A Consortied Ledger-based Cloud Computing Provider Users
Reputation Architecture
Gabriel Escobar Vasques and Adriano Fiorese
a
Graduate Program in Applied Computing (PPGCAP), Department of Computer Science (DCC),
Santa Catarina State University (UDESC), Joinville, Brazil
Keywords:
Blockchain, Cloud Computing, Credibility, Reputation.
Abstract:
Several users and organizations have been attracted by the benefits offered by the cloud computing services.
However, there are still several security issues and challenges in these environments. Controlling access to
providers is an important task and must be carried out safely. In this sense, this paper presents a ledger-based
collaborative user reputation architecture for a cloud providers’ consortium. The reputation is based in two
indicators: objective and subjective ones. The objective data corresponds to the user’s session data. Subjective
data, on the other hand, corresponds to the providers’ feedback about users and data obtained from external
sources. The combination of these two indicators defines the reputation value. It aims to avoid bias in the
evaluations carried out by the providers and possible conflict of interests. In order to evaluate the proposed
architecture, a user cloud providers scenario was developed. Evaluation results show applicability of storing
and providing users reputation values to participating providers.
1 INTRODUCTION
Several users and organizations have been attracted
to Cloud Computing (CC) services in the last years.
Public health questions can result in social isolation
preventing organizations from carrying out their ac-
tivities normally. As main benefits in the adoption
of cloud services we can emphasize its ubiquitous
model of working and the on demand services deliv-
ery. Besides that, the cloud services make possible
cost reduction on hardware acquisition, installation,
and maintenance of the hardware equipment, etc. (Liu
et al., 2012). Despite the benefits offered by the CC
there are several security issues and challenges that
still need attention, offering research opportunity for
development of proposals to solve them.
The NIST (National Institute of Standards and
Technology) in its cloud reference architecture iden-
tified several issues of security (Liu et al., 2012), and
had made them public through a report. It estab-
lishes necessary recommendations so that the secu-
rity and privacy of services are guaranteed. Among
the issues raised by NIST, the identity management
and access control is an important security issue for
service providers. Malicious users represents men-
ace and are cause for great concerns in cloud envi-
a
https://orcid.org/0000-0003-1140-0002
ronments (Nkosi et al., 2013). In the literature it is
possible to find several techniques and tools focused
on authentication and control access of users. The
Public Key Infrastructure (PKI) and Reputation Sys-
tems (RS) are examples of used techniques to per-
form these tasks. Particularly, reputation systems
are widely used to estimate the credibility of service
providers, e-commerce users, among others. In such
cases, the risks assumed by providers and users in
their relationships on computational systems are pro-
portional to the reputation of the provider/user they
are dealing with. To encourage users good behaviour
in their interactions with the service providers and
help on providers access control it is possible to use
credibility indicators. Such indicators can be defined
by a reputation calculation based in the actions taken
by users in their interactions with providers. Coping
with a scenario in which CC service users utilize ser-
vices of multiple service providers in their activities,
it is possible to utilize a collaborative reputation sys-
tem shared between a consortium of providers aim-
ing the users reputation calculation based on the per-
formed actions on participants providers.
In this sense, this paper aims to present a cloud
provider service users reputation architecture that
uses as basis for the reputation value calculation,
the relations of users with consortium CC service
374
Vasques, G. and Fiorese, A.
A Consortied Ledger-based Cloud Computing Provider Users Reputation Architecture.
DOI: 10.5220/0010527403740381
In Proceedings of the 23rd International Conference on Enterprise Information Systems (ICEIS 2021) - Volume 1, pages 374-381
ISBN: 978-989-758-509-8
Copyright
c
2021 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
providers and their overall behaviour catch by ex-
ternal information sources. This architecture has
the function of providing participating providers with
users’ reputation/credibility value to help them con-
trol access to their services and contribute to the man-
agement of risks and threats that users may repre-
sent. Given the scenario characteristics we intend to
work on, which relies on the distributed collabora-
tion of consortium cloud providers, blockchain tech-
nology will be used to store the values used to cal-
culate the reputation value as well as the reputation
value itself. This technology acts as a Timestamp-
ing Authority (TSA), which is responsible for ensur-
ing the order of chronological transactions and user
behaviour data integrity. Blockchain offers benefits
such as data redundancy, transparency, consensus on
demand, which differs from traditional reputation ar-
chitectures/systems. In addition, it uses the P2P (peer-
to-peer) communication model. That is, blockchain
does not present a single point of failure that is com-
mon in reputation systems that relies on client/server
communication.
