Policy-based Security Provisioning and Performance Control in the
Cloud
Viswanath Nandina, Jos
´
e Marcio Luna, Edward J. Nava, Christopher C. Lamb, Gregory L. Heileman
and Chaouki T. Abdallah
Department of Electrical and Computer Engineering, University of New Mexico, Albuquerque, U.S.A.
Keywords:
Cloud Computing, Security, Control Theory, Access Control, Usage Management.
Abstract:
In this paper we describe the development of a system that provides security provisioning and performance
controls over content in a cloud environment. Using an approach grounded in Usage Management and Control
theory we are able to successfully provision resources in multiple cloud systems. Providing increased security
comes with a cost of reduced performance. Therefore a variable performance control model for different levels
of security is proposed. This control model allocates Virtual Machines adaptively so that a desired performance
measure lies between predefined upper and lower bounds as agreed in the Service Level Agreement.
1 INTRODUCTION
The bane of the existence of cloud computing has
been the inability to ensure security while maintain-
ing performance. Despite the rising popularity sur-
rounding cloud computing, disengaging users from
hardware needs has long been criticized for being un-
able to provide a trustworthy solution to the problem
of managing the trade-off between security and per-
formance (Zissis and Lekkas, 2012). Organizations
and users have been hesitant to adopt cloud comput-
ing because of current cloud systems’ inability to pro-
vide varying levels of security to various types of data.
Currently, many organizations have requirements to
assure that data cannot be altered by intentional at-
tacks. Thus some organizations invest in their own
computing enterprise systems with rigorous security
and reliability protections. There exists a global de-
mand for an automatic control system to enforce se-
curity rules so that operations can be directed to pri-
vate or public cloud systems. Cloud computing con-
tinues to be plagued by infrastructural failures and po-
tential security problems (Takabi et al., 2010). Over
the past few years major infrastructure providers rang-
ing from Amazon to Microsoft have suffered from ex-
tensive and unexpected system downtime. This has
led to the development of tools like Amazon’s Cloud-
Watch, which can monitor provisioned infrastructure
for failures and help secure additional infrastructure
when required.
We are pursuing a novel approach for applying
Usage Management (UM) (Jamkhedkar et al., 2010)
to resources (Virtual Machines and data sets) in a
cloud computing environment (Jamkhedkar et al.,
2011). We demonstrate how UM can be applied to
cloud computing by controlling the resource based on
a particular policy set. Likewise others in the area
(Takabi and Joshi, 2012) have their own policy man-
agement framework. We are incorporating multi-level
security on our UM policies. The overhead of the
application service provided by the cloud increases
as the security requirements become more stringent
(e.g., military applications). Therefore, we propose a
control theoretic approach to regulate application per-
formance and guarantee that the negotiated SLAs and
the security requirements are satisfied.
Though cloud-dependent systems are becoming
more stable as a result of these and other efforts,
they still have yet to integrate secure provisioning into
these kinds of bursting schemes. We present a mecha-
nism that provides robust system and information us-
age management in tandem with performance moni-
toring and alerting. This mechanism can securely pro-
vision systems on demand based on a unified perfor-
mance and sensitivity profile. We describe notionally
how we expect this system to work, the primary con-
cepts it embodies, and describe our experiences im-
plementing the system using Eucalyptus, Amazon’s
Elastic Compute Cloud, and Amazon’s Simple Stor-
age Service.
In a cloud computing environment an application
may have different performance measures, such as,
502
Nandina V., Luna J., Nava E., Lamb C., Heileman G. and Abdallah C..
Policy-based Security Provisioning and Performance Control in the Cloud.
DOI: 10.5220/0004378705020508
In Proceedings of the 3rd International Conference on Cloud Computing and Services Science (CLOSER-2013), pages 502-508
ISBN: 978-989-8565-52-5
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
response time, throughput or latency. In previous
works such as (Nathuji et al., 2010), an approach as-
suming that the cloud can be modeled as a Multi-
Input-Multi-Output (MIMO) system is implemented
to regulate performance in the cloud. From this per-
spective the cloud is seen as a black box with in-
puts and outputs. The outputs are the different per-
formance measures of interest, while the inputs are
values related to the amount of available virtual re-
sources in the cloud, e.g., CPU utilization and avail-
able memory in a Virtual Machine (VM), as well as
the amount of available VMs in a cluster. The au-
thors assume that the hypervisor allows for a fine-
grained control of the parameters of the Virtual Re-
sources in place. However, based on (Vliet and Pa-
ganelli, 2011; Lim et al., 2009), in a public cloud ser-
vice, specifically in Infrastructure-as-a-Service (IaaS)
such as EC2 from Amazon, the variation of the inter-
nal parameters of the VMs and therefore the response
of the system are coarse-grained. This imposes sev-
eral difficulties to apply methodologies such as Linear
Quadratic Regulation as in the CPU problem in (Yao
et al., 2010) which is effective to regulate MIMO sys-
tems.
