipulate the market. Therefore, designing an auction
scheme with the first three properties has the highest priority.
The secondary priority involves system performance, specif-
ically, maximizing the resources allocated to meet the user’s
delay requirements.

IV. TAFS S CHEME D ESIGN

This section begins with a description of the issues inherent
to applying McAfee’s double auction [27]. We then describe
our solution, TAFS, which consists of two relatively indepen-
candidate UAs. Clearly, |Uc | = |Fc |. FB fjk belongs to the
dent stages. Stage 1, is the assignment stage, which achieves
candidate FN fj , the entity assigned to the kth candidate UA,
goals 4 and 5 associated with system performance. We for-
which can be expressed as a one-to-one mapping φ : {i : ui ∈
mulate it as an optimization problem named the candidate
Uc } → {k : fjk ∈ Fc }. φ −1 (k) is the index of the UA to which
FN assignment (CFA). Stage 2, is the pricing stage, which
fjk ∈ Fc is assigned. In this article, FB is the so-called seller accomplishes goals 1–3 related to economic characteristics.
entity, and FN is the so-called seller.
Let Uw ⊆ U be the winning set of buyers, i.e., the winning
UA set. Let Fw ⊆ Fc be the winning set of seller entities, i.e., A. McAfee Double Auction
j
the winning FB set. Fw = {fjk | fjk ∈ Fw } is the winning set of In the McAfee double auction [27], each buyer submits a
seller entities in the candidate FN fj . Let Pbi be the price that bid Bi , and each seller submits an ask Aj . The auctioneer sorts
the winning buyer ui needs to pay and Psj,k be the payment all bids in the descending order, Bi1 ≥ Bi2 ≥ · · · ≥ Bin , and
the auctioneer remits the winning seller entity fjk . The utility sorts all asks in the ascending order, Aj1 ≤ Aj2 ≤ · · · ≤ Ajm .
of buyer ui ∈ U is defined as The auctioneer then makes pricing decisions according to the
 (i) sorted bid and ask lists. The sorting process is similar to the
Vi − Pbi , if ui ∈ Uw
Ui =
b
(1) TASC scheme designed by Yang et al. [28].
0, otherwise. Applying the McAfee double auction to our local fog
The utility of seller fj ∈ F is the sum of all utilities of seller computing system offloading model creates two problems.
j
entities fjk ∈ Fw , which is defined as Q1: If the FN acts as a seller, each seller may trade with
   multiple buyers at the same time. How then should the seller
j
Psj,k − cj · dφ −1 (k) , if fjk ∈ Fw sets its ask price? It is unreasonable to ask for the same total
Uj =
s k (2)
0, otherwise. price for multiple buyers with different resource demands.
Q2: On the other hand, if a unit price is asked, then how
The notations in this article are summarized in Table II. to reasonably and fairly sort the prices?
Our answers to these questions are twofold.
C. Scheme Design Objectives 1) We introduce the FBs as the seller entities as in
The objectives of our design are as follows. Definition 2. After stage 1, each candidate FN is split
1) Truthfulness: To keep the auction truthful, the scheme into several FBs as seller entities, corresponding to
should guarantee that any buyer (resp. the seller) cannot buyers one-to-one, as shown in Fig. 3.
improve its utility by submitting a bid (resp. ask) dif- 2) The seller submits the ask (i.e., unit price) while the
ferent from its true valuation (resp. the true unit cost) seller entities participate in sorting; therefore, the price
regardless of what others submit. waiting to be sorted is an ask multiplied by the amount
2) Individual Rationality: No buyer is charged more than of resource demand of the corresponding buyer assigned
its bid. No seller offers a service at a payment less than to this seller.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 02,2024 at 11:41:02 UTC from IEEE Xplore. Restrictions apply.
 SUN et al.: TAFS: TRUTHFUL AUCTION FOR IoT APPLICATION OFFLOADING IN FOG COMPUTING NETWORKS 3257

Algorithm 1: Candidate Assignment (U, F, G)

1 Uc ← ∅ , F c ← ∅, (i) ← 0 for all ui ∈ U;
// Find Feasible FN for each UA
2 for each ui ∈ U do
3 Fi ← ∅ ;
4 for each fj ∈ F do
j
5 Take wi from pre-computed all-pair shortest
paths;
j
6 if wi ≤ ti then
7 Fi ← Fi ∪ {fj };
Fig. 4. Example of feasible FN. 8 end
9 end
10 end
For the bid sequence B(i i1
1)
≥ B(i
i2
2)
≥ · · · and the ask 11 Solve Eq. (3)-(6) by using Algorithm 2 or other
sequence Aj1 · dφ −1 (k1 ) ≤ Aj2 · dφ −1 (k2 ) ≤ · · · , the auctioneer algorithm;
(im ) for each ui ∈U, each fj ∈ Fi do
finds the largest m that satisfies Bim ≥ Ajm · dφ −1 (km ) , and 12