This work is organized as follows: Section 2
presents the concepts related to cloud computing, rep-
utation and blockchain technology. Section 3 exhibits
works related to this research. Section 4 presents the
users reputation architecture proposal. Section 5 dis-
cusses the scenario and experiments preformed, as
well as results obtained. Lastly, Section 6 present the
final consideration and future work.
2 BACKGROUND
This section presents characteristics and descriptions
of the related technologies to the content covered by
this work.
2.1 Cloud Computing
Cloud computing is a paradigm for more efficient use
of computational infrastructure, being defined as a
technologically ubiquitous model. Cloud computing
service models are defined according to how the con-
figurable set of computational resources can be deliv-
ered to end users. In this work, we adopted the NIST
nomenclature, which is used by several researchers
and organizations.
According to the NIST reference model, an im-
portant feature in computational clouds regardless of
their classification and deployment model is security.
NIST has developed a security guide, which provides
an overview of the security and privacy challenges in-
volved in CC (Jansen and Grance, 2011). Among the
security issues pointed out by NIST, access control
is an important task and requires care by providers.
This work presents a cloud providers’ users reputa-
tion architecture that aims to assist the addressment of
issues raised by the security guide prepared by NIST.
Among these issues, the proposed architecture aims to
assist service providers at managing identity and ac-
cess. In addition, the architecture reputation value can
indirectly assist the management of cloud providers
internal risks and threats.
2.2 Reputation
Reputation is a well-known concept and applied in
several areas. A widely used definition of reputation
to describe its use in computer systems was devel-
oped by (Mui et al., 2002). There, the authors define
reputation as: “the perception that an agent creates
through past actions on his or her huge intentions”.
Typically, reputation values are defined according to
the feedback provided by users and/or comprising the
results of the transactions. However, applications that
use only feedback in their calculations, which are sub-
jective in nature, can present loopholes for malicious
users and common reputation system attacks such as
self-promotion or defamation (collusion) (Hendrikx
et al., 2014). So, in fact, the final reputation values
are not affected only by feedback to prevent malicious
behavior, it is possible to use other information to-
gether with them to compose the calculations. In this
sense, it is possible to use objective information about
users to compose the reputation calculation, such as:
behavior classification, use of resources, etc. In addi-
tion, information related to the payment of monthly
fees/annuities regarding the used computational re-
sources, can also be part of the final reputation value.
This way, the final reputation values will be less sus-
ceptible to malicious feedback provided by dishonest
users.
2.3 Blockchain
Blockchain is a disruptive technology for conducting
transactions between two entities without the need for
a responsible third party to establish trust (Nakamoto,
2009). The blockchain concept consists of six main
characteristics: decentralization, transparency, open
source, autonomy, immutability and anonymity (Lin
and Liao, 2017). In general, applications developed
with blockchain can be implemented following three
main models: public, private and consortium (Lin and
Liao, 2017). In public blockchains, the entry of new
participants in the network happens in a simple way.
Users can choose to participate in the consensus pro-
A Consortied Ledger-based Cloud Computing Provider Users Reputation Architecture
375
cess and consequently to the addition of new blocks
where all participants in the blockchain can check the
transactions. A blockchain network is a technical in-
frastructure that provides ledger and smart contracts
services to applications.
In implementations that follow the private model,
the control of the blockchain network is performed
by a single organization which is responsible for al-
lowing or not allowing the entry of new participants,
in addition to defining the actions that each partici-
pant can perform. The consortium model works in
a similar way to the private one. However, two or
more organizations have control of the network. In
these models, the mechanisms used to obtain consen-
sus and to add new blocks tend to be less computa-
tionally costly, and consequently consumes less en-
ergy. Furthermore, in these models, participants are
usually well identified and only participants defined
by the organization(s) can verify transactions and add
new blocks (Greve et al., 2018).
Blockchain technology is used in the development
of the reputation architecture proposed in this work.
In turn, the architecture suits a scenario where CC ser-
vice providers and their users integrate a blockchain
network at the consortium model.