In this work, we follow the trend proposed by
(Lim et al., 2009) called proportional threshold-
ing which is suitable for Single-Input-Single-Output
(SISO) systems and overcomes the coarse-granularity
of this problem. The main goal is to confine the de-
sired performance to an interval of values rather than
regulate it to a specific reference value. However,
in order to improve the response of the controller,
we have added on-line model identification (Kulhav
´
y,
1987; Haykin, 2002; Hellerstein et al., 2004), which
allows the update of the model parameters to over-
come the typical workload variations of the cloud.
We implement a policy based usage management
system in multiple cloud environments. This system
treats resources with specific attributes that may be
provisioned according to usage policies. We regulate
the performance of running virtual machines to meet
the service level agreements applying control theory.
This approach is conceptually similar to multi-level
security system implemented in a cloud computing
environment. We will consider an operational en-
vironment that includes a mix of public and private
cloud computing resources, each of which can be cat-
egorized as having a certain level of trust.
The rest of the paper is as follows: Section 2, de-
scribes how usage management goes beyond access
control. Section 3 explains the Usage Management
system for Cloud computing, as well as the system
configuration and the technologies used to build the
system. Section 4 deals with performance control and
finally Section 5 provides concluding remarks.
2 USAGE MANAGEMENT
A conceptual UM system is described in Figure 1
for a cloud computing environment. The UM sys-
tem determines if a user can be granted access to a
resource, such as a data file. In a cloud computing
environment, once a user is granted access to a re-
source, there is a need to control not only how the
resource can be used, but also where. Our primary
goal is to develop and implement a secure, robust,
and inter-operable attribute-based usage management
system for data transactions in cloud computing. This
system will merge usage management systems with
modern cloud computing technologies. Given a data
resource with an associated sensitivity characteristic,
one part of the proposed cloud UM process will be to
determine on which type of cloud computing system
the iresource will be made available.
2.1 Beyond Access Control
Usage management (UM) is defined as the manage-
ment of the usage of resources (and data) across and
within computing environments.(Jamkhedkar et al.,
2010) Usage management incorporates characteris-
tics of traditional access control and digital rights
management. In order to be effective, it must be flex-
ible enough to be scalable and interoperable enough
to provide services across different computational en-
vironments. In an implementation of UM, the first
action provided by the system is access control. A
user logs into a system with credentials and with con-
text that is provided by the user, and/or determined
automatically. The next role of the UM is to ensure
that data is used in an environment that complies with
security policies that are either specified in an associ-
ated license, or are applicable to a security category
for unclassified data.
For instance, a user might specify at which sen-
sitivity level, or classification level he or she wishes
to operate and UM checks the contextual information
before providing access to any resources. Based on
the examination, UM commands a cloud computing
system to instantiate a VM and load an image that
contains the necessary security mechanisms. Then
it transfers the actual data to the VM. In this paper,
we use VM images as the controlled resources, where
each image has a set of policies associated with it that
describe the circumstances under which that image
can be used.
Policy-basedSecurityProvisioningandPerformanceControlintheCloud
503
Likewise, each user has an identity that describes
their credentials and environment information, also
known as context. Usage management occurs when
a user’s identity information is compared with a re-
sources’ policies to determine if that user is qualified
to use that particular resource. Once he gets access
to the resource the UM system monitors how the re-
source is being used based upon the policy require-
ments.
Global Usage
Manager
Mechanism
Trust Module
Usage Management Framework
UM
Virtual Machine
Applications &
Data
UM
Virtual Machine
Applications &
Data
Cloud Computer Storage
or Content Network
Data & Licenses
User and/or
Applications
Figure 1: Hierarchical UM operation concept.