then selects the buyer winners and the seller entity winners to 13 if xij = 1 then
determine the price. 14 Uc ← Uc ∪ {ui }; Fc ← Fc ∪ {fj }; (i) ← j;
15 end
16 end
B. Stage 1: Assignment Stage 17 return (Uc , F c , )
Definition 3 (CFA Problem): Given fog network G, UA set
U, and FN set F, the CFA problem is to assign the applications
in U to FNs in F, satisfying three conditions: 1) the total Algorithm 2: Urgency+Greedy-Based Algorithm
amount of resource demand of UAs assigned to an FN does not
exceed the amount of its available resources; 2) each assigned 1 for each Fi do
UA’s delay requirement is satisfied; and 3) the total amount 2 sort all the FNs in Fi in descending order per their
of resources allocated are maximal. The CFA problem can be available CRBs;
3 end
formulated as follows:
4 Sort all the UAs in U in ascending order per their

max di · xij (3) maximum tolerance delay ti ;
ui ∈U fj ∈Fi 5 for each ui ∈ U do
 6 Take the FN from Fi in turn, find the first j, such that
s.t. di · xij ≤ rj for j = 1, 2, . . . , m (4) di ≤ rj ;
ui ∈U
 7 xij ← 1, rj ← rj − di , repeat sorting for Fi
xij ≤ 1 for i = 1, 2, . . . , n (5) 8 end
fj ∈Fi
xij ∈ {0, 1} (6)

where the 0-1 variable xij denotes whether ui is assigned to the delay requirement of ui . Thus, the path ui → f2 → f1 is
fj . The objective function (3) serves to maximize the allocated excluded, and ui cannot be offloaded to f1 .
resources, which is the design objective 5 referenced above. Definition 4 (Feasible FN): Given any ui ∈ U with a
Constraint (4) bounds the total amount of resources assigned maximum tolerance delay ti , and minimal latency vector
j
to any FN by its amount of available resource. fj ∈ Fi in Wi =(Wi1 , Wi2 , . . . , Wim ), wi is the shortest latency from ui to fj .
Constraints (3) and (5) state that the feasible assignment of Let Fi ⊆ F be the Feasible FN set, where each fj ∈ Fi satisfies
j
ui to fj satisfies condition 2. For the definition of a feasible wi ≤ ti .
assignment, see Definition 4 below. Constraints (5) and (6) We use the heuristic algorithm (Algorithm 1) to implement
specify that the application to be offloaded is indivisible, and stage 1.
so it can only be assigned to at most one FN. We can prove that Lines 2–10 find the Feasible FN set Fi for each ui ∈ U.
the CFA problem is NP-hard by a polynomial-time reduction Previous researchers have reported the multiple knapsack
from the 0-1 knapsack problem. Therefore, we seek effective problem with assignment restrictions (MKARs) problem solu-
heuristic algorithms rather than trying to design polynomial- tions corresponding to line 11 [29]. The auctioneer can choose
time optimal algorithms. different matching algorithms for different purposes. However,
To satisfy condition 2, the IoT provider needs to compute under the real-time conditions of the fog computing scenario,
j
the latency wi from ui ∈ U to all fj ∈ F by using the shortest it is necessary to select a method with low computational
path algorithm. As shown in Fig. 4, the minimal latency from complexity. Here, we use three fast heuristic algorithms:
ui to f4 is no more than the maximum tolerance delay of ui , 1) urgency + greedy based (UGB); 2) urgency-only based
i.e., w4i ≤ ti , while the shortest path from ui to f1 cannot meet (UOB); and 3) random assignment (RA). The UGB algorithm

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 02,2024 at 11:41:02 UTC from IEEE Xplore. Restrictions apply.
 3258 IEEE INTERNET OF THINGS JOURNAL, VOL. 10, NO. 4, 15 FEBRUARY 2023

Algorithm 3: Winner Determination & Pricing Algorithm 4: TAFS(U, F, B, A, G)

(Uc , Fc , , B, A) c , )←Candidate Assignment(U, F, G);
1 (Uc , F
1 Fc ← ∅, k ← 1, φ(i) ← 0 for all ui ∈ Uc , b s
2 (Uw , Fw , P , P )←Winner Determination & Pricing
Uw ← ∅, Fw ← ∅; 
(Uc , Fc , , B, A);
// Split each FN in Fc into FBs 3 return (Uw , Fw , P , P );
b s
2 for each ui ∈ Uc do
3 Fc ← Fc ∪ {f(i) k }; φ(i) ← k; k ← k + 1;

4 end
5 Sort all the buyers in Uc to get an ordered list
U = ui1 , ui2 , · · · such that B(i i1
1)
≥ B(i
i2
2)
···;
6 Sort all the seller entities in Fc to get an ordered list
F = fjk1 1
, fjk2
2
, · · · such that Aj1 · dφ −1 (k1 ) ≤ Aj2
·dφ −1 (k2 ) · · ·
(im )
7 Find the largest m, such that Bi
m
≥ Ajm · dφ −1 (km )
8 if m < 2 then
9 return( Uw , Fw , 0, 0);
10 end
11 (x, y, z)← (im , jm , km );
// Determine price and payment
(x)
12 Pb ← Bx , Ps ← Ay · dφ −1 (z)
// Sacrifice one buyer and one seller
entity to ensure the truthfulness
13 Uw ← Ux \ {ux }, Fw ← Fx \ {fy };
z

// Determine the final winners

14 for each ui ∈ Uw do
φ(i)
15 if f(i) ∈ / Fw then
16 Uw ← Uw \ {ui };
17 end
18 end
19 for each fj ∈ Fw do
k