3 RELATED WORK
Some works such as (Xu et al., 2019) are aimed at
assisting the selection of personalized CC services,
based on Quality of Service (QoS) assessments car-
ried out by users. To encourage users to provide re-
liable QoS assessments, the authors propose a mech-
anism for calculating users’ reputation and reliability
based on the assessments provided by them and com-
paring it with other assessments on the same service
provided by other users. In (Zheng et al., 2018) the
authors present an approach to improve the speed of
classifying users’ normal behavior in a cloud environ-
ment. For this, those authors present a reputation clas-
sifier, which is responsible for calculating the reputa-
tion values based on the user behavior and service use
classifications obtained by means of machine learn-
ing.
The work of (Thakur and Breslin, 2019) proposes
a mechanism for managing the reputation of IaaS-
type service providers in a federated/hybrid cloud.
The proposal aims to encourage providers to clas-
sify their users’ behavior correctly, differentiating
them between good and malicious users. The reputa-
tion of providers is estimated through feedback from
providers about other providers, feedback from users
about providers and feedback from providers about
users.
The work developed by (Wu et al., 2015) in-
troduces a reputation mechanism and designs a
reputation-based identity management model for CC.
The authors developed a model, in which users re-
ceive pseudonyms generated from a reputation sig-
nature in order to guarantee the non-traceability of
pseudonyms and proposes a reputation calculation
mechanism to help identify malicious users. The
work of (Du et al., 2019) proposes an online repu-
tation calculation method to efficiently provide a per-
sonalized reputation for each user, within the context
of cloud applications based on service-oriented archi-
tectures. The work (Donghong et al., 2016) presents
a dynamic access control framework based on repu-
tation and attributes for privacy protection in the CC
environment.
These works analysis shows that there is a gap for
works that present approaches related to the reputa-
tion of users of cloud providers. Moreover, it is also
noticeable a lack of works where providers’ users rep-
utation is shared with other providers, allowing them
to assess the risks that eventually malicious users can
to offer.
4 PROPOSAL
This paper models and proposes a reputation archi-
tecture for users of cloud providers, as a collaborative
tool between them.
4.1 Reputation Architecture
The proposed architecture is presented in Figure 1. It
is composed of several modules that perform the func-
tions related to the calculation of the reputation value.
For the implementation, blockchain technology was
used as a mechanism for data persistence, taking ad-
vantage of its benefits as immutability to protect data
against attacks on the reputation system, as well as
ensuring the traceability and auditing of any conflicts.
The reputation architecture has three main compo-
nents: the reputation system (RS), the cloud provider
and external data sources. The RS consists of the rep-
utation calculation module and the blockchain itself.
The data is validated and stored in a distributed and
redundant way using a blockchain in the consortium
model through distributed consensus. The blockchain
has the important role of acting as a distributed TSA,
being responsible for ensuring the temporal ordering
of the records inserted in the chain of blocks.
The proposed architecture usage and behaviour is
given by the following interactions between its mod-
ICEIS 2021 - 23rd International Conference on Enterprise Information Systems
376
ules. Upon receiving the request, the access control
module of the cloud provider can perform a query to
the reputation calculation module to obtain the user
reputation value. Upon receiving the user’s reputa-
tion value, the provider may or may not allow the
user access to its services, applying access and secu-
rity policies according that value. If access is allowed,
user uses the provider’s services and at the end of his
session provider makes available the data from that
session to the reputation system so that they can be
used in the proper credibility indicator calculations.
After receiving the session data from the user, the RS
performs consultations with external sources to obtain
information from public databases that are also used
in the calculations. After receiving all the previously
mentioned data, they are saved in the ledger and cal-
culations of the objective and subjective credibility in-
dicators are performed. The values that compose the
objective credibility indicator come from resources
utilization offered by the providers to the users like
RAM memory, processors, the user connection sta-
bility and payments. Thus, the higher these values,
which indicates an active, stable and well payer user,
the higher the objective credibility indicator and con-
sequently the reputation value. The contrary is also
directly related. The values that make up the sub-
jective credibility indicator come from the external
sources and user evaluation given by the providers
feedback. The final reputation value is calculated by
combining these two indicators.
