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

 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

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

Virtual Machine

Applications &

Data

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

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

action

(act)

activity

(acv)

true/

false

allowed (iav)

iav = <acv, iC>

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

Resources

Policy Sets

Amazon

EC2

Context

Access

Control

Resource

Manager

Policy

Manager

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.

Usage Management

Mechanism

Nodes

Eucalyptus

Cloud

Controller

Nodes

Figure 4: Component diagram of Usage Management for

hybrid Cloud-Based System.

browser user interface, the user is presented with a

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

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.

Sec. Level 1

Sec. Level 2

Sec. Level 3

Sec. Level 4

Sec. Level k

Ctrl 1

Ctrl 2

Ctrl 3

Ctrl 4

Ctrl k

Resource

Allocation

Unit

User

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

and a state signal y

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,

n+1

= a

+ b

, (1)

CLOSER2013-3rdInternationalConferenceonCloudComputingandServicesScience

506

where y

is the state signal representing the perfor-

mance measure at the n-th discrete time instant, u

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

∈ ℜ and b

∈ ℜ 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

and up-

per y

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

, y

] at all times

while the lower bound changes and the upper bound

remains ﬁxed.

Average CPU Utilization per cluster

Time (s)

+1VM +1VM -1VM

௛

௟ଵ

௟ଶ

௟ଷ

௟ଶ

Figure 8: Plot of the desired response. The performance is

conﬁned to the interval [y

, y

The bound that remains ﬁxed y

is chosen based

on the SLA. To determine the appropriate value of

the changing bound y

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

= f (y

, M) with M being the number of currently

activated VMs to construct the interval [y

, y

]. Notice

that the ﬁxed bound can be y

and by a similar proce-

dure y

= g(y

, M) can be calculated as well.

Once the empirical model has been determined a

modiﬁed integral control is proposed as follows,

n+1







+ K

− y

) if y

< y

+ K

− y

) if y

< y

otherwise

(2)

with u

being the number of active VMs in the cluster

i and the gain K

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

and

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

, then the system activates more VMs. If the per-

formance measure gets less than the lower bound of

the interval y

, 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