(IJACSA) International Journal of Advanced Computer Science and Applications,

Vol. 11, No. 2, 2020

Machine Learning based Access Control Framework

for the Internet of Things
Aissam Outchakoucht 1, Anas Abou El Kalam2, Hamza Es-Samaali3, Siham Benhadou4
LISER Laboratory, Hassan II University, ENSEM School, Casablanca, Morocco, IPI, Paris, France1, 3, 4
Cadi Ayyad University, ENSA School, Marrakech, Morocco2

Abstract—The main challenge facing the Internet of Things Motivated by the above, we perceive the relationship
(IoT) in general, and IoT security in particular, is that humans between IoT and ML much like the relationship between the
have never handled such a huge amount of nodes and quantity of human body and its brain. Our bodies gather sensory input
data. Fortunately, it turns out that Machine Learning (ML) such as sight, sound, smell, taste and touch while our brains
systems are very effective in the presence of these two elements. focus on gathering that data and making sense of it.
However, can IoT devices support ML techniques? In this paper,
we investigated this issue and proposed a twofold contribution: a The remainder of this paper is presented as follows:
thorough study of the IoT paradigm and its intersections with Section II exposes an overview of ML applications in IoT
ML from a security perspective; then, we actually proposed a scenarios; then Sections III and IV reveals the building blocks
holistic ML-based framework for access control, which is the of the ML-based framework aiming to handle IoT AC, as well
defense head of recent IT systems. In addition to learning as all the required concepts to understand it. Next, Section V
techniques, this second pillar was based on the organization and provides the theoretical and technical details of
attribute concepts to avoid role explosion problems and applied implementation which are applied to a smart city case study,
to a smart city case study to prove its effectiveness. before moving to the last section in which we discuss and
evaluate the results.
Keywords—Access control; internet of things; machine
learning; security; smart city II. RELATED WORKS
I. INTRODUCTION Basically, ML algorithms are computer programs that can
essentially learn and improve their accuracy by looking at data
Access Control (AC) plays a pivotal role in the security
without being explicitly programmed. In a more formal
world given its mission of protecting digital and physical
wording: “A Computer program is said to learn from an
accesses by delimiting and enforcing who has access to what
experience E with respect to some task T and some
and in which conditions [1]. However, most of the AC
performance measure P, if its performance on T, as measured
solutions we find in the literature tend to consider the IoT as a by P, improves with experience E” [2, 3]. In this section, we
single block that is characterized mainly by the limited storage
will be exploring the most promising and latest ML
and computing capacities. In this paper, we will come back to
applications used in order to secure IoT environments.
this unfair and unrealistic view that slows down the
elaboration of a holistic approach to address AC in IoT A. Learning Algorithms for Constrained Environments
environments. Moreover, relying on a single technique to Despite the common sayings that build a delusive wall
address an issue that is as complex as IoT is also a weakness between ML algorithms and constrained nodes given the
that confines the performance of many IoT security-oriented computational and storage limitations of the latter, many
models. studies combined these two worlds in order to answer both
To fulfil the AC requirements, this paper will application and security issues. In this section we are about to
progressively build a global framework that not only focuses discuss the most recent and relevant ones.
on policy management and AC models, but also digs deeper Let us begin with a recent work [4] that combined the
into the mechanisms that accurately fit them; which leads to a strengths of current neural and tree-based learning techniques
smooth and coherent Machine Learning (ML) integration in conjunction with ternary (-1, 0, 1) quantization to enable
going down to highlight what and where ML algorithm(s) computation and size compression of NN models in IoT
should be implemented. platforms. This technique outperformed the state-of-the art by
To do so, we first need to delimit the perimeter covered by 11.1%, 52.2% and 30.6% in the number of computations, the
the IoT by giving a much more representative definition of the model size, and the overall memory footprint respectively,
term, which will allow us later to tackle the question of AC without losing too much in terms of accuracy.
with a much more appropriate vision, and above all, will lead Another study [5] focused on the IoT device side rather
us to know where and how we can use the power of ML to than the cloud. The proposed model is developed on a
take advantage of the large amount of objects and data we are relatively simple tree learnt in a low-dimensional space for
handling. efficient prediction on IoT nodes like Arduino UNO or
ATmega328P boards with 16 MHz, 2 KB RAM and 32 KB
ROM. The authors executed their model using several datasets
This work was supported by the IGS Group, Paris, France