20 if uφ −1 (k) ∈ / Uw then
21 Fw ← Fw \ {fjk }; Fig. 5. Example with five UAs and seven FNs: Stage 1. (a) Bid matrix B
and the assignment. (b) Resource demand and maximum tolerance delay of
22 end UAs. (c) Amount of available resources and asks of FNs. (d) Shortest latency
23 end matrix.
24 return (Uw , Fw , P , P )
b s

seller entities in F. (U, F) denotes the set of matches induced

by U and F, i.e., (U, F) = {(ui , fjk ) : ui ∈ Uc , fjk ∈ Fc , j =
is shown in Algorithm 2. The UOB algorithm only sorts UAs (i), k = φ(i)}. The UA-to-FB pair, per which the auctioneer
in U and finds the satisfying FN for each sorted ui in turn (i.e., determines the clearing price, is called the boundary pair. In
lines 4–8 of Algorithm 2) The RA algorithm finds the satisfy- Algorithm 3, (uim , fjkmm ) is such a pair. As the buyer winner
ing FN for each UA in U in a random manner (i.e., lines 5–8 of φ(i)
Algorithm 2). The returned values include the candidate UA and the seller entity winner are generated in pairs, (ui , f(i) )
set Uc , the candidate FN set Fc , and the assigned mapping or (uφ −1 (k) , f(φ
k
−1 (k)) ) is the so-called winning pair.

(·). The main TAFS algorithm is shown in Algorithm 4.

C. Stage 2: Pricing Stage D. Illustrative Example

After the assignment result is gathered in stage 1, we can We take a simple example to demonstrate the core idea
apply a double auction to determine the final winners Uw and of TAFS. Fig. 5(a) and (b) shows the bid, resource demand,
Fw and clearing prices Pb and Ps via Algorithm 3. and maximum tolerance delay that each buyer submits to the
Exploration: U denotes a sorted candidate buyer sequence auctioneer. Fig. 5(c) shows the available resources and the ask
from the stage 1 output, which is in the nonincreasing order submitted by each seller to the auctioneer. Fig. 5(d) shows the
of their bids. F denotes a sequence of candidate seller entities, precomputed shortest latency matrix. Cases that fail to satisfy
which are sorted in nondecreasing order of their asks according the delay requirement of ui → fj are highlighted in the figure,
to the method described in Section IV-A. Ux denotes the sublist where fj ∈/ Fi . The assignment result of stage 1 is marked in
of the first x buyers in U. Fx denotes the sublist of the first x Fig. 5(a), where u1 is assigned to f2 as an example.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 02,2024 at 11:41:02 UTC from IEEE Xplore. Restrictions apply.
 SUN et al.: TAFS: TRUTHFUL AUCTION FOR IoT APPLICATION OFFLOADING IN FOG COMPUTING NETWORKS 3259

Fig. 7. Illustration for Lemma 2.

j
the result holds even after increasing the bid Bi or decreasing
the ask Aj . Under Lemmas 4 and 5, if fjk (resp. ui in Lemma 5)
is still the winner after giving a different ask (resp. different
bid in Lemma 5), the clearing price remains the same.
Due to space limitations, we prove the lemmas for sellers
(Lemmas 2, 4, and 6) and only one (Lemma 3) for buyers here.
Other properties for buyers (Lemmas 5 and 7) can be proven
Fig. 6. Example with five UAs and seven FNs: Stage 2. (a) Assignment
in a similar (and easier) manner as we prove the lemmas for
stage result: (1) = 2, (2) = 4, (3) = (4) = (5) = 3; φ(1) = 1, sellers.
φ(2) = 5, φ(3) = (2), φ(4) = 3, φ(5) = 4; Uc = {u1 , u2 , u3 , u4 , u5 }, Fc = Lemma 1: Assuming ui is assigned to fj in stage 1, the
{f2 , f3 , f4 }, Fc = {f21 , f32 , f33 , f34 , f45 }. (b) Bipartite graph showing the pricing −j
stage. auction result for ui is independent of its bid on Bi , which
j
includes the bids in Bi except Bi .
Proof: As mentioned in Section IV, the assignment and
In stage 2, the candidate FNs are further split into the can- pricing stages are entirely separate. The whole assignment pro-
didate FBs [Fig. 6(a)]. The candidate buyers and seller entities cess is independent of any bids. On the other side, once ui is
sorting and pricing results can be represented by a bipartite assigned to fj , whether ui is determined to be a winner or not
j
graph, as shown in Fig. 6(b). The auctioneer finds m = 4, the depends only on Bi . Therefore, our lemma holds.
k
winning pairs being (u5 , f34 ) and (u4 , f33 ). In this case, each Lemma 2: If fj wins with ask Aj , then it still wins with
buyer pays 4.5, each seller gains 4, and the auctioneer earns Ãj < Aj .
2 × (4.5 − 4) = 1. Proof: According to Lemma 1, if (i) = j and φ(i) = k
for winner fjk with Aj , then the same is true for f˜jk (i.e., similar
to FB as fjk only with a different ask Ãj ). Similarly, in the
V. A NALYSIS
pricing stage, Uc , Fc , Fc , φ, and  are all unchanged even if
In this section, we prove that TAFS has the three economic the ask of fj decreases, which moves the position of fjk forward
properties mentioned in Section III-C. in the nondecreasing order set F.
Theorem 1: The TAFS scheme is individually rational. Fig. 7 shows that the order of FBs between the new posi-
Proof: Since Ux is a nonincreasing ordered list, and Fx tion of f̃jk and the position of fjk are not affected by Ãj .
is a nondecreasing ordered list, for each winning buyer ui ∈
(i) (i ) We take another winner FB fjkaa in F, where Ãj · dφ −1 (k) ≤
Uw ⊆ Ux , we know that Bi ≥ Bix x = Pb . For each win-
j Aja · dφ −1 (ka ) ≤ Aj · dφ −1 (k) , as an example. Previously,
ning seller entity fjk ∈ Fw ⊆ Fx , Psk,j = Ps = Ajy · dφ −1 (z) ≥ (i )
 s  Aja · dφ −1 (ka ) ≤ Bia a ; now, its position has moved forward,
Aj · dφ −1 (k) , and we have k Pj,k ≥ Aj · k dφ −1 (k) . This (i )
completes ours proof. Aja · dφ −1 (ka ) ≤ Aja+1 · dφ −1 (ka+1 ) ≤ Bia+1a+1 so that (uia+1 , fjkaa )
Theorem 2: The TAFS scheme is budget balanced. is the winning pair. The boundary pair selection remains the
Proof: Every winning buyer ui ∈ Uw needs to pay same as before, which implies that f̃jk still wins.
(i)
B(i
ix
x)
= Pb and every winning seller entity receives Ajy · Lemma 3: If ui wins with bid Bi , then it still wins with
(i) (i)
dφ −1 (z) = Ps . We know that winners are one-to-one UA-to- B̃i > Bi .
(i )
FB pairs, i.e., |Uw | = |Fw |, and Bix x ≥ Ajy · dφ −1 (z) , and so Proof: According to Lemma 1, we know that ui is still
 (i)

