ISSN 2278-3091

Urooj Yousuf Khan et al., International Journal of Advanced

Volume Trends in Computer
10, No.3, MayScience
- Juneand Engineering, 10(3), May - June 2021, 2611– 2622
2021
International Journal of Advanced Trends in Computer Science and Engineering
Available Online at http://www.warse.org/IJATCSE/static/pdf/file/ijatcse1541032021.pdf
https://doi.org/10.30534/ijatcse/2021/155032021

A Comparative Study of Various Machine Learning

Algorithms in Fog Computing
Urooj Yousuf Khan1, Muhammad Mansoor Alam2
1
Institute of Business Management, Pakistan, [email protected]
2
Institute of Business Management, Pakistan, [email protected]

 in the informed decision-making process and ultimate

ABSTRACT integration of the smaller systems into lager one [2]. Thus,
this system of intelligent Fog nodes and Cloud provides a
Internet of Things is the reality of the future. It encompasses a promising platform to realize IoT.
vast prototype. The applications are limitless and the research If envisioned completely, there are huge undiscovered areas
potential is huge. Several unanswered questions still stand in where IoT can bring about revolutionary changes[3]. These
the way. One of these challenging questions is the Network applications have a huge influence on human lives,
layer implementation of IoT. Cloud offers one perspective particularly in the health monitoring systems [4]. The paper
solution. Cloud services face the down side of connectivity is organized as follows: Section 2 covers the literature
and latency. Fog computing, as a complementing technology review. Section 3 covers the Fog paradigm in machine
for Cloud answers the issues of latency and connectivity in the learning. Section 4 covers the application areas where
Cloud. Being novel idea in itself, Fog computing offers machine learning can enhance the functionality of Fog
promising future growth. A major concept in Fog networks is Networks. Section 5 points to newer research directions and
its ability to learn and adopt to changing environment. This Section 6 concludes the paper. References are covered in the
ability of the Fog devices to self-learn and improve itself is the last section.
core of machine-learning. Hence, machine-learning
algorithms occupy central place in visualizing robust and 2. LITERATURE REVIEW
effective Fog networks. This paper presents a comparative
study of various machine-learning algorithms and their
applications in the Fog network. The paper also projects the 2.1 Cloud and Fog
undiscovered research paradigms and future directions Cloud computing has emerged as one of the most dominant
technologies. It is immensely capable of storing, analyzing
and managing huge amount of data. The focus of Cloud
Key words: Cloud Computing, Fog Computing, Internet of
computing is largely on centralized and remote data centers
Things, Machine Learning
[2]. This design comes with various shortcomings. These
1. INTRODUCTION include security, reliability and latency[2],[1]. Fog, as a
compliment to the existing Cloud network answers these
drawbacks[1]. It permits better resource utilization and lower
Fog as a probable platform for envisioning Internet of Things
(IoT) is a viable and sustainable idea. Visualization of being latency. Localized data processing enhances reliability and
connected everywhere is largely supported by the Cloud and security.
Fog computing[1] . Another way to realize IoT are the Fog is a term coined by Cisco[3],[1] in 2012. It is an
extension of the Cloud. It has many similar equivalents such
cyber-physical systems [2]. A prominent feature of any
as edge computing, mobile computing, grid computing or
smart cyber-physical network is its ability to learn from the
environment and the presence of feedback loops. These cloudlets[1]. Concisely, it is a localized data center with
feedback loops ensure that data analysis and study take computational capabilities. The aim is to provide control and
central stage [2]. In the perspective of IoT, this data analysis communication to the nodes in real time. It answers many
takes place at the Cloud. Data collection and immediate downsides of Cloud such as security, bandwidth
processing takes place at the Fog node. For long- term data consumption and latency. These challenges are expected as
the networks expand to 5G. The introduction of Cloud and
analysis and processing, a feedback system and a
mathematical model is needed. Embedding the ability of Fog can further envision and enhance IoT and artificial
self-learning in the Fog nodes would serve similar purpose. intelligence [3].
There are many arguments to support idea of smart nodes. In 2015, OpenFog Consortium was founded. It is a group of
These include deep knowledge from data inspection, support distinguished companies namely CISCO, Dell and academia
such as Princeton University. A reference architecture was
released in 2017. It is built upon guiding pillars of agility,