331 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 11, No. 2, 2020

and proved that it was able to make predictions within Strength Indicator values from multiple iBeacons. It was able
milliseconds, had lower battery consumption than the state-of- to learn better action policies with at least 23% improvement
art and could fit in KB of memory. in terms of distance to the target as well as almost 67% more
gathered rewards compared to the supervised Deep RL
Additionally, a model called MorphNet was suggested by approaches.
Google [6] to automate the design of Deep Neural Networks
(DNNs). This approach is specifically adjustable to meet B. Learning Applications for IoT Security
constrained environments’ requirements without Now that we have seen many ML-based applications in the
compromising the performance, it actually optimizes the DNN IoT, let us move to some studies that tackled security
by iteratively shrinking and expanding. The study showed that problems (always in IoT environments) through ML tools.
MorphNet is simple to implement and fast to apply, which is
why it is a better choice for IoT scenarios. In fact, Support Vector Machines (SVMs) are one of the
first and most used ML models. They represent standard
ShuffleNet [7] is another contribution in this direction. It is classification models, generally known for splitting
a particularly computation-efficient CNN designed specifically hyperplanes. Data sorting is achieved through maximizing the
for mobile devices which are essentially characterized by their distance between the hyperplane and the nearby training
limitations in terms of computational power. It is mainly based samples of each class. SVMs are more adapted to datasets with
on the power of 1×1 convolutions combined with channel a large number of features but a relatively small number of
shuffle with the aim of reducing the required cost for samples [13]. In the IoT world, a study [14] proposed a linear
computation without neglecting the accuracy. Being SVM-based Android malware detection system to secure IoT
implemented on an ARM-based mobile device, the model platforms. The comparison that was led between the
attains up to 13× actual speedup over AlexNet. performance of the model and other ML algorithms
Besides, the authors in [8] suggested a NN-based outbalanced the SVM method. Besides, SVM was also used to
implementation that takes benefit from the communications compromise cryptographic devices [15, 16]. However, one of
passed between IoT nodes. Theoretically, this work is founded the big challenges in multidimensional SVM problems is the
on the UAT theorem affirming that a NN with a single hidden tough task of selecting a suited kernel.
layer is enough to compute a bounded approximation of a Another generic method is Random Forest (RF): an
generic continuous function. In fact, this remark has led to accumulation of Decision Trees (DTs), which means that they
integrate intelligence into IoT constrained platforms by means are built and trained in order to vote for the output class [17]. A
of some (local and on-the-fly) computations as the data study [18] over 17 IoT devices belonging to nine categories
navigate between the IoT devices using the collective behavior affirmed that RF (among other ML methods) presents
of such networks. significant improvements in the identification of unauthorized
A Mobile and Edge Computing (M/EC) solution was IoT nodes. Another ML-based study [19] was performed on
proposed in [9] to bring computation near the IoT end-nodes IoT environments to detect DDOS attacks. In this regard, RF
by applying CNNs, RNNs and RL at the edge of IoT provided slightly superior results compared to other ML
networks. The very idea of this work is to implement methods. That being said, it is important to emphasize that RF
Information-Centric Networking on top of the IoT via some methods are not always feasible, specifically over large
techniques namely shared weights, pooling, and in-network datasets as they require the construction of a -relatively- large
caching to solve storage issues on IoT nodes. This approach number of DTs.
led to remarkable reduction in latency for time-critical In another direction, UL is represented by the popular K-
applications. Means with the key objective of Data clustering (k being the
Another study [10] digs deeper into the technical hardware number of clusters). The algorithm consists of assigning each
requirements to implement DL algorithms over IoT devices. data sample to one of the k clusters based on their (similar)
The authors implemented several models inside three boards: features. Usually UL models are privileged when the dataset is
Qualcomm Snapdragon 800 used for phones and tablets (4- not labelled. In IoT, k-means clustering was used to
core 2.3 GHz CPU, 1GB of RAM and 8MB DSP), Intel distinguish Sybil attackers from normal sensors through
Edison principally oriented to wearables and form-factor clustering the channel vectors in industrial WSNs [20].
sensitive IoT (500MHz dual-core CPU, 1 GB of RAM) and Nevertheless, this technique has many limitations, namely the
finally Nvidia Tegra K1 used for example in June IoT Oven need to have roughly equal numbers in each cluster for the
[11], Nexus 9 phone and IoT-enabled cars (up to 1.7GB of algorithm to properly work, as well as the non-trivial task of
RAM). The study proved, inter alia, the feasibility of choosing k [21].
implementing DL techniques on IoT oriented boards. Now, let us move to the deep sphere, and begin our survey
A more general approach was proposed in [12], in which by Convolutional Neural Networks (CNNs). Actually, the
the authors came with a semi-supervised deep RL model basic idea of a CNN is to put a bit of structure in NNs [22] by
designed to fit smart cities. Its inference engine exploits shrinking the enormous number of connections between
Variational Auto Encoders (VAE) to generalize optimal layers, thus optimizing the training time. One of the main
policies. The model was implemented to handle localization benefits of CNNs is their end-to-end nature ensured by their
issues in a smart building case study portrayed as an ensemble built-in “features extraction” ability. Yet CNNs still have a
of labeled positions associated with the Received Signal high computational cost; hence the difficulty of implementing