3 USAGE MANAGEMENT
SYSTEM FOR CLOUD
COMPUTING
In order to provide an assured information sharing ca-
pability in a cloud computing environment, a UM sys-
tem capable of interacting with, and controlling, dif-
ferent provider systems is required. The high-level
conceptual view is shown in Figure 2. Each system
will likely be slightly different, so a common frame-
work, or ontology, is required in order to use common
policy semantics so that the policies can be applied
consistently in each system. A policy generator will
be required in order to generate the licenses that will
be needed for each data set, and the policies will use
the common ontology as a basis.
Based on the policies that are specified in a license
Data
Policy
Data
Policy
Cloud
Provider 1
Cloud
Provider
2
Data
Policy
Usage Management Framework
Common Cloud Ontology
conform conform
aggregation
Policy
Generator
Figure 2: Usage Management in a Multi-Cloud Environ-
ment.
for a particular resource, the UM first grants access
to the resource. Then, based on the policies, the UM
instantiates a VM in an appropriate cloud system and
copies the data to the VM for processing. Once a data
set has been transferred to a VM, without additional
controls, the user and associated applications can use
the data with no restrictions. In order to assure that
all resources are used in accordance with the policies,
UM capability is also required within the VM for con-
trol.
3.1 System Architecture/Model
For a complete UM system in a cloud computing en-
vironment, a hierarchical design is required. Our pro-
posed system implementation includes the following
key elements:
Usage Management Mechanism: This component
acts as a central controller that manages com-
munication between all components of the usage
management framework and with external ser-
vices, such as storage and content networks. The
usage management mechanism provides an inter-
face to the user. Different applications can use the
common interface provided by the usage manage-
ment mechanism.
Trust Module: This module includes a capabil-
ity to dynamically update the trust values of the
cloud resources and update the policies within the
licenses. The global Usage Management mecha-
nism in the trust module will then force actions in
response to changes, as necessary.
Virtual Machine Usage Management Mechanism:
As mentioned previously, UM within each VM is
necessary to ensure that users and applications use
data and other resources only in ways that are al-
lowed by the policies specified within the licenses.
3.2 Usage Management within a Virtual
Machine
The key elements of a usage management mechanism
that will be used in a VM are shown in Figure 3.
When we consider UM at the VM level, the control
is implemented with a finer level of granularity than
is possible when implementing UM with a central-
ized mechanism because the centralized mechanism
cannot control the actions within the individual VMs.
The process of license enforcement within a VM
using a variant of our existing UM design takes place
in six steps. In step 1, an application queries the usage
management mechanism about whether or not a par-
ticular action act on a given resource by a given user
CLOSER2013-3rdInternationalConferenceonCloudComputingandServicesScience
504
APPLICATION
OPERATING
SYSTEM
USAGE MANAGEMENT
FRAMEWORK
CONTEXT
INSTANCE
GENERATOR
USAGE MANAGEMENT
MECHANISM
INTERPRETERLICENSE
user info
resource info
action = act
allowed/
not allowed
system info
user, system
resource info
context instance
iC
action
(act)
activity
(acv)
true/
false
allowed (iav)
iav = <acv, iC>
1
2
3
4
5
6
Figure 3: Operation of the Usage Management Framework
within a Virtual Machine.
is allowed. As a part of the request, the application
provides information about the user, the resource, and
the action that is to be performed. In step 2, the usage
management mechanism obtains the current state of
the computing environment from the operating sys-
tem. The type of information obtained by the usage
management mechanism depends upon the manner in
which the context is specified. This information may
include current location, day, date, time, IP address,
etc. In step 3, the usage management mechanism gen-
erates the context instance iC, by using the current
values of system parameters, user information, and re-
source information. This is followed by step 4, where
the usage management mechanism queries an inter-
preter to determine the activity acv that corresponds to
the act information provided by the application. Once
the activity corresponding to the action is obtained,
the usage management mechanism generates the ac-
tivity instance denoted by the tuple iav =< acv, iC >.
In step 5 the usage management mechanism invokes
the allowed iav function provided by the license, the
license executes an allowed iav function, and returns a
Boolean value to the usage management mechanism.
Finally, in step 6, the usage management mechanism
notifies the application whether or not the action is
authorized.