− fjk ∈Fw Aj · dφ −1 (k) = |Uw | · (Pb − Ps ) ≥ 0. This assigned to f(i) , in stage 1, and all candidate winners, i.e.,
ui ∈Uw Bi (i)
completes the proof. FBs and UAs, are unchanged even if the bid changes to B̃i .
(i) (i)
Theorem 3: The TAFS scheme is truthful. As the bid increases, i.e., B̃i ≥ Bi , the position of ui
This proof necessitates a group of lemmas. First, note that in the sorted set U should move forward or be unchanged as
the results of the auction remain the same regardless of how shown in Fig. 8. We know B̃(i) i > B(i)
i ≥ B(x)
x = Pb , so the
the winners change their bids or asks. Given that ui is assigned value of m remains constant and, thereby, the boundary pair
to fj in stage 1, Lemma 1 shows that as long as ui does not is the same; ui still wins.
change its bid on fj , the result for ui does not change. Lemmas Lemma 4: If fjk wins after asking Aj or Ãj , then it is given
2 and 3 show that if we already have a winning pair (ui , fjk ), the same payment: Ps = P̃s .

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 02,2024 at 11:41:02 UTC from IEEE Xplore. Restrictions apply.
 3260 IEEE INTERNET OF THINGS JOURNAL, VOL. 10, NO. 4, 15 FEBRUARY 2023

For f̃jk , as the ask decreases, the boundary pair

may move forward giving P̃s two possible states:
P̃s = Ps , where Ũj,k s = P̃s − c · d
j φ −1 (k) < 0, or
P̃ > P , where another boundary pair is selected
s s