2611
Urooj Yousuf Khan et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 2611– 2622

autonomy, availability, hierarchy programmability, Incorporating machine-learning algorithms in Fog nodes to

serviceability, reliability, openness and security [5]. There make them more intelligent implies introduction of
are enormous potentials for the Fog as a complement to the “plasticity” within these nodes[6].
Cloud. These include great geographical distribution, task Many studies suggest that machine learning is resolving key
offloading and latency minimization[5]. issues within the Fog Computing architecture. These include
Utilizing the smart interconnected Fog-Cloud network to data aggregation, computational overhead and
realize IoT offers great advantages. This novel idea has far communication efficiency among many others [10]. This
reaching influence in human lives. A study proposed a health study suggests a method for data aggregation and analysis in
monitoring system based on IoT and machine learning [4]. Fog nodes using Machine Learning[10]. A
The study used three-tier architecture of IoT to monitor process-monitoring and prediction system based on Machine
significant parameters for heart diseases [2]. According to learning offers an efficient alternate to existing designs [11].
another study, an interesting analogy exists between human This system utilizes the Random forest algorithm to make
brain and these self-learning Fog networks[6]. Plasticity is scalable, high performance predictive model[11] . The proof
the capability of human brain to learn from its environment of concept provides a stronger foundation to the concept[11].
and adapt to the changes. Moreover, the parallel execution of The task of quickly learning from the surroundings and
concurrent tasks, selection of the best-fit algorithm for a utilizing the limited bandwidth is further elaborated with the
given set of constraints and job scheduling are tasks of the container designs [12]. The study utilizes machine learning
human brain. Human brain presents one of the most powerful algorithms such as Q-learning and deep learning to handle
and efficient processing systems[6]. large-scale migrations in Fog nodes [12]. The results
Technically, Fog as a communicating layer between Devices concluded that deploying machine learning techniques such
and the Cloud is a vastly supported idea. This multi layered as Deep learning significantly reduced the delay, power
architecture chiefly includes User-Equipment layer, consumption and migration cost.
Network layer and Data Processing layer [7]. There are Another significant area where machine learning is changing
different implementation details and challenges for each of the paradigm is the security and safety in Fog Computing
these layers. Fog nodes collect user data from system[13].The heterogeneous nature of the Fog nodes and
User-Equipment layer, process it locally at the Network their scalability leads to the problems of authentication and
Layer and route it to Cloud at the Data Processing Layer [7]. security within the local network. Deploying machine
Conferring to authors, this hierarchical network of User learning techniques for detecting anomalies and mitigating
devices, Fog nodes and Cloud has certain limitations[8]. attacks is discussed here [13].
Major challenges in this layered implementation for IoT Table 1 presents machine-learning solutions for numerous
include security, trust and privacy [8]. Inclusion of Fog IoT security threats[13].
nodes has majorly redefined the bounds. These Fog nodes Table 1: Machine learning solutions for numerous IoT security
now provide localized solutions with better bandwidth threats[13]
conservation, decentralized processing and real-time
analysis [8]. Use-Case Machine Learning Algorithm
Another example of including Fog nodes to enhance Anomaly Naïve Bayes
productivity exists in literature[9] . According to the study Detection ANN
conduct, one such scenario is discussed bysimulating the Finding Infrequent Support Vector Machine
data stream in Fog instead of routing it to the Cloud. The Data points Random Forest
results showed that the proposed model achieved 97% PCA
accuracy with an end-to-end delay of 180ms[9]. Naïve Bayes
KNN
2.2 Machine Learning Forecast Values Linear Regression
and Class Support vector Regression
Machine learning aims at improving the performance of the CART
computers by learning through familiarity. It enables FFNN
computers to improve their response to problems with Feature Extraction PCA
experience. This response can be applied to solving similar CCA
problems in the future. Classically, there are following three Malware SVM
major divisions of machine-learning algorithms: supervised, Detection Random Forest
unsupervised and reinforcement learning. Supervised Intrusion PCA
learning includes mapping. It charts input values to Detection Naïve Bayes
predefined output. These problems majorly fall under two Unearthing of K-Means
sub-classes: classification and regression. Unsupervised pattern DB-SCAN
learning implies finding data patterns in an unsupervised
manner. Reinforcement learning largely comprises of a The security threats in the smart city come naturally since it
series of feedback loops for reward and penalty. In this is a fully connected entity. These threats can have great
environment, the intelligent agent (learner) is an independent impact on human life. Another machine learning algorithm is
decision-maker. This decision maker takes decisions based applied to detect cyberattacks in the Smart City [14]. AD-IoT
on the awareness of the environment [2].
2612
Urooj Yousuf Khan et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 2611– 2622