The sources of external information are used in
the reputation calculation together with the providers’
feedback to compose the subjective credibility indi-
cator. These kind of external sources should provide
info that can be used to marginally infer a particular
portion of the user social behavior. Queries to these
external (to the architecture) databases return the re-
spective number of issue items that user has. Origi-
nally, as proposed by default, it should be four exter-
nal data sources and each one of them represents 25%
of the external issues indicator.
However, providers can assign agreed weights to
external data sources at their own discretion.
Figure 1: Reputation Architecture.
Thus, the reputation (R) of users is given by the
combination of two historical indicators of credibil-
ity: the objective and subjective ones, as observed in
Equation 1.
According to the definition adopted for reputation,
I
obj
is the historical and current objective credibility
indicator (if any) for user s, I
sub
is the historical sub-
jective credibility indicator and ω
obj
, ω
sub
are weights
defined for the respective indicators.
Equation 2 represents the user’s objective cred-
ibility indicator, which is composed of the average
percentage of computational resources use (upr) in n
sessions, the connection stability (EoC) calculated by
the n sessions latency (ping) value average and the ra-
tio calculated between the number of payments made
on time and the total number of monthly/annuity pay-
ments (pay), for each user. These indicators corre-
spond respectively to 35%, 25% and 40% of the ob-
jective credibility indicator value.
The upr variable corresponds to the percentage of
resource use, which is calculated by means of the
arithmetic average between the percentage averages
of CPU usage (mpcp) and corresponding RAM mem-
ory usage (mmr) in the last n user sessions. Equa-
tion 6 is responsible for calculating the percentage
value of resource usage according to the user’s history
(upr). The percentage of use of each CPU (process-
ing core) (pcp(s)) is captured during the user’s session
and through it the average percentage of CPU usage
of that session is calculated. Thus, the mpcp vari-
able, which corresponds to the average percentage of
CPU usage, is obtained by adding the percentages of
use of each CPU (pcp) used, from the first core (i =
1) to the last one (npc), and dividing that sum by the
total number of cores (npc). Equation 3 is responsi-
ble for calculating the user’s CPU usage percentage
(mpcp(s)).
The mmr variable corresponds to the average
RAM memory usage in MB, which is calculated as
the ratio between the sum of the memory usage value
every second during the user session, and the total du-
ration of the session tf, in seconds. Equation 4 is re-
sponsible for calculating the average RAM memory
usage for that session. The pur variable corresponds
to the percentage of RAM memory usage, calculated
by the ratio between the average RAM memory used
(mmr - Equation 4) and the amount of RAM memory
allocated (amr). The amount of allocated RAM (amr)
is the amount originally purchased by the user. Equa-
tion 5 is responsible for determining the percentage of
RAM memory usage.
In turn, variable (EoC) present in Equation 2 cor-
responds to the historical average of the user’s con-
nection latency. To obtain its value, the user’s connec-
A Consortied Ledger-based Cloud Computing Provider Users Reputation Architecture
377
tion latency value should be initially obtained from
each user’s sessions. In this case, the ping appli-
cation is used for polling, by means of the Internet
Control Message Protocol (ICMP), between the cloud
provider and a host on the user’s network, making it
possible to obtain the latency value in the communica-
tion between provider and user. Thus, every ping op-
eration during the user’s session at the provider adds
value 1 for the latency accumulator, if its value is less
than or equal to 150 ms. At the end of the session, this
accumulated value is divided by the number of times
that the latency value was recovered (n), according to
Equation 7. This is the user connection latency when
he is using a cloud provider. The user s historical av-
erage latency (EoC(s)), used in the composition of the
objective indicator, according to Equation 2, is cal-
culated using the p last (therefore historical) values
of the user’s connection latency (EoC), as shown in
Equation 8.
In turn, variable (pay), also Equation 2, corre-
sponds to the user’s punctuality in paying the monthly
fee for the cloud provider’s services. Variable (pay)
corresponds to the ratio between the number of pay-
ments up to on time (npd) and the total of payments
made (ntp), as described by Equation 9.
On the other hand, Equation 10 is responsible for
calculating the subjective credibility indicator value.
This value results from the difference between the
feedback from the providers and the results of con-
sultations with external sources.