332 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 11, No. 2, 2020

them on IoT constrained nodes. Many studies managed to ones. Though, up to now the training phase of GANs still
bypass this limitation though, especially using distributed unstable and a tough task [21].
architectures [23]. Another study showed that CNNs could
Providing a large amount of training data is not always an
help in Android malware detection by means of raw sequence
static analysis (RSSA) of disassembled programs [24]. easy task; hence, finding alternatives is a matter of serious
concern for ML experts. Reinforcement Learning (RL)
One more giant pillar of ML nowadays goes under the consists of learning behavior only through interactions
name of Recurrent Neural Network (RNN). It is undoubtedly between an agent (usually represented by the algorithm) and
one of the ML big discoveries thanks to their particularity of its surrounding environment; in fact this learning process
having memory. They can read inputs in sequence, and consists of increasing the rewards it receives from the
“remember” some information/context thanks to their hidden environment. Many researchers focus on the application of RL
layer activations that get passed from a given time-step to the to IoT security; for instance, the work in [28] opted for an RL
following [22]. Accordingly, RNNs can achieve excellent approach to learn a sub-band selection policy so that it could
results in classifying network traffic and detecting malicious avoid both jammer signals as well as interference from other
behavior. Besides, RNNs could be a good choice for IoT since radios in wideband autonomous cognitive radios (WACRs).
it produces massive sequential data from different nodes. For Two of our previous works [1, 29] tackled the Access Control
instance, a previous work [25] proved the worth of using (AC) in IoT scenarios, the two building blocks were: first
RNNs to detect network traffic behavior by modeling it as a taking into account the smart devices’ context while making
sequence of states changing over time. Yet, vanishing and an AC decision; and proposing AC policies that can be
exploding gradient problems still the ultimate nightmare of improved and optimized over time. However, given the
RNNs. enormous and heterogeneous amount of data generated by IoT
devices, the proposition benefits from the power of RL, to
In another direction, many researchers consider that there accomplish this task. The problem with RL algorithms is that
was various contributions in ML in recent times, but they require a large number of practice run (given their trial-
Generative Adversarial Networks (GANs) are the only error nature) before they can make significant progress.
contribution that could be called a breakthrough in the last
decade. GAN trains two models at the same time: a generative It is worth noting that, in addition to provide an explicit
model G to identify the data distribution, and a discriminative survey of the latest and relevant works in IoT and ML, one of
model D to predict the probability that a sample came from the motivations of this related works section is to prove that
the training data rather than G [26]. A recent work [27] ML is already used in the IoT world, and consequently to
realized a GAN-based architecture in order to secure IoT disprove the idea claiming that IoT and ML are two parallel
systems by detecting abnormal behavior. GANs may have a universes. In the following section, the proposition of this
potential application in IoT security especially in zero-day- paper is presented with all the necessary details.
like threats given their ability to learn diverse attack scenarios
Table I summarizes and compares these studies especially
and then to generate innovative attacks beyond the existing
based on their achievements in a number of IoT situations.
TABLE. I. COMPARISON AND SUMMARY OF ML STUDIES FOR IOT

Application domains Used algorithms Achievements Studies

Semi-supervised D-RL Bring computing next to IoT devices

Networks traffic optimization in
Auto Encoders Reduction in latency [9], [10], [12]
indoor or smart city scenarios
D-NN, CNN, RNN More gathered rewards
Local & on-the-fly computations
(Relatively) Simple processing D-NN, CNN Up to 13× actual speedup over AlexNet
[4], [5], [6], [7], [8]
situations Low-dimensional trees Decreasing model size, memory consumption
Prediction within milliseconds
Significant improvements in the identification of
[14], [15], [16], [18], [19],
Malware & Intrusion Detection SVM, RF, CNN unauthorized IoT nodes
[23], [24],
Compromising cryptographic devices
Zero-day attacks
Network threats &
RNN, GAN Good results in time-based environments [25], [27]
Network traffic behavior
Data augmentation
Policy optimization
Security policy improvements RL [1], [28], [29]
Policy efficiency
Zero-day & Sybil attacks detection
New and unprecedented attacks K-means, RL [1], [20], [28], [29]
Avoid jammer signals

333 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 11, No. 2, 2020

III. PRELIMINARIES oriented chips, which are mainly due to the excessive demand
and hot market of AI applications these days. This subsection
We begin this section by exposing the research questions exposes three reasons to motivate researchers –and investors–
standing behind our work, together with all the essential not to draw a spontaneous correlation between IoT and ML
details required to understand our proposition. ineptness:
A. Problematic  IoT > constrained nodes: As explained before, one key
The nature of a large portion of IoT devices (e.g. idea to clarify when talking about IoT is that it is more
Healthcare, critical infrastructures) makes security the number than just a collection of constrained nodes. Not to
one priority: it is, literally, a matter of life and death. In confound with Wireless Sensor Networks.
addition, the density, heterogeneity and autonomy are intrinsic
characteristics of these systems that not only expend the  Hardware progress: The AI market is in an exponential
perimeter of potential attacks but also their magnitude [21]. growth, which leads to more investments, then to more
innovations. This climate could only be beneficial for
However, it turns out that treating all security aspects in ML community. With this in mind, one can take a look
one proposition is not a reasonable task; hence, this paper is at Amazon store for example to see how the ratio of
focusing on the AC cornerstone because of its nature of hardware size to its storage capacity is decreasing
protecting the access to digital and physical resources by faster than ever before. Regarding computation
delimiting, managing and enforcing who has (has not, is capacities the progress is astonishing as well, for
obliged to have) access to what, when and under which instance, just few months ago, NVIDIA announced a
conditions [1]. 70mm x 45mm AI computer, 4 GB memory, Quad-
Furthermore, the abovementioned IoT features impose core ARM® A57 CPU and 128-core NVIDIA Maxwel
intelligent and dynamic management instead of traditional and GPU [30].
impractical one; we believe that IoT must benefit from these  Software evolution: What is true for hardware also
features considered so far as obstacles, IoT nodes need to holds for software. Section II is an illustration of the
“learn to look after each other”. To do so, our research has active race in proposing new and suitable ML
confronted many speed bumps that we have tried to demolish algorithms for IoT. Besides, many dedicated and
in this article. extremely optimized ML libraries are already easing
First, the misleading idea that reduces the IoT to programmers’ life. For instance, Google’s Tensorflow
constrained devices, but more importantly: The need to exploit Lite transforms heavy TensorFlow models into
the quantity of IoT devices and data as a catalyst for security compressed flat buffers, which are then loaded into a
to emerge. mobile or embedded device [31].
Not to mention the necessity for models that go beyond Furthermore, there are several active research directions
simply defining AC policies to understand the context of each that could lead to further findings (even breakthroughs) in
smart device and continuously improving these policies, IoT-adapted ML applications: (i) parallel computing in
without falling into the trap of static management or role training phase using Graphical Processing Units and Tensor
explosion. Processing Units (GPUs/ TPUs), (ii) transfer learning in order
to swiftly transfer the knowledge from pre-trained models,
To the best of our knowledge, there seems to exist no (iii) fog computing to decrease communications overhead size,
previous work presenting a holistic ML-based framework for data traffic, user-side latency, (iv) fast optimization algorithms
IoT answering these problematics. [32].
B. IoT and Computation Paradigm C. Background
At first sight, IoT is a concatenation of two words: The aim of this section is to explore two concepts that are
“Internet”, which refers to connectivity and communication essential to understand our proposition, namely, AC and IoT
aspects; and “Things”, which is a generic and global term that architecture.
includes all kinds of objects, whether large or small, powerful
or not. In that sense, every “thing” endowed with communication In fact, AC is of paramount importance being the entry
capacity is an IoT device. Put this way, one can easily classify point of every system after the identification/ authentication
the aforementioned constrained nodes, together with mobile phase, thus securing any system must pass through (if not
phones, a Raspberry Pi board and cloud servers as IoT begin with) controlling its accesses. In the literature, tens of
devices; and can also set traditional TVs, calculators or models are handling this issue, one of the most popular is Role
pillows outside the IoT scope (unless they are connected). Based Access Control [33] (RBAC); without going into too
much details: instead of granting (or removing) a separate
Actually, no one can deny the difficulty (sometimes even permission to every subject in the network, the model
the impossibility) of implementing complex ML tools on aggregates these subjects by roles and thus gets a lightweight
several types of constrained nodes. Yet, these latters remain version of its AC policy. Yet, end devices are not involved in
ambiguous especially given the recent innovations in ML- AC decision, also, even with the aggregation of subjects into