The UM design provides the basis for deciding
which operations are allowed and which are not. The
outcome of these decisions must be transmitted, in a
verifiable manner, to enforcement mechanisms in the
operating system executing within the VM. The de-
cision must be verifiable in order to ensure that the
decision is properly enforced.
3.2.1 Operational Details
The operational details of our usage management sys-
tem are visualized in Figure 4. By means of a web
Amazon
S3
Resources
&
Policy Sets
Amazon
EC2
Context
Access
Control
Resource
Manager
1
4
5b
3
5b
Policy
Manager
5a
6
1. User logs in & the access control verifies his credentials.
2.a. User credentials are passed to resource manager.
2.b. Context information is passed to resource manager.
3. Resource manager compares the context with the policy set
and gets the list of references to the available resources and
displays it to the user.
4. User selects a resource.
5a. Policy manager retrieves the selected resource from Amazon S3.
5b. Policy manager creates appropriate VM instance either in
public(Amazon EC2) or private cloud(Eucalyptus)
5c. Control monitoring & processing.
6. Returns results of processing.
2b
2a
Usage Management
Mechanism
Nodes
Eucalyptus
Cloud
Controller
Nodes
5c
5c
6
Figure 4: Component diagram of Usage Management for
hybrid Cloud-Based System.
browser user interface, the user is presented with a
login dialog box. The user enters his/her credentials
to log into the system. The credentials and the con-
text information are forwarded to the Resource Man-
ager. Once the user has been authenticated, a list of
available resources is presented on the user’s browser.
This list of resources varies depending on user con-
text. If the user selects a resource, the Policy Manager
retrieves the desired resource from a cloud computing
repository like Amazon S3. The Policy Manager then
instantiates an appropriate Virtual Machine specified
to handle resources with the same security context.
Depending on the context and policy requirements,
resources can be moved from public to a private cloud
or vice versa. The Policy Manager also monitors the
instance and returns references about the performance
of the node to the Policy Manager. Using our control-
theoretic approach, we regulate the performance mea-
sures of the system that are covered by SLA require-
ments.
3.2.2 Implementation
The implementation of the system is shown in Fig-
ure 5. The application is written using Ruby on Rails
and we use MySQL for storing user context. When
a user logs in, we use Rails to authenticate the user.
Once the user is authenticated we use Resource Man-
ager written in Ruby to decide what resources the user
Policy-basedSecurityProvisioningandPerformanceControlintheCloud
505
Usage
Management
Mechanism
Ruby/Rails
Application
Eucalyptus
Cloud Controller
Node
User
Node
Amazon
EC2
Amazon
S3
MySQL
Figure 5: Technology Architecture.
can access by comparing context information with the
policy set of the resources and the list of resource ref-
erences are returned to the user interface. Then the
user selects a resource and the policy manager re-
trieves the resource from Amazon S3 based on the
policy information. The policy manager sends the re-
source to a VM running appropriate image either to
Eucalyptus or to Amazon EC2 cloud. It is important
to note the actual resources and corresponding pol-
icy set are stored in Amazon S3 and some of the fea-
tures of the Eucalyptus cloud system have been omit-
ted from our diagram for the sake of brevity. With the
help of our software we monitor the VMs for perfor-
mance as defined in the SLA requirements and control
the VMs appropriately.
4 PERFORMANCE CONTROL
As mentioned before, different levels of security are
considered in the Cloud. Therefore, some virtual re-
sources will be equipped with more complex security
infrastructure than others. In Figure 6 we illustrate
several VMs, indicated by the hollow circles, that are
grouped in clusters depending on the different levels
of security. As the demand for security on the avail-
able VMs increases, the overhead and therefore the
cost of more secure VMs increases as well. That is
why our assumption is that there are more VMs with
lower security guarantees (Sec. Level K) than the
ones that are more secure (Sec. Level 1). One of our
main assumptions is that the VMs within each sub-
group are identical in resources. Therefore, in order
to get an estimate of the model consistent with the
available resources we propose to identify a different
model for each level of security. In a similar way, we
propose to implement a different controller for each
level of security as shown in the blocks labeled Ctrl
in Figure 6.
Figure 6: Performance Control of virtual resources for dif-
ferent levels of security. The level of security increases
from bottom to top. Notice that the amount of resources
decreases as its level of security increases.