The second component of the subjective credibil-
ity indicator comprises the results from consultation
with external sources, and it is called external issues
indicator (I
Pex
(s)). Equation 11 is responsible for cal-
culating that indicator. For the calculation of the ex-
ternal issues indicator, normalization is carried out in-
volving the sum of the values obtained from external
sources (Fe
j
) weighted by the cloud providers (default
weight is 0.25 each external source), and the sum of
weighted maximum values from each external source,
according each user s. Equation 12, on the other hand,
is responsible for calculating the average of n values
of the external issues indicator for each user, making
the final external issues indicator a historical indica-
tor. The n value, representing the historical period
(how many external issue indicator values are used),
should be agreed among the cloud providers belong-
ing to the consortium.
Cloud providers send feedback values, regarding
their users, to allow the user behaviour assessment
by the reputation architecture. These feedback val-
ues comprise different activities users perform at the
cloud provider. Each of these activities is represented
by an indicator, for instance (I
1
, I
2
, ..., I
k
), where k is
the total number of activities/indicators. This work
proposes three so called feedback indicators. First
indicator (I
1
) corresponds to the provider’s feedback
due to the user’s login attempts. Second indicator
(I
2
) corresponds to the provider’s feedback due to
the login path, in which the IP address and the user
agent are verified. According to (Yu et al., 0108)
and (Berrached and Korvin, 2006) these two indica-
tors represent factors that influence the user’s login
behavior and credibility. Third indicator (I
3
), on the
other hand, corresponds to the provider’s feedback
due to the attempts of users to access improper data.
Providers assign values from 0 to 10 for each of these
indicators according the user attempts to execute the
corresponding activities. All providers must be in line
with the feedback evaluation model in order to avoid
indicator value bias. Thus, a user feedback provided
by all nfp providers, in a time period, corresponds
to the ratio between feedback indicators’ sum from
all providers [1..nfp], and the sum of maximum val-
ues for each of these indicators, from all providers
[1..nfp]. Equation 13 shows user feedback (Fp(s))
calculation.
In turn, Equation 14 transforms the single period
of time user feedback indicator (Fp(s)) into a histor-
ical (average) user feedback (Fp(s) ). To accomplish
that, it performs the ratio between the sum of np single
user feedback indicators and the number of already
(past) calculated single user feedback indicators for
that historical period (np). This np value also must be
agreed among providers’ consortium participants. It
is important to note that the historical user feedback is
one of the subjective credibility indicator components
(Equation 10).
R
s
= ω
ob j
I
ob j
(s) + ω
sub
I
sub
(s). (1)
I
ob j
(s) = (upr 0.35 +EoC 0.25 + pay 0, 40). (2)
mpcp(s) =
npc
i=1
pcp(s)
i
npc
(3)
mmr(s) =
t f
i=1
umr(s)
i
t f
(4)
pur(s) =
mmr(s)
amr(s)
(5)
upr(s) =
n
i=1
pur(s)
n
+
n
i=1
mpcp(s)
n
2
(6)
EoC(s) =
n
i=1
1[ping
i
150ms]
n
(7)
EoC(s) =
p
i=1
EoC(s)
i
p
(8)
ICEIS 2021 - 23rd International Conference on Enterprise Information Systems
378
pay(s) =
npd
nt p
(9)
Isub(s) = F p I
Pex
(10)
I
Pex(s)
=
N f e
i=1
Fe
i
ωi
N f e
j=1
max(Fe
j
ω j)
(11)
I
Pex(s)
=
n f
i=1
I
Pex(s)
i
n f
(12)
F p(s) =
n f p
i=1
I
1,i
+ I
2,i
+ I
3,i
n f p
j=1
max
j
(I
1
) + max
j
(I
2
) + max
j
(I
3
)
(13)
F p(s) =
np
i=1
F p
i
np
(14)
5 SCENARIO & EXPERIMENTS
The performed experiments are based on a sce-
nario, in which 15 users use the services of 3 dif-
ferent cloud providers. These cloud providers com-
pose a blockchain network coping with the consor-
tium model, and they use the proposed reputation
architecture. The experiments were carried out us-
ing Hyperledger Fabric version 2.2 framework de-
veloped by the Linux Foundation. To create and
configure the blockchain network, Docker contain-
ers and the container orchestrator Docker-Compose
were used. The blockchain network reaches consen-
sus by means of the Byzantine Fault Tolerance (BFT)
mechanism, which allows the consensus process to
be carried out even if some nodes fail or act ma-
liciously. To develop chaincodes (smart contracts),
JavaScript programming language was used. Node.js
framework runs JavaScript code. To manage pack-
ages for JavaScript language, the Node Package Man-
ager was used, which is already included in Node.js
framework. In addition, cURL (Client URL) software
that provides a library and a command line tool for
transferring data, providing support for various proto-
cols, was used.