334 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 11, No. 2, 2020

roles, IoT systems still have astronomical number of objects

and actions to manage. Attribute Based Access Control [34]
(ABAC) is another model that is taking up more and more
space recently, its basic concept is to identify subjects and
objects through attributes (characteristics), then the
permissions are granted according to these attributes, which
could be any relevant security-characteristics, this makes
ABAC, unlike RBAC, more adapted to afford fine-grained AC
highly valued in IoT circumstances. Still, it comes with a
terrifying and intolerable drawback: complexity. Other models
[35] tried to make many tradeoffs but generally fall into one of
the aforesaid limitations.
Fig. 1. Simplified Presentation of OrBAC Layers.
Reasonably, Organization Based Access Control [36]
(OrBAC) is a remarkable model that we believe it could be
used as a foundation to an IoT-oriented solution. It inherits
and extends the benefits of RBAC by proposing abstractions
to all the AC elements (subject to role, object to view and
action to activity – see Fig. 1), then it adds two more
dimensions to the decision making process: context (crucial
for IoT) and organization; this latter makes it a centralized
model, but we will see in the next section how to turn this
limitation into a strength.
In the other hand, and given what is been elucidated earlier
in this paper, IoT is not a single homogenous block; hence, in
order to propose reasonable solutions, there is no other way
but to have a conception of the main building blocks
composing standard IoT environments. To do so, of course
one could suggest several subdivisions but since the intention
of this paper is to conciliate IoT and ML, the proposed
categorization needs be centered on computing capabilities. Fig. 2. Intersections between IoT Layers.
In that sense, many researchers [37, 38] agree that every
IoT platform could be molded into one or more of the IV. CONTRIBUTION
following categories: C1: the constrained layer, here is where In this section we present the AC framework that takes
the constrained devices are located (physical constraints on into consideration all the previously discussed requirements.
many characteristics such as size, weight, available power and
energy [39] which make them unable to accomplish more than A. Global Questions Need Global Answers
basic tasks); C2: this category includes more powerful nodes, One can fairly claim that IoT is the largest and most
which are capable of executing relatively serious heterogeneous artificial network humans have ever made, it is
computations. In fact, the vast majority of the everyday smart becoming earth’s nervous system. Therefore it would be
devices fit into this category (e.g. smartphones, smart homes naive, if not irrational, to search for elementary narrowed
components); finally, C3: Computational or offloading layer, solutions to such a complex and multifaceted problem; instead
it is the most powerful one, it could be Cloud or local servers, what this paper is suggesting is a multi-layer AC solution, that
computers, GPUs/TPUs and so on. exploits ML and OrBAC strengths to answer IoT burning
questions exposed in Section III-A.
Generally, complex environments (like smart cities) are a
combination of the three layers. Another point to underline is First thing to consider when treating AC in IoT is the
the absence of explicit and rigid boundaries between these overall architecture, whether it is distributed (like blockchain
layers; this is mainly due to the dynamic nature of the IT based solutions [40]) or centralized as long as it is equipped
domain, in fact what we call powerful today (or for a given with the collaboration aspect. In fact, each of these two
task) might be considered constrained tomorrow (or for architectures has its advantages and drawbacks [29]; however,
another task) and vice versa. Based on some examples, Fig. 2 without loss of generality, this paper mainly focus on the
exposes the borders and intersections between these categories . second one for the following reasons: (i) IoT, at least as we
know it today, requires that each device has an owner, whether
As shown above, typically what most people call IoT is in a person, in the case of a smart home for example or an
general the intersection between the three layers, it is where organization in industry or in smart cities. So practically an
all sorts of devices are interacting with each other to create approach that takes this aspect into consideration will be
this large universe of smart devices. closer to reality. (ii) Even in distributed architectures, there
always should be an entity in charge of defining AC policies
for the IoT nodes, unless the device is open to everyone (and