but must still be in front of the previous position of

Fig. 8. Illustration for Lemma 3. fjk , unless fjk won previously. Thus, P̃s < cj ·dφ −1 (k) ;
s = P̃s − c · d
Ũj,k j φ −1 (k) < 0.
j j
c) fj and f̃j all lose, i.e., fjk ∈
k k / Fw , and f̃jk ∈/ F̃w . We
Proof: According to Lemma 2, we know that even if the have
ask of fjk changes, Uc , Fc , Fc , φ, and  all remain unchanged. s
Ũj,k = Uj,k
s
= 0.
Moreover, both fjk and f̃jk win, which means Ãj · dφ −1 (k) ≤ Ajy ·
dφ −1 (z) and Aj · dφ −1 (k) ≤ Ajy · dφ −1 (z) , while the boundary pair j j
In subcases a) and c), |Fw |= |F̃w |; a reduction in
is unchanged. According to the pricing strategy, the payment
the ask does not change the utility of every fjk . In
should be the same, i.e., Ps = P̃s = Ajy · dφ −1 (z) .
(i) (i)
subcase b), the utility cuts back. In general,
Lemma 5: If ui wins by bidding Bi or B̃i , then it is  
Ũjs = f̃jk ∈F̃wj Ũj,k ≤ fjk ∈Fwj Uj,k = Uj .
s s s
charged the same price; Pb = P̃b .
Lemma 6: TAFS is truthful for sellers.
Proof: We prove this lemma by showing that no seller fj In summary, a seller cannot improve its utility by submitting
can improve its utility by asking Aj = cj , i.e., Ũjs ≤ Ujs for any a different ask from its true cost. This completes the proof.
Aj = cj , when Ũjs and Ujs are utilities with fj asking Aj and Lemma 7: TAFS is truthful for buyers.
cj , respectively. Any seller fj can include several seller entities Proof of Theorem 3: Lemmas 6 and 7 together prove that
j TAFS is truthful.
fjk . For simplicity, let f̃jk and F̃w be any seller entity and the
Theorem 4: The time complexity of TAFS is O(T +nm+l2 ),
set of winning seller entities in fj with Aj , respectively; let fjk where T is the time complexity of line 11 in Algorithm 1.
j s and Ũ s be the utilities of
and Fw be those with cj . Let Uj,k j,k The assignment stage (Algorithm 1) takes place both offline
k k
fj and f̃j . and online. Lines 2–10 maintained by the IoT provider can run
1) Case 1: Aj > cj . offline and the time complexity can be ignored. Therefore, the
For fjk ∈ Fc , there are three subcases. assignment stage has a time complexity of O(T + nm).
a) fjk wins by asking cj and f̃jk wins by asking Aj , i.e., In the pricing stage (Algorithm 3), the input is the result
j j
fjk ∈ Fw , and f̃jk ∈ F̃w . Under Lemma 4, we have from the assignment stage (Algorithm 1), i.e., the number of
s = Ps − c · d candidate buyers. Let l = |Uc |. Clearly, l ≤ n. Sorting both
φ −1 (k) = P − cj · dφ −1 (k) = Ũj,k .
Uj,k s s
j,k j
k k
sellers and buyers takes O(l log l) time (lines 5 and 6). Line 7
b) fj wins by asking cj but f̃j loses by asking Aj , i.e., takes O(l) time. Lines 13–18 take O(l2 ) time. Therefore, the
j j
fjk ∈ Fw , and f̃jk ∈ / F̃w . We have time complexity of the pricing stage is O(l2 ).
Ũj,k = 0 ≤ Uj,k = Ps − cj · dφ −1 (k) .
s s In summary, the overall time complexity of TAFS is O(T +
c) fjk and f̃jk all lose, i.e., fjk ∈
j
/ Fw , and f̃jk ∈
j
/ F̃w . We nm + l2 ).
have
s = U s = 0.
Ũj,k j,k VI. N UMERICAL R ESULTS
j
In this case, for each fjk ∈ Fw , increase in the ask We ran a series of simulations to evaluate the performance
does not improve utility. The lemmas above indi- of TAFS in MATLAB. We focused on economic effects and
cate that it is impossible that fjk loses by asking cj system efficiency.
j j
but f̃jk wins by asking Aj . Therefore, |Fw | ≥ |F̃w |.
We have   A. Experiment Settings
Ũjs = f̃jk ∈F̃wj Ũj,k ≤ fjk ∈Fwj Uj,k = Uj .
s s s
We used a random fog network topology generated by the
2) Case 2: Aj < cj . Watts–Strogatz small-world graph model [30], where K = 3,
There are also three subcases for fjk ∈ Fc . p = 0.2, and the number of FNs varies between 10 and 25 with
a) fjk wins by asking cj and f̃jk wins by asking Aj , i.e., the link weights generated randomly in [1, 5] ms. For each
j j
fjk ∈ Fw , and f̃jk ∈ F̃w . Under Lemma 4, we have setting, we randomly generated 100 instances and averaged
the results. All tests were implemented on a Windows PC
s
Uj,k = Ũj,k
s
= Ps − cj · dφ −1 (k) . with an Intel-Core 1.60-GHz CPU and 8-GB memory.
For the auction, the bids and asks are the prices that buyers
and sellers can accept, respectively. In practical applications,
b) fjk loses by asking cj and f̃jk wins by asking Aj , i.e., buyers and sellers can be provided with a guided quotation
j j
fjk ∈
/ Fw , and f̃jk ∈ F̃w . For fjk , we have interval to help with their valuations. In this article, the nor-
malized range of asks is set to (0, 1], and the normalized
s
Uj,k = 0 and Ps < cj · dφ −1 (k) . range of bids is set to (0, Vmax ], similar to the settings in [28]

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 02,2024 at 11:41:02 UTC from IEEE Xplore. Restrictions apply.
 SUN et al.: TAFS: TRUTHFUL AUCTION FOR IoT APPLICATION OFFLOADING IN FOG COMPUTING NETWORKS 3261

Fig. 9. Utilities of [(a)–(c)] a buyer and [(d)–(f)] a seller in auctions with different assignment algorithms, where n = 10 and m = 30. (a) UGB. (b) UOB.
(c) RA. (d) UGB. (e) UOB. (f) RA.