Experiments were carried out on a computer with
an Intel Core-i5-7200U 2.50 GHz processor, 8 GB of
RAM memory with Linux Mint 19.2 operating sys-
tem. A simulation of a hypothetical case was con-
ducted, in which each user performs 10 sessions at
one of the participating providers. In addition, ve
different users are assigned to each provider. The val-
ues used to compute the objective and subjective indi-
cators like memory use, processors use, user latency,
external issues values, and providers feedback were
generated randomly, following a normal probability
distribution function using the Math.random floating
point random number generation function available
in the JavaScript programming language. All values
were generated 10 times to enable averages to be cal-
culated. Values between 35 and 95 were used for the
CPU and RAM memory percentage indicators.
Average of 10 values between 20 and 200ms is
used for the connection stability indicator (user la-
tency). Values between 0 to 10 were generated for
each feedback indicator from the providers. These
values were generated 10 times and added, then these
values were normalized to obtain values between 0
and 1, according Equation 13. In this case, nfp equals
to 10. Comprising data from each external sources,
values between 0 and 10 were generated taken into
account one external data request per session and a
historical period of ten sessions (n in Equation 11 is
10). The payment amounts correspond to the pay-
ments on time divided by the total number of total
payments (historical period) resulting in a value be-
tween 0 and 1. The values were generated for a pe-
riod of 10 months, where the value 1 corresponds to
the payment on time and 0 for late payment. When
the payment was made the value 1 is added, otherwise
the value 0 is added. The values of the indicators that
make up the objective and subjective indicators are
obtained through the summation and subsequent cal-
culation of averages of data of 10 last sessions. To
calculate the averages of the indicators, values corre-
sponding to 10 session of each user were generated.
These values were used to calculate the historical av-
erage values of the I
pex
, F p and EoC indicators.
5.1 Results
Tables 1, 2, 3 show results from simulation. In all
tables, user identifiers (1 to 15) and provider iden-
tifiers (1 to 3) are present in the first two columns.
Table 1 presents the information about the CPU use
percentage averages, RAM memory use average, re-
source usage percentage average, ratio between the
number of payments on time and total number of
payments, and connection stability indicator histori-
cal value, i.e., user session latencies average. Table 2
presents average (historical) user feedback, and the
average values returned by the external data sources
(F1, F2, F3 and F4) presented in Section 4.1. These
values represent the historical average obtained by the
summation of the 10 simulations data to each external
data source.
Table 3 presents Objective and Subjective credi-
bility indicator values, according data from previous
A Consortied Ledger-based Cloud Computing Provider Users Reputation Architecture
379
Table 1: User session data provided by cloud providers.
User ID Provider ID CPU aver-
age
RAM aver-
age
Resources use average Payment Connection stability
U1 P1 66% 64% 65% 1 0,92
U2 P3 74% 70% 72% 0,83 0,87
U3 P3 71% 66% 68,2% 0,75 0,64
U4 P1 60% 53% 56,5% 1 0,98
U5 P2 79% 71% 75% 0,9 0,86
U6 P2 58% 50% 54% 0,6 0,90
U7 P3 64% 58% 61% 0,87 0,69
U8 P1 70% 64% 67% 1 0,79
U9 P2 78% 73% 75,5% 0,9 0,93
U10 P1 59% 50% 54,5% 0,83 0,99
U11 P1 62% 57% 59,5% 0,8 0,99
U12 P3 40% 30% 35% 0,5 0,99
U13 P2 53% 44% 48,5% 0,85 0,97
U14 P3 80% 78% 79% 0,88 0,87
U15 P2 73% 70% 71,5% 1 0,923
Table 2: Data obtained from external sources.