is a proposed system for detecting anomalies in the system Table 2: Differences between Cloud and Fog [20]
using Machine learning. This model also deploys Random Feature Description
Forest technique for anomaly detection. The paper also Heterogeneity Located at the edge of the network with
suggests enhancements such as deploying Conventional rich, heterogenous end-user support
Neural Networks (CNN) to control traffic activities in the Capability Supports a wide range of industrial
city. This effective algorithm gains the highest classification applications deo to instant response
accuracy of 99.34% with lowest false positive rate[14]. capability
Another interesting case study for deploying machine Storage and Has its own computing, storage and
learning algorithms is found in BEMS [15]. In this paper, the Services networking services
authors propose a self-learning algorithm using Neural Operation Areas Operates locally single hop from a device
Networks. The smart buildings automatically control the to a fog node)
temperature, air quality and the humidity within the Platform Has a highly virtualized platform
premises. Deploying an effective machine learning
algorithm such as Neural Network has enabled examination Additional Portable deployment in terms of hardware
and higher accuracy rate.it has also led to reduced energy Features and software
consumption and increase of comfort level for the residents
[15]. Table 3: Computational aspects for Cloud and Fog[20]
An application of task offloading in Fog Networks using Aspects Fog Computing vs Cloud Computing
classification and regression tree algorithm is also found in Resources Relatively fewer computing resources in
literature[16]. This study suggests that usage of classification Fog
algorithm has reduced power consumption of CPU, RAM Capability Both can process data generated from a
and Wi-Fi by 44%,71% and 55% respectively[16] . Usage of diverse set of resources
machine learning in vehicular computing and connected Distribution Based on geographical location, both
vehicles is another significant paradigm [17]. The effective Strategies can distribute data densely or sparsely
usage of machine learning methods aims at improving the Connectivity Both support machine-to-machine
mobility and connectivity. The loss of connectivity when communication and wireless
transitioning from one fog node to the other can be connectivity
minimized using recurrent neural networks (RNN)[17] . The Flexibility Fog can be installed on low
selection of optimal node and its effective cost were specification devices such as switches
calculated given the time and location. This optimal node and Internet protocol cameras
minimized the handover cost associated with loss of Usability One of the primary uses is currently for
connectivity [17]. It is expected that more than 9 billion mobile and portable devices
people will live on the planet by 2050 and about two thirds of
them will live in cities [18]. An important implementation of Table 3 above provides a clear picture for different
the Fog computing lies in Smart cities. Fog computing computational aspects of Cloud and Fog. There is a separate
enables huge data generated by the nodes to be processed paradigm for Fog devices. These Fog devices can be further
locally. Incorporating machine intelligence in the system classified as mission-critical, real-time, interactive,
comes with challenges such as the capability to handle such conversational, streaming, CPU-bound and Best effort[21].
diverse data resources and physical constraints imposed by Each class clearly specifies the significance of Fog
device limitations [19]. computing in its own terms[21].
Table 4 below signifies the architectural differences between
2.3 Fog Architecture Cloud and Fog.
The precise location and platform selection of the Fog nodes Table 4: Differences between Fog and Cloud [5]
depend on the application scenario and required number of Features Cloud Fog Computing
hosts. There are key differences between Cloud and Fog. The Computing
major difference lies in utilization of resources. Fog is a Architecture Centralized Decentralized
resource-constrained environment with stringent time Access Fixed and Wireless Mainly Wireless
requirement. Cloud, on the other hand, virtually has Latency High Very Low
unlimited storage and processing capability. Moreover, Mobility Limited Support Supported
peer-to-peer communication is much faster in Cloud due to Operation Operated by large Often operated by
high-speed connectivity. It does not happen in Fog where the companies small companies
networks vary quickly and communication is majorly Server Nodes Few Very Large
wireless. Vertical communication, however, takes another Number
form. In Cloud, this communication is slower while Server nodes The Internet The local network
accessing remote devices. Cloud computing is a stable, location
market-oriented paradigm while Fog is still an emerging Size Very large data Large number of
concept [1]. centers small fog nodes
Table 2 below identifies the key differences between Cloud Scalability Average High
and Fog. Vulnerability High probability Very low probability