where Vmax is set to 4. We varied Vmax to observe the influ- IoT provider, and social welfare when the number of sellers
ence of different values on the auction success rate, shown m = 10. Fig. 10(e) shows the effects of changes in the number
in Fig. 11(b). Finally, Vmax was set to 6 in other experi- of buyers on the IoT provider’s utility by taking the RA algo-
ments, because the median performance was obtained when rithm as an example with m = 10, 15, 20, 25, 30, respectively.
the bid is (0, 6]. Regarding the delay settings, since some pop- It is interesting to find that in the case of a small number of
ular applications, such as augmented reality and autonomous users (when smaller than 70, 50, and 60 for users, fogs, and
driving, require millisecond response time, the maximum toler- social welfare, respectively), the no-delay requirement had no
ance delay was randomly generated between [20, 30] ms. The obvious effect on the utility. The utility under UGB-IDT was
number of CRBs requested by the buyer and available CRBs even lower than that under the other three algorithms, whereas
by the seller was randomly generated in [2, 3] and [30, 50], when the number of users was larger, the utility, especially of
respectively. fogs under the UGB-IDT algorithm, was significantly better
The performance metrics in the experiments included buyer than other algorithms. This is due to the sealed auction.
utility, seller utility, the utility of IoT provider, the ratio of 3) Impact on Efficiency: Here, system efficiency is
successful transactions to buyer requests, and running time. expressed as the ratio of the number of successful transactions
We compared the three algorithms mentioned in Section IV-B. to the total number of requests. As discussed in Section III-C,
In order to observe the effect of delay tolerance on TAFS the system efficiency of TAFS may also be affected due to
performance, we also considered UGB (Algorithm 2) in the a mismatch between the bids of buyers and asks of sellers.
infinite delay-tolerance (IDT) scenario, where all applications’ To capture the economic impact on system efficiency, we
maximum delay tolerance was infinite. Throughout all the plotted the ratio of successful transactions to buyer requests
experiments, we denoted these algorithms as UGB, UOB, RA, over the three assignment algorithms as shown in Fig. 11(a).
and UGB-IDT, respectively. Surprisingly, the ratio was not significantly affected by the
number of sellers. With the increase in the number of users,
the success rate slightly decreased, but even the number of
B. Results users climbed 200, that is, when the total demand for offload-
1) Truthfulness of TAFS: To verify the truthfulness of ing was greater than the amount of available resources that the
TAFS, we randomly selected one buyer and one seller to test fog could provide within the area, the success rate could still
how their utility change if they change their bid or ask. As be maintained at more than 70%.
shown in Fig. 9, in each auction, V was the true valuation of Next, we examined the impact of the bid distribution on
the buyer and C was the true cost of the seller. Two different system efficiency. For different Vmax , Fig. 11(b) shows the
values were tested for both the buyer’s true valuation and the successful transaction rate of TAFS when using the three
seller’s true cost. For each different true value (resp. cost), the respective algorithms. The success rate under the UGB, UOB,
buyer (resp. seller) could not improve its utility by submitting RA, and UGB-IDT algorithms all increased as Vmax increased,
a bid (resp. ask) different from its true valuation (resp. cost). and even approached 100% when the bid interval reached to
2) Impact on Utility: The original TAFS design does not [0, 100]. That is, when the offload service offered by the FN
prioritize profitability, but we did run simulations to verify brought a greater actual value to the user, TAFS reached the
the impact of changes in the number of participants on the desired economic characteristics with guaranteed delay and
utility of each party to observe the performance of this mech- system efficiency.
anism. Fig. 10(a)–(d) shows the effects of changes in the 4) Running Time: To confirm the time complexity analysis
number of buyers on the total utility of buyers, sellers, the of TAFS given in Section V, we tested its running time with

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 02,2024 at 11:41:02 UTC from IEEE Xplore. Restrictions apply.
 3262 IEEE INTERNET OF THINGS JOURNAL, VOL. 10, NO. 4, 15 FEBRUARY 2023

Fig. 10. (a)–(d) Utility of buyers, sellers, IoT provider, and social welfare in auctions with different assignment algorithms, where m = 10. (e) Utility of
IoT provider with RA algorithm, where m =10, 15, 20, 25.

scheme encourages the FN (as a seller) to provide an offload-

ing service with idle resources in exchange for a reward. The
UA (as a buyer) bids to obtain an offloading service. TAFS
guarantees that both buyers and sellers make truthful offers,
preventing any malicious manipulation of market behavior
with strong economic characteristics (e.g., individual ratio-
nality and budget balance). Implemented under a three-layer
cloud-fog-IoT system framework based on the FoT paradigm,
TAFS reduces operation costs while taking into account the
Fig. 11. System efficiency of TAFS over different assignment algorithms. dynamics of the FNs. Furthermore, different assignment meth-
(a) m = 10. (b) m = 10 and n = 30.
ods can be applied according to different scenarios. Simulation
experiments validated our theoretical analysis, showing that
the desired economic characteristics can be realized without
sacrificing system efficiency.