335 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 11, No. 2, 2020

therefore no need for AC at all). For simplicity reasons, in 2) Inference stage request processing: Now that AC
both scenarios this entity is given the name “object owner” policies are initiated and stored in the PIP, phase 2 begins: a
(OO). subject is willing to execute an action on an object , to
As shown in Fig. 3, the object owner i ( ) owns several do so the following request is sent to the Policy Enforcement
IoT devices which can belong to any category (C1, C2 or C3). Point (PEP):
He defines their AC policies and stores them either locally, in
(4)
a distributed manner on multiple servers or even in large
networks (like blockchain) using smart contracts. Henceforth, When the PEP gets the access request, it triggers the
the location of these policies is called Policy Information process of matching (phase 3); it is the step where we go up
Point (PIP), as defined by the ISO/IEC standard for the access from concrete to abstract entities in order to reduce the
control framework [41] and the XACML related architecture complexity. So the PEP transfers (4) to the Policy Decision
[42]. Point (PDP), which in its turn requests the PIP by an:
As mentioned before, the choice of OrBAC as the (5)
background model to an IoT-oriented framework is defended
by the following reasons: First, it has the concept of After that, the PIP, responds by the corresponding
organization by design thus no need for extra dimensions to matches:
designate the OO. Also, OrBAC is distinguished by two other (6)
features crucial for IoT environments, namely an advanced
level of abstraction required to alleviate the complexity Up till now, the PDP has all the static features to take the
produced by the colossal number of devices; together with the decision. However, even if we have already handled two IoT
context incarnated in all OrBAC rules, which will ease the major worries, namely context and complexity, the process
collect of real time contextual information from the end nodes still lacking dynamism. In fact, policies are statically stored in
for better AC decisions. In a more formal sense, the AC policy the PIP without any learning from past experiences. To answer
is stored in the PIP as rules presented in this form: this, (6*) will be coupled with an extra feature: a ratio
reflecting the probability of a safe access granting given the
(1) aforementioned characteristics:
Where org stands for organization or the view owner, r for (7)
role (aggregation of subjects), v for view as a collection of
objects, ay for activity which is an abstraction of actions To do so, (1*) needs to be joined with the probability
whereas c is the context. Thus the previous rule declare that: feature which is set by default (in the policy definition phase)
In the organization org the role r has permission to execute the to p=1, then it is updated over time:
activity ay on the view v under c circumstances.
(8)
1) Pre-request stage: Policy initiation: In step one, when Now, and based on a threshold fixed by the OO, the PDP
the OO have to define a new rule, either the device o fits in can decide (phase 4) and inform the PEP, and consequently
one of the existing views v so the affectation : the requester, about the final decision. Note that this threshold
(2) is executed; if not the role is automatically created when has two essential benefits: first it gives the organization a
declaring a new permission and the subject and role get the better personalization of the framework, in addition to
same name. allowing it to define even several thresholds given the
criticality of certain resources or context.
Another key concept of this framework is the process of
matching abstract and concrete entities, it also begins in phase
one: In fact, besides the aforementioned rules, the PIP
contains two types of match functions:
Usual correspondences in the form of ( )
as shown in (2) are employed for the frequently used entities,
while generic match functions based on attributes, for
instance:
(3)
Which means if a subject s, has the
attributes then it belongs to the role . Yet since the
designation of could as well be considered as an attribute of
itself, we can use (3) for both. Fig. 3. Step 1: to define AC Policies.

336 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 11, No. 2, 2020

3) Post-request stage learning and upgrading: The final Note that the proposed framework is a decentralized one.
phase -5- in our process is a post access one, it consists of In fact, the concept of organization is introduced to
calculating a feedback rating of the experience, which will be decompose complex IoT environments into reasonable and
used by the PDP to generate AC policy updates leading to manageable groups, not to turn them into one giant centralized
one. For instance, if we have to manage AC in a smart city
more accurate decision in the future (a concrete example is
situation, the framework will treat this multidimensional
given in the case study section). The output will be stored in platform as a collection of organizations interacting and
the “learning matrix” which will have 6 columns collaborating with each other.
(organization, role, view, activity, context, feedback) and as
many line as the number of experiences the system could Several studies have examined the collaboration issues in
store; while feedback is a rational number between 0 and 1. OrBAC [37, 44, 45, 46], either by creating further abstract
entities, web services, or even through prior agreements
The learning algorithms run in phase 5 varies according to between the involved organizations, however the definition of
the hardware resources of the system but also according to the AC relationships using attributes that we saw in 1.b. (Table II)
layers defined in Section III-C. However, generally in is a better alternative in IoT situations since with one tool we
complex and multifaceted IoT environments we propose answer both intra and inter organizations AC concerns. An
using: RL and many resource consuming SL scheduled, for example of this scenario will be discussed in the following
example, periodically in category C3; SL up to a reasonable section, where we are exploring a smart city case study.
size of the learning matrix in C2; while leaving normal
equation stechnique or no ML at all to C1. Table II
summarizes the previously detailed steps.
For simplicity reasons, we were focusing in permission
formulas, however what goes for permissions is also valid for
obligations and prohibitions [43].
B. The Algorithm
The steps discussed in Table II are compressed in the
following Fig. 4 to explain the overall functioning of the
algorithm. In fact, the framework could be segmented into
three main time frames: (i) pre-request tasks, which handle the
definition of the AC policies; (ii) request processing,
involving all the actions triggered after an access request up
till the subject receives back a permission/rejection; (iii) post-
request actions that are responsible for the learning and policy
improvements.
Fig. 4. Overall Functioning of the Framework.