The control method we use in our approach relies
upon the availability of a mathematical model relat-
ing virtual resources and performance measures of the
system. Based on (Nathuji et al., 2010), given an in-
put signal u
n
and a state signal y
n
of the cloud, the
general model of the cloud interactions is given by
a set of nonlinear difference equations. The existent
identification techniques for nonlinear systems are not
suitable to be implemented in real-time, furthermore
the amount of required learning data for the identi-
fication algorithm turns out to be impractical except
for the simple cases. In previous works on control of
CPU utilization, control of power in data centers and
cloud computing described in (Nathuji et al., 2010;
Yao et al., 2010; Luna and Abdallah, 2011), a lin-
ear time-invariant realization of the system is accurate
enough as long as the control inputs of the system are
constrained to a relatively ”small“ interval. The con-
trol loop including the actual plant, the model identi-
fication unit and the controller are shown in Fig. 7.
Model
Identification
Controller
Computing
Infrastructure
Figure 7: Block diagram of the control loop.
Based on (Hellerstein et al., 2004), our model is
approximated by the following difference equations,
y
i
n+1
= a
i
n
y
i
n
+ b
i
n
u
i
n
, (1)
CLOSER2013-3rdInternationalConferenceonCloudComputingandServicesScience
506
where y
i
n
is the state signal representing the perfor-
mance measure at the n-th discrete time instant, u
i
n
is
the input signal which represents the number of active
VMs in the i-th cluster where i = 1, 2, . . . , L, where L
is the total number of clusters of VMs. The system
parameters a
i
and b
i
are parameters to be
estimated by Recursive Least Square (RLS).
4.1 Proportional Thresholding
Following the methodology proposed by (Lim et al.,
2009), the desired result of our controller is illustrated
in Figure 8. In the plot, the chosen performance mea-
sure is the CPU utilization of an entire cluster of VMs.
The green dashed lines determine the lower y
l
and up-
per y
h
bounds of the desired interval where the perfor-
mance measure should be confined. The blue vertical
lines indicate whether a VM is turned on (+1VM) or
turned off (-1VM) in the system. Notice that the per-
formance remains in the interval [y
l
, y
h
] at all times
while the lower bound changes and the upper bound
remains fixed.
Average CPU Utilization per cluster
Time (s)
+1VM +1VM -1VM
ݕ
ݕ
௟ଵ
ݕ
௟ଶ
ݕ
௟ଷ
ݕ
௟ଶ
Figure 8: Plot of the desired response. The performance is
confined to the interval [y
l
, y
h
].
The bound that remains fixed y
h
is chosen based
on the SLA. To determine the appropriate value of
the changing bound y
l
the authors in (Lim et al.,
2009) proceed to take a single VM to obtain an em-
pirical mathematical expression relating the perfor-
mance measurement of interest (CPU utilization in
this example) and the workload. In this case, it is
not difficult to expect this expression to be a mono-
tonically increasing function. Once this model has
been approximated, it is possible to approximate the
total workload of the cluster given that a specific
number of VMs are activated. Therefore, depend-
ing on the empirical model we can obtain a function
y
l
= f (y
h
, M) with M being the number of currently
activated VMs to construct the interval [y
l
, y
h
]. Notice
that the fixed bound can be y
l
and by a similar proce-
dure y
h
= g(y
l
, M) can be calculated as well.
Once the empirical model has been determined a
modified integral control is proposed as follows,
u
i
n+1
=
u
i
n
+ K
i
n
(y
i
n
y
h
) if y
h
< y
i
n
u
i
n
+ K
i
n
(y
i
n
y
l
) if y
i
n
< y
l
u
i
n
otherwise
(2)
with u
i
n
being the number of active VMs in the cluster
i and the gain K
n
calculated based on the desired tran-
sient response of the linear approximation in (1). Dif-
ferent than (Lim et al., 2009) we propose an adaptive
gain that changes depending on the values of a
i
n
and
b
i
n
to keep the system closer to the desired transient re-
sponse. Notice from (2) that if the performance mea-
sure gets greater than the upper bound of the interval
y
h
, then the system activates more VMs. If the per-
formance measure gets less than the lower bound of
the interval y
l
, then the system deactivates VMs. Oth-
erwise, the current number of activated VMs remains
the same. The empirical models that define the rela-
tionship between the workload defined by the number
of threads and CPU utilization and latency are being
currently developed.