R EFERENCES
[1] Indian Health Service. “Internet of Things (IOT) Connected Devices
Installed Base Worldwide from 2015 to 2025 (in Billions).” Statista-
The Statistics Portal. Statista.com. [Online]. Available: https://
www.statista.com/statistics/471264/iot-number-of-connecteddevices-
worldwide/ (accessed Nov. 2016).
Fig. 12. Running time of TAFS. (a) m = 10 and n is varying from 20 to
100, while (b) n = 100 and m is varying from 10 to 30. [2] “Data Our New Natural Resource.” [Online]. Available:
https://www.demandcaster.com/blognews/data-our-new-resource/
(accessed Oct. 2018).
[3] L. Mearian. “Self-Driving Cars Could Create 1GB of Data a Second.”
UGB, UOB, and RA algorithms as shown in Fig. 12. The RA [Online]. Available: http://www.computerworld.com/article/2484219/
emerging-technology/self-driving-cars-could-create-1gb-of-data-a-seco
algorithm shows the lowest running time. nd.html (accessed Apr. 7, 2016).
[4] F. Bonomi, R. Milito, J. Zhu, and S. Addepalli, “Fog computing and
its role in the Internet of Things,” in Proc. 1st Edition MCC Workshop
VII. C ONCLUSION Mobile Cloud Comput., 2012, pp. 13–16.
[5] M. Aazam, S. Zeadally, and K. A. Harras, “Offloading in fog comput-
We designed an incentive mechanism, TAFS, in this study ing for IoT: Review, enabling technologies, and research opportunities,”
for application offloading in fog network environments. The Future Gener. Comput. Syst., vol. 87, pp. 278–289, Oct. 2018.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 02,2024 at 11:41:02 UTC from IEEE Xplore. Restrictions apply.
 SUN et al.: TAFS: TRUTHFUL AUCTION FOR IoT APPLICATION OFFLOADING IN FOG COMPUTING NETWORKS 3263