TABLE. II. SUMMARY OF THE FRAMEWORK’S PHASES

Location/
Action Responsible Example
destination

1.a. AC policy definition OO PIP

1.b. AC relationship definition (either explicit or

Phase 1 OO PIP
through attributes)

1.c. Threshold definition OO PDP

2.a. The subject request executing an action over a
Phase 2 The requester/ subject PEP
resource/object

3.a. Request for decision PEP PDP

Phase 3 3.b. Request matching information PDP PIP

3.c. Matching response PIP PDP

Phase 4 4.a. Make decision PDP PEP

5.a. Update learning matrix PDP PDP

Phase 5 5.b. Learning process PDP PDP

5.c. Send updates periodically PDP PIP

337 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 11, No. 2, 2020

Now that we have exposed all the stakeholders, first thing

V. CASE STUDY to do as the owner of CRA is to define a primary AC policy,
The IoT is growing by leaps and bounds, thus creating this so in the PIP there will be two sorts of rules: (i) Those of the
large, smart and autonomous system requiring less and less 1.a. step, which will be in the form of:
human intervention. Smart city (SC) was always the case in
point that portrays this vision where devices not only interact
but depend on (even control) each other. This section presents
a SC situation to depict how the previously described
framework could be implemented in such complex * Star stands for all, i.e. any organization.
environment. Then the definition of the AC relationships, which are
Actually, the choice of SC has been motivated by many responsible for matching abstract to concrete entities, for
reasons: First, because it covers many of the other IoT use example a car is considered luxury if its release date comes
cases. Secondly, in addition to the IoT world, it is also a after 2017 and its price exceeds 100,000$:
typical case in the AI domain. Thirdly, it presents an active
research field [47], and last but not least it is becoming a
necessity in order to efficiently deal with the massive growth
of urbanization that is estimated to reach 66% by 2050 [48].
To better illustrate this example, let us take three
For luxury cars the threshold is set to 1, so that access is
organizations in a SC, namely: a car rental agency (CRA),
not granted to any role without a 100% confidence about its
which represents the organization we are willing to secure, a
previous experiences. While for the less critical view (normal)
smart parking (SP) and a police station (PoS).
the threshold is reduced to 0.8 allowing more usability. These
As a CRA, our organization is mainly composed of self- thresholds are set in the car side’s PDP.
driving cars (decomposed into two views: luxury and normal
Let’s move to phase 2. At the moment, an access request
cars); its customers are generally normal clients, VIP clients or
arrives to the PEP of the car id=79, it is a subject presenting
blacklisted ones (which makes respectively three roles: NC,
his membership badge and demanding this resource for 5 days
VIP, BC); regarding the activities it is more realistic to
starting from August the 1st, 2019.
categorize them by rental period (a1: 1 day, a2: between 1 and
3 days, and a3: more than 3 days). Finally, the context
represents the time of the year during which the request is
made: is it a peak season (peak) like summer for example, or
off season (off).
After the previous request is forwarded to the PDP, this
In this case, each self-driving car is equipped with its own latter requests matching information from the PIP via:
PEP, which means it is the car itself that receives the access
request and responds the requester by the final decision. PIP
may ideally be one (or several) local server(s) storing the AC
policy. Regarding the PDP, in general it needs to be spread out
over two layers: one part dealing with the steps (1.c., 3, 4 and
5.a.) that are executed within the smart car, and another part Then it gets the response:
that needs to be run on C3 category (namely 5.b. and 5.c.).
Nevertheless, as we will see later some lightweight versions of
these two steps could also be run within the smart car. The And since the threshold for luxury cars is set to 1, the PDP
following Fig. 5 portrays the building blocks forming this makes and informs the PEP about its final decision allowing
platform: the requester to access/use the car:

Now comes the learning phase. Actually, it is up to the

organization either to rate the experience after its end (i.e. rate
just the physical state of the car for example) or to perform an
online rating so that it gets periodic feedback from the smart
car to compute more sophisticated inferences (e.g.
geolocation, fuel consumption, speed) and then the PDP ends
up with a weighted average. To simplify our case, we use the
first option.
Therefore, after the car is back, the PDP receives the
Fig. 5. The Building Blocks of the Case Study. required ingredients to compute the feedback. Let us consider