5 CONCLUSIONS AND FUTURE
WORK
In this paper, we first described the primitives and
approaches currently available that we have used to
enable simple SLA-centric control over resource uti-
lization and processing based on performance and se-
curity sensitive attributes. Thereafter, we described
in some detail a system architecture currently realiz-
able with modern open-source tools that enables this
kind of dynamic information control. Finally, we dis-
cussed our experiences with a prototypical implemen-
tation of our proposed system architecture.
We have adopted a control systems approach
based on model identification and proportional
thresholding. It is capable of guaranteeing perfor-
mance regulation while satisfying the security re-
quirements of the applications served by clouds. One
of the main contribution of the proposed methodol-
ogy is the addition of control theory to UM policies.
It opens a set of possibilities supported by well known
computer science tools which are complemented with
mathematically supported control theory approaches.
We have multiple clouds because different cloud
systems will have different security characteristics.
We have demonstrated that we can move data from
a cloud repository to an appropriately configured VM
within a cloud system based upon the policies spec-
ified in the license. We enforce the Policies within
Policy-basedSecurityProvisioningandPerformanceControlintheCloud
507
the VM so that the use of those resources is regulated
by the policies of the resource. This approach helped
us to design and develop a policy based usage man-
agement system in conjunction with control theory to
regulate performance and provide security to multiple
cloud systems.
In the future, we plan to move away from our cur-
rent web-centric model, examining more data-centric
tooling, though we expect to remain committed to
open source tools. We will also incorporate more
standards, like the Extensible Access Control Markup
Language (XACML), to describe policies and con-
trols, and work to establish a clear model behind our
work in order to more deeply understand the intrinsic
limitations of this problem domain. Currently, we are
developing the necessary models, as well as design-
ing the strategies involving the migration of resources
to guarantee the stability of the controller.
REFERENCES
Haykin, S. (2002). Adaptive Filter Theory. Prentice Hall.
Hellerstein, J., Diao, Y., Parekh, S., and Tilbury, D. M.
(2004). Feedback Control of Computing Systems.
John Wiley & Sons.
Jamkhedkar, P. A., Heileman, G. L., and Lamb, C. C.
(2010). An interoperable usage management frame-
work. In Proceedings of the tenth annual ACM work-
shop on Digital rights management, DRM ’10, pages
73–88, New York, NY, USA. ACM.
Jamkhedkar, P. A., Lamb, C. C., and Heileman, G. L.
(2011). Usage management in cloud computing. In
IEEE CLOUD, pages 525–532.
Kulhav
´
y, R. (1987). Restricted exponential forgetting in
real-time identification. Automatica, 25(5):589–600.
Lim, H., Babu, S., Chase, J., and Parekh, S. (2009). Au-
tomated control in cloud computing: Challenges and
opportunities. In Proc. of 1st Workshop on Autom. Ctrl
for Datacenters & Clouds., pages 13–18, Barcelona.
Luna, J. M. and Abdallah, C. T. (2011). Control in com-
puting systems: Part ii. In IEEE Multi-Conference on
Systems and Control, Denver, CO.
Nathuji, R., Kansal, A., and Ghaffarkhah, A. (2010). Q-
clouds: Managing performance interference effects
for qos-aware clouds. In Proceedings of the ACM
European Society in Systems Conference 2010, pages
237–250, Paris, France.
Takabi, H. and Joshi, J. B. D. (2012). Policy management
as a service: An approach to manage policy hetero-
geneity in cloud computing environment. In HICSS,
pages 5500–5508. IEEE Computer Society.
Takabi, H., Joshi, J. B. D., and Ahn, G.-J. (2010). Security
and privacy challenges in cloud computing environ-
ments. IEEE Security and Privacy, 8(6):24–31.
Vliet, J. V. and Paganelli, F. (2011). Programming Amazon
EC2. O
´
reilly Media Inc.
Yao, J., Liu, X., Chen, X., Wang, X., and Li, J. (2010). On-
line decentralized adaptive optimal controller design
of cpu utilization for distributed real-time embedded
systems. In Proceedings of the 2010 American Con-
trol Conference (ACC’10), pages 283–288, Baltimore,
MD.
Zissis, D. and Lekkas, D. (2012). Addressing cloud com-
puting security issues. Future Gener. Comput. Syst.,
28(3):583–592.
CLOSER2013-3rdInternationalConferenceonCloudComputingandServicesScience
508