[6] R. Yu, G. Xue, V. T. Kilari, and X. Zhang, “The fog of things paradigm: [30] D. J. Watts and S. H. Strogatz, “Collective dynamics of ‘small-world’
Road toward on demand Internet of Things,” IEEE Commun. Mag., networks,” Nature, vol. 393, no. 6684, pp. 440–442, 1998.
vol. 56, no. 9, pp. 48–54, Sep. 2018.
[7] P. Mach and Z. Becvar, “Mobile edge computing: A survey on architec-
ture and computation offloading,” IEEE Commun. Surveys Tuts., vol. 19,
no. 3, pp. 1628–1656, 3rd Quart., 2017.
[8] H. Shah-Mansouri and V. W. S. Wong, “Hierarchical fog-cloud com-
puting for IoT systems: A computation offloading game,” IEEE Internet
Things J., vol. 5,no. 4, pp. 3246–3257, Aug. 2018.
[9] J. Du, L. Zhao, J. Feng, and X. Chu, “Computation offloading
and resource allocation in mixed fog/cloud computing systems with Lijun Sun (Member, IEEE) received the B.S.
min-max fairness guarantee,” IEEE Trans. Commun., vol. 66, no. 4, degree from Qingdao University of Science and
pp. 1594–1608, Apr. 2018. Technology, Qingdao, China, in 2000, and the
[10] Z. Chang, Z. Zhou, T. Ristaniemi, and Z. Niu, “Energy efficient M.S. degree in computer engineering and the Ph.D.
optimization for computation offloading in fog computing system,” in degree from the Institute of complexity science,
Proc. IEEE Global Commun. Conf. (GLOBECOM), Jan. 2018, pp. 1–6. Qingdao University, Qingdao, in 2003 and 2008,
[11] L. Liu, Z. Chang, X. Guo, S. Mao, and T. Ristaniemi, “Multiobjective respectively.
optimization for computation offloading in fog computing,” IEEE She was a Visiting Scholar with Arizona State
Internet Things J., vol. 5, no. 1, pp. 283–294, Feb. 2018. University, Tempe, AZ, USA. She is currently an
[12] Z. Zhou, H. Liao, X. Wang, S. Mumtaz, and J. Rodriguez, “When vehic- Assistant Professor with the College of Computer
ular fog computing meets autonomous driving: Computational resource Science and Technology, Qingdao University of
management and task offloading,” IEEE Netw., vol. 34, no. 6, pp. 70–76, Science and Technology. Her current research interests include fog/edge com-
Nov./Dec. 2020. puting, Internet of Things, crowdsourcing and network economics, and smart
[13] A. Yousefpour, G. Ishigaki, R. Gour, and J. P. Jue, “On reducing IoT traffic.
service delay via fog offloading,” IEEE Internet Things J., vol. 5, no. 2,
pp. 998–1010, Apr. 2018.
[14] Y. Xiao and M. Krunz, “QoE and power efficiency tradeoff for fog
computing networks with fog node cooperation,” in Proc. IEEE Conf.
Comput. Commun. (INFOCOM), May 2017, pp. 1–9.
[15] Z. Chang, L. Liu, X. Guo, and Q. Sheng, “Dynamic resource allocation
and computation offloading for IoT fog computing system,” IEEE Trans.
Ind. Informat., vol. 17, no. 5, pp. 3348–3357, May 2021.
[16] X. Wang, Z. Ning, and L. Wang, “Offloading in Internet of Vehicles: Guoliang Xue (Fellow, IEEE) received the Ph.D.
A fog-enabled real-time traffic management system,” IEEE Trans. Ind. degree in computer science from the University of
Informat., vol. 14, no. 10, pp. 4568–4578, Oct. 2018. Minnesota, Minneapolis, MN, USA, in 1991.
[17] W. Sun, J. Liu, Y. Yue, and H. Zhang, “Double auction-based resource He is a Professor of Computer Science and
allocation for mobile edge computing in Industrial Internet of Things,” Engineering with Arizona State University, Tempe,
IEEE Trans. Ind. Informat., vol. 14, no. 10, pp. 4692–4701, Oct. 2018. AZ, USA. His research interests span the areas
[18] Y. Liu, C. Xu, Y. Zhan, Z. Liu, J. Guan, and H Zhang, “Incentive mech- of QoS provisioning, machine learning, wireless
anism for computation offloading using edge computing: A Stackelberg networking, network security and privacy, crowd-
game approach,” Comput. Netw., vol. 129, no. 2, pp. 399–409, 2017. sourcing and network economics, Internet of Things,
[19] R. Hasana, M. Hossainaand, and R. Khan, “Aura: An incentive-driven smart city, and smart grids.
ad-hoc IoT cloud framework for proximal mobile computation offload- Prof. Xue has received the IEEE Communications
ing,” Future Gener. Comput. Syst., vol. 86, pp. 821–835, Sep. 2018. Society William R. Bennett Prize in 2019 (the Best Paper Award for
[20] F. Zhang, Z. Tang, M. Chen, X. Zhou, and W. Jia, “A dynamic resource IEEE/ACM T RANSACTIONS ON N ETWORKING and IEEE T RANSACTIONS
overbooking mechanism in fog computing,” in Proc. IEEE 15th Int. ON N ETWORK AND S ERVICE M ANAGEMENT in the previous three years).
Conf. Mobile Ad Hoc Sens. Syst. (MASS), Oct. 2018, pp. 89–97. He was a Keynote Speaker at IEEE LCN’2011 and ICNC’2014. He
[21] Y. Jiao, P. Wang, D. Niyato, and K. Suankaewmanee, “Auction was the TPC Co-Chair of IEEE Globecom’2020 and INFOCOM’2010
mechanisms in cloud/fog computing resource allocation for public and the General Co-Chair of IEEE CNS’2014. He has served on the
blockchain networks,” IEEE Trans. Parallel Distrib. Syst., vol. 30, no. 9, TPC of many conferences, including ACM CCS, ACM MOBIHOC,
pp. 1975–1989, Sep. 2019. IEEE ICNP, and IEEE INFOCOM. He is an Associate Editor of IEEE
[22] Z. Chang, W. Guo, X. Guo, Z. Zhou, and T. Ristaniemi, “Incentive T RANSACTIONS ON M OBILE C OMPUTING. He served on the Editorial Board
mechanism for edge computing-based blockchain,” IEEE Trans. Ind. of IEEE/ACM T RANSACTIONS ON N ETWORKING and an Area Editor of
Informat., vol. 16, no. 11, pp. 7105–7114, Nov. 2020. IEEE T RANSACTIONS ON W IRELESS C OMMUNICATIONS, overseeing 13
[23] Y.-Y. Shih, C.-Y. Wang, and A.-C. Pang, “Fog computing service provi- editors in the Wireless Networking area. He is the Steering Committee Chair
sion using bargaining solutions,” IEEE Trans. Services Comput., vol. 14, of IEEE INFOCOM.
no. 6, pp. 1765–1780, Nov./Dec. 2021.
[24] L. Pu, X. Chen, J. Xu, and X. Fu, “D2D fogging: An energy-efficient
and incentive-aware task offloading framework via network-assisted
D2D collaboration,” IEEE J. Sel. Areas Commun., vol. 34, no. 12,
pp. 3887–3901, Dec. 2016.
[25] H. Zhang, Y. Xiao, S. Bu, D. Niyato, F. R. Yu, and Z. Han, “Computing
resource allocation in three-tier IoT fog networks: A joint optimization
approach combining Stackelberg game and matching,” IEEE Internet
Things J., vol. 4, no. 5, pp. 1204–1215, Oct. 2017. Ruozhou Yu (Senior Member, IEEE) received the
[26] P. Samimi, Y. Teimouri, and M. Mukhtar, “A combinatorial double auc- Ph.D. degree in computer science from Arizona State
tion resource allocation model in cloud computing,” Inf. Sci., vol. 357, University, Tempe, AZ, USA, in 2019.
pp. 201–216, Aug. 2016. He is an Assistant Professor of Computer Science
[27] R. P. McAfee, “A dominant strategy double auction,” J. Econ. Theory, with North Carolina State University, Raleigh, NC,
vol. 56, no. 2, pp. 434–450, Apr. 1992. USA. His research interests include Internet of
[28] D. Yang, X. Fang, and G. Xue, “Truthful auction for cooperative commu- Things, cloud/edge computing, smart networking,
nications,” in Proc. 12th ACM Int. Symp. Mobile Ad Hoc Netw. Comput. algorithms and optimization, distributed machine
(MOBIHOC), 2011, p. 9. learning, security and privacy, and blockchain.
[29] M. Dawande, J. Kalagnanam, P. Keskinocak, R. Ravi, and F. S. Dr. Yu was a recipient of the NSF CAREER
Salman, “Approximation algorithms for the multiple knapsack problem Award in 2021. He has served on the Organizing
with assignment restrictions,” J. Combinatorial Optim., vol. 4, no. 2, Committee of IEEE INFOCOM 2022 and IEEE IPCCC 2020–2021 and as a
pp. 171–186, 2000. member of the Technical Committee of IEEE INFOCOM 2020–2022.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 02,2024 at 11:41:02 UTC from IEEE Xplore. Restrictions apply.