338 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 11, No. 2, 2020

that what is important for this CRA is whether the car is back humans learn, they always have some innate skills before
in time and its mechanical situation, and that for both features anyone comes to teach them anything.
this experience was negative, so the feedback was set to 0.6.
Finally, we believe that having the ability to personalize
The learning matrix is now updated:
the threshold, not forcing an immediate update after each
experience, and allowing explicit as well as extracted
attributes bring this framework with further flexibility and
Now we can imagine that if after a while similar feedbacks adaptability to fit the IoT requirements discussed in
comes from the same organization ( ) or under the same SectionIII-A.
context or whatever, the algorithm will detect the pattern when
the learning model is run over the collected data REFERENCES
using , then the AC policy could [1] A. A. El Kalam, A. Outchakoucht, H. Es-samaali, “Emergence-Based
eventually be updated, thus improved over time. Access Control: New Approach to Secure the Internet of Things”.
DTUC ’18 Paris. 2018.
A final point that we want to highlight is about the [2] T. M. Mitchell, “Machine learning”, 7th ed. NY, McGraw Hill, ISBN:
attributes used in the matching phase. In fact, depending on 0070428077. 1997.
their nature there are two ways to collect them; either [3] Y. LeCun, Y. Bengio, G. Hinton, “Deep Learning”, Nature, volume 521,
explicitly as we saw in the previous example (i.e. as 2015.
parameters within the access request) or by extracting them [4] D. Gope, G. Dasika and M. Mattina, “Ternary Hybrid Neural-Tree
Networks for Highly Constrained IoT Applications,” in the 2nd
directly from the object. For instance, if we want to use the Conference on Systems and Machine Learning (SysML), 2019.
license plate number we could eventually use the cameras [5] A. Kumar, S. Goyal, M. Varma, “Resource-efficient Machine Learning
from the smart parking. in 2 KB RAM for the Internet of Things”; Proceedings of the 34th
International Conference on Machine Learning, 2017.
VI. DISCUSSION AND CONCLUSIONS [6] A. Gordon et al. “MorphNet: Fast & Simple Resource-Constrained
The motivation behind this work was to come up with a Structure Learning of Deep Networks,” arXiv:1711.06798. 2017.
smart, decentralized and IoT-suited AC framework. In this [7] X. Zhang, X. Zhou, M. Lin and J. Sun, "ShuffleNet: An Extremely
section we discuss theoretical and practical contributions of Efficient Convolutional Neural Network for Mobile Devices,"
IEEE/CVF Conference on Computer Vision and Pattern Recognition,
this paper as well as their strengths over existing solutions. pp. 6848-6856, 2018.
First thing to remember is that IoT are not all constrained. [8] N. Kaminski et al., "A neural-network-based realization of in-network
To prove this we went back to define this paradigm and to computation for the Internet of Things," IEEE International Conference
on Communications (ICC), pp. 1-6, 2017.
delimit its boundaries and layers; in addition to overview
[9] H. Khelifi et al., "Bringing Deep Learning at the Edge of Information-
several IoT propositions that have used ML techniques. Centric Internet of Things," in IEEE Communications Letters, vol. 23,
However, to the best of our knowledge, the existing solutions no. 1, pp. 52-55, 2019.
are narrowed ones, each one focus either on proposing a [10] N. D. Lane et al., "An Early Resource Characterization of Deep
model, managing the policy or tackling specific techniques. Learning on Wearables, Smartphones and Internet-of-Things Devices,"
The problem with this approach shows up when an in the Proceedings of the 2015 International Workshop on Internet of
organization wants to put it all together, generally the Things towards Applications, pp. 7-12, 2015.
concatenation of these uncoordinated solutions do not give [11] Nvidia. [online] Available at: https://blogs.nvidia.com/blog/2015/06/
09/gpu-powered-june-oven/, [Accessed 2 December 2019].
acceptable results. Thus proposing a holistic framework to
manage AC in IoT environments is another key contribution [12] M. Mohammadi, A. Al-Fuqaha, M. Guizani and J. Oh, "Semisupervised
Deep Reinforcement Learning in Support of IoT and Smart City
of this paper. Services," in IEEE Internet of Things Journal, vol. 5, no. 2, pp. 624-635,
2018.
Equally important, the introduction of the notion of
[13] A. L. Buczak and E. Guven, "A survey of data mining and machine
organization in IoT is notably benefic since it helps drastically learning methods for cyber security intrusion detection," IEEE
decreasing the problem of role explosion, which is the number Communications Surveys & Tutorials, vol. 18, no. 2, pp. 1153-1176,
one problem challenging role-based and attribute-based AC 2015.
solutions. In fact, an IoT environment could always be broken [14] H.-S. Ham, H.-H. Kim, M.-S. Kim, and M.-J. Choi, "Linear SVM-based
off into several organizations and therefore, depending on android malware detection for reliable IoT services," Journal of Applied
their mission, the roles will be manageable; and of course Mathematics, vol. 2014, 2014.
what goes for roles also goes for views, activities and context. [15] A. Heuser and M. Zohner, "Intelligent machine homicide," in
Note that the organization aspect has by no means been a International Workshop on Constructive Side-Channel Analysis and
Secure Design, pp. 249-264, 2012.
source of centralization, it is rather a push toward more
[16] L. Lerman, G. Bontempi, and O. Markowitch, "A machine learning
decentralization and collaboration. approach against a masked AES," Journal of Cryptographic
Engineering, vol. 5, no. 2, pp. 123-139, 2015.
Another interesting and captivating point: the learning
process we have introduced actually differs from the [17] L. Breiman, "Random forests," Machine learning, vol. 45, no. 1, pp. 5-
32, 2001
traditional procedure commonly used in the current AI
[18] Y. Meidan et al., "Detection of Unauthorized IoT Devices Using
applications, which consist first of a learning phase followed Machine Learning Techniques," arXiv preprint arXiv:1709.04647, 2017.
by the prediction phase. In our framework it is rather a [19] R. Doshi, N. Apthorpe, and N. Feamster, "Machine Learning DDoS
minimum of basic knowledge is initiated in the beginning then Detection for Consumer Internet of Things Devices," arXiv preprint
learning came to fine-tune this expertise. It is actually how arXiv:1804.04159, 2018.