User ID Provider ID Feedback F1 F2 F3 F4
U1 P1 0,7 0 0 0 0
U2 P3 0,8 1 3 2 5
U3 P3 0,7 2 0 3 0
U4 P1 0,6 1 1 1 1
U5 P2 0,6 0 2 0 1
U6 P2 0,3 0 3 2 1
U7 P3 0,7 2 1 2 3
U8 P1 0,4 1 1 2 0
U9 P2 0,8 3 2 3 3
U10 P1 0,7 1 2 2 5
U11 P1 0,8 6 1 1 3
U12 P3 0,2 0 1 2 2
U13 P2 0,3 2 9 2 3
U14 P3 0,9 3 2 0 0
U15 P2 0,7 5 3 4 0
Table 3: Indicator results and reputation.
User ID provider ID Objetive Subjective Reputation
U1 P1 0,85 0,7 0,810
U2 P3 0,8 0,7 0,775
U3 P3 0,7 0,71 0,706
U4 P1 0,84 0,84 0,843
U5 P2 0,83 0,57 0,758
U6 P2 0,65 0,525 0,615
U7 P3 0,73 0,62 0,702
U8 P1 0,83 0,44 0,715
U9 P2 0,85 0,72 0,819
U10 P1 0,77 0,61 0,726
U11 P1 0,77 0,7 0,755
U12 P3 0,57 0,74 0,624
U13 P2 0,76 0,50 0,683
U14 P3 0,87 0,82 0,860
U15 P2 0,77 0,6 0,728
tables. In addition, Table 3 also shows the reputation
values for each user.
Summing up, Figure 2 shows the users reputation
values. The x axis of Figure 2 shows distribution of
users at the providers, through the users’ identifiers
(U1 to U15) associated with the providers’ identi-
fiers (P1 to P3). The y axis presents reputation values
calculated for each user, ranging from 0 to 1, where
value 1 represents the best reputation. By observing
values presented, it is possible to notice that user 14
(U14) who used the services of provider 3, obtained
the highest reputation value among all users, whose
value is close to 85%.
This value is related to the historic average use
of computational resources, since U14 presents the
highest resources use average (Equation 6 when com-
pared to the other users, according Table 1. More-
over, it performs most of the monthly/annuity pay-
ments on time and presents good connection stability,
as it is possible to see in Table 1. In addition, user
14 obtained high average comprising providers’ feed-
back values, despite having a total of 5 pending is-
sues at external sources, as shown in Table 2. Thus,
for this user, objective credibility indicator is 0.87 and
subjective credibility indicator is 0.82, demonstrating
balancing between objective and subjective behavior.
Values shown by Figure 2 for user 12 (U12), who also
used services of provider 3, obtained the lowest repu-
tation value among all users. User 12 reputation value
is close to 62%.
Figure 2: Reputation results.
This value is related to the use of computational
resources, which obtained the lowest average when
compared to the other users, with the accomplishment
of half of the monthly/annuity payments on time, but
with a high value for the connection stability indica-
tor, as can be seen in Table 1. In addition, user 12
obtained a low average of the provider’s evaluation
feedback values, and has a total of 5 pending issues
with external sources, as shown in Table 2. Thus, for
this user, the objective credibility indicators obtained
a value of 0.57 and the subjective credibility indicator
obtained a value of 0.74 Thus, the reputation value
R(s) for this user is 0.444. Weight of each of these
ICEIS 2021 - 23rd International Conference on Enterprise Information Systems
380
indicators was defined as 0.7 and 0.3 respectively for
all participating providers.
Results show that it is possible to evaluate the
users reputation based on the use of computational re-
sources.
6 FINAL CONSIDERATIONS
This work proposed a reputation architecture for
cloud providers’ users reputation architecture based
on blockchain consortium model. The proposed ar-
chitecture aims to help cloud providers to prevent re-
source access to malicious users i.e., those whose rep-
utation value is low enough to be considered mali-
cious according to providers security policy defini-
tions. However, it is not restricted to this context and
it can be adapted and used by other computer sys-
tems. Users have their reputations calculated by com-
bining objective and subjective credibility indicators.
The objective credibility indicator is calculated with
the user resources utilization like memory, process-
ing, connection stability and payment. The subjec-
tive credibility indicator is calculated with the data
obtained in external data sources and the providers
feedback regarding user’s activities performed on the
provider. These two indicators make up the reputa-
tion value assigned to each user. Thus, reputation
value can be used by participating providers to as-
sist in their access control decisions. Results were
obtained through simulations performed with hypo-
thetical scenarios. In this way, participating providers
have access to user’s reputation values based on in-
teractions with themselves and with other participat-
ing providers through a shared and collaboratively
maintained reputation architecture. As future work,
tests are being developed with a greater number of
providers and users to assess architecture scalability
and performance.
REFERENCES
Berrached, A. and Korvin, A. (2006). Reinforcing access
control using fuzzy relation equations. In Security and
Management.
Donghong, S., Wu, L., Ping, R., and Ke, L. (2016). Reputa-
tion and attribute based dynamic access control frame-
work in cloud computing environment for privacy pro-
tection. In 2016 12th International Conference on
Natural Computation, Fuzzy Systems and Knowledge
Discovery (ICNC-FSKD), pages 1239–1245.
Du, X., Xu, J., Cai, W., Zhu, C., and Chen, Y. (2019). Oprc:
An online personalized reputation calculation model
in service-oriented computing environments. IEEE
Access, 7:87760–87768.
Greve, F., Sampaio, L., Abijaude, J., Coutinho, A. A.,
Valcy, I., and Queiroz, S. (2018). Blockchain e a
revoluc¸
˜
ao do consenso sob demanda. In Simp
´
osio
Brasileiro de Redes de Computadores e Sistemas Dis-
tribu
´
ıdos (SBRC) - Minicursos, chapter 5, page 30. So-
ciedade Brasileira de Computac¸
˜
ao - SBC.
Hendrikx, F., Bubendorfer, K., and Chard, R. (2014). Rep-
utation systems: A survey and taxonomy. Journal of
Parallel and Distributed Computing, 75.
Jansen, W. and Grance, T. (2011). Sp 800-144. guide-
lines on security and privacy in public cloud comput-
ing. Technical report, U.S Departament of Commerce,
Gaithersburg, MD, USA.
Lin, I.-C. and Liao, T.-C. (2017). A survey of blockchain
security issues and challenges. I. J. Network Security,
19:653–659.
Liu, F., Tong, J., Mao, J., Bohn, R., Messina, J., Badger, L.,
and Leaf, D. (2012). NIST Cloud Computing Refer-
ence Architecture: Recommendations of the National
Institute of Standards and Technology (Special Publi-
cation 500-292). CreateSpace Independent Publishing
Platform, USA.
Mui, L., Mohtashemi, M., and Halberstadt, A. (2002).
A computational model of trust and reputation. In
Proceedings of the 35th Annual Hawaii International
Conference on System Sciences, pages 2431–2439,
Big Island, HI, USA. IEEE.
Nakamoto, S. (2009). Bitcoin: A peer-to-peer electronic
cash system,” http://bitcoin.org/bitcoin.pdf.
Nkosi, L., Tarwireyi, P., and Adigun, M. O. (2013). Detect-
ing a malicious insider in the cloud environment using
sequential rule mining. In 2013 International Confer-
ence on Adaptive Science and Technology, pages 1–
10, Pretoria, South Africa. IEEE.
Thakur, S. and Breslin, J. G. (2019). A robust reputation
management mechanism in the federated cloud. IEEE
Transactions on Cloud Computing, 7(3):625–637.
Wu, L., Zhou, S., Zhou, Z., Hong, Z., and Huang, K. (2015).
A reputation-based identity management model for
cloud computing. Mathematical Problems in Engi-
neering, 2015:1–15.
Xu, J., Du, X., Cai, W., Zhu, C., and Chen, Y. (2019).
Meurep: A novel user reputation calculation approach
in personalized cloud services. PLOS ONE, 14(6):1–
15.
Yu, X., Gao, N., and Jang, W. (2010(8)). Research on the
identity authentication technology in cloud comput-
ing. In Information Network Security, pages 71–74.
Zheng, R., Chen, J., Zhang, M., Wu, Q., Zhu, J., and
Wang, H. (2018). A collaborative analysis method
of user abnormal behavior based on reputation voting
in cloud environment. Future Gener. Comput. Syst.,
83(C):60–74.
A Consortied Ledger-based Cloud Computing Provider Users Reputation Architecture
381