339 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 11, No. 2, 2020

[20] Q. Li, K. Zhang, M. Cheffena, and X. Shen, "Channel-based Sybil [34] E. Yuan, J. Tong, “Attributed based access control (ABAC) for Web
Detection in Industrial Wireless Sensor Networks: a Multi-kernel services,” in: IEEE Int. Conf. Web Serv., IEEE, 2005.
Approach," in IEEE Global Communications Conference, pp. 1-6: [35] A. Ouaddah, H. Mousannif, A. A. Elkalam, A. Ait Ouahman,
IEEE, 2017.
[36] “Access control in the Internet of Things: Big challenges and new
[21] M. Al-garadi, A. Mohamed, A. Al-ali, et al. “A survey of machine and opportunities”, Computer Networks 112, pp. 237–262, 2017.
deep learning methods for internet of things (IoT) security”. arXiv
[37] A. Kalam, et al., “Organization based access control,” in: IEEE 4th Int.
preprint arXiv:1807.11023, 2018.
Work. Policies Distrib. Syst. Networks, IEEE Comput. Soc, pp. 120–
[22] M. Ford, “Architects of Intelligence: The truth about AI from the people 131, 2003.
building it”, book, Packt Publishing, 2018.
[38] S. El Bouanani, M. A. El Kiram, O. Achbarou and A. Outchakoucht,
[23] E. De Coninck et al., "Distributed neural networks for Internet of "Pervasive-Based Access Control Model for IoT Environments," in
Things: the Big-Little approach," in International Internet of Things IEEE Access, vol. 7, pp. 54575-54585, 2019. doi:
Summit, pp. 484-492, 2015. 10.1109/ACCESS.2019.2912975.
[24] N. McLaughlin et al., "Deep android malware detection," in Proceedings [39] M.R. Abdmeziem, D. Tandjaoui, I. Romdhani, “Architecting the
of the Seventh ACM on Conference on Data and Application Security Internet of Things: State of the Art”, Robots and Sensor Clouds. Studies
and Privacy, pp. 301-308, 2017. in Systems, Decision and Control, vol 36. Springer, 2015.
[25] P. Torres, C. Catania, S. Garcia, and C. G. Garino, "An analysis of [40] C. Bormann, M. Ersue, A. Keranen, "Terminology for Constrained-
Recurrent Neural Networks for Botnet detection behavior," in Biennial Node Networks", RFC 7228, May 2014.
Congress of Argentina, IEEE pp. 1-6, 2016.
[41] A. Ouaddah, A. Abou Elkalam and A. Ait Ouahman, “FairAccess: a new
[26] I. Goodfellow et al., "Generative adversarial nets," in Advances in Blockchain-based access control framework for the Internet of Things”,
neural information processing systems, pp. 2672-2680, 2014. Security and Communication Networks, pp. 1-22, 2017.
[27] R. E. Hiromoto, M. Haney, and A. Vakanski, "A secure architecture for [42] ISO/IEC JTC 1, Information technology, iso/iec 29146:2016, A
IoT with supply chain risk management," in 9 th IEEE International framework for access management, 2016.
Conference on Intelligent Data Acquisition and Advanced Computing
[43] eXtensible Access Control Markup Language (XACML) Version 3.0,
Systems: Technology and Applications (IDAACS), vol. 1, pp. 431-435,
2017. OASIS Standard, January 2013.
[28] M. A. Aref, S. K. Jayaweera, and S. Machuzak, "Multi-agent [44] A. Ameziane El Hassani et al., “Integrity-OrBAC: a new model to
preserve Critical Infrastructures integrity”, International Journal of
Reinforcement Learning Based Cognitive Anti-jamming," in Wireless
Communications and Networking Conference (WCNC), IEEE pp. 1-6, Information Security, Springer, vol. 14, Issue 4, pp 367–385, 2014.
2017. [45] A.E. Kalam, Y. Deswarte, “Multi-Orbac: a new access control model for
distributed, heterogeneous and collaborative systems,” in: 8th IEEE
[29] A Outchakoucht, ES Hamza, JP Leroy, “Dynamic access control policy
International Symposium on Systems and Information Security, p. 1,
based on blockchain and machine learning for the internet of things,” in
2006.
International Journal of Advanced Computer Science and Applications,
Vol. 8, No.7, 2017. [46] A. Abou El Kalam, Y. Deswarte, A. Baïna, M. Kaâniche, “PolyOrBAC:
a security framework for critical infrastructures,” Int. J. Crit. Infrastruct.
[30] NVIDIA Corporation, Jetson Nano Developer Kit, March 2019, [online]
Prot. pp. 154–169, 2009.
Available at: https://developer.nvidia.com/embedded/jetson-nano-
developer-kit [accessed 2 December 2019] [47] I. Bouij-Pasquier, A. A. El Kalam, A. A. Ouahman, and M. De
Montfort, “A Security Framework for Internet of Things,” Springer
[31] Google Brain, Deploy machine learning models on mobile and IoT
devices, 2019, [online] Available at: https://www.tensorflow.org/lite International Publishing, pp. 19–31, 2015.
[accessed 2 December 2019]. [48] M. Mahdavinejad, M. Rezvan, M. Barekatain, P. Adibi, P. Barnaghi, P.
Sheth, “Machine learning for Internet of Things data analysis: A
[32] C. Zhang, P. Patras, and H. Haddadi, "Deep Learning in Mobile and
Wireless Networking: A Survey," arXiv preprint arXiv:1803.04311, survey”. Digital Communications and Networks. 2017.
2018. [49] U. Nations, “World urbanization prospects: The 2014 revision,
highlights,” Department of economic and social affairs. Population
[33] R.S. Sandhu, “Role-based access control,” Adv. Comput. 46, pp. 237–
286, 1998. Division, United Nations, 2014.

340 | P a g e
www.ijacsa.thesai.org