2613
Urooj Yousuf Khan et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 2611– 2622

Figure 1 below shows the Fog architecture and its different Figure 2 further explains Cloud-to-Fog continuum. The Fog
classes. Evident from the figure is the placement and nodes, placed in user proximity provide continuity along the
application of the Fog nodes. The end devices could be any Things-to-Cloud. Optimal placement of nodes in the Fog
user-defined input scheme. The Fog layers 1 and 2 provide the network also promotes job scheduling and deploys digital
layered implementation for varied types of requirements. resources at the Edge of the network. It provides a local pool
of resources for immediate and optimal utilization[24].
Figure 3 below clearly depicts the placement of Fog Nodes
and their functionality. A typical Fog network is identified
by a large number of heterogeneous nodes, placed in close
proximity to the user or test case. These nodes vary largely
from each other in terms of data streams, protocols and
networking capabilities. These may include PCs, routers,
switches, proxy servers, mobile devices or source inhibited
System on Chip (SoC). These heterogeneous nodes
communicate with each other via Ethernet or Wi-Fi based
networks. These are resource stringent nodes that provide
more options for wireless communications. In other words,
the great diversity among nodes requires flexible and elastic
framework, with real time response. All thesefeatures are
intrinsic to Fog. Thus, Fog is becoming a pivotal point in
future networks [25].

Figure1: Fog Computing Architecture and Class of

Service [21] Figure 3: Architecture of a Fog Network [25]
Another example of layered approach for implementing
CPU-bound, real-time Fog paradigm exists in IoTPulse
where Fog devices play major role in health monitoring Figure 3: Architecture of Fog Networks[25]
sector[22] . The system comprised of four layers: sensing Fog device acts as an intermediary layer that “sits” between
layer, network layer, servicing layer and interfacing layer. Cloud and the heterogeneous ends. The challenges for the
The sensing layer gathered the patient parameters while the Fog nodes include dynamic nature of the input data, the
network and servicing layer utilized the Fog networks for varied data syntax and semantics and minimum latency time.
decision making. Classification of Fog nodes into All these scenarios are clearly depicted in the figure above.
heterogenous cellular networks further promotes latency Moreover, the significance of machine learning can be
reduction to achieve better Quality of Service (QoS) [23]. visualized clearly. A robust self-learning mechanism would
enhance the productivity and efficiency of these networks.
Figure 3 complete architecture of Fog. It explains the
communication between IoT, Fog and Cloud. The layered
architecture of Fog is clearly visible. The Fog layer provides
hierarchical continuum along Things-to-Cloud for limited,
time-bound data storage. This heterogeneous arrangement
provides varied services for different user requirements.
These include scenarios such as connected cars, smart grids,
intelligent buildings etc. These nodes are essentially a huge
collection of geographically smart devices such as routers,
switches or even an RFID enabled device [11],[5]. These
nodes receive and send data to nearest available nodes
horizontally in their network [5].
All figures above identify yet another interesting scenario. It
is highlighting the IoT paradigm. The topological
arrangement of the Fog devices and the Cloud networks is
Figure 2: Fog Computing Infrastructure[24]
also enhancing IoT visualization. Here the varied devices
2614
Urooj Yousuf Khan et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 2611– 2622

and different physical objects interact and communicate with are capable to represent complex relationships among the
each other using a common communication platform. elements of the input set [2]. Another variation of the ANN is
Initially the concept was to utilize the IPV4 addressing. in multilayer perceptron (MLP) that deploys nonparametric
However, later it was upgraded to IPV6, thereby making it estimator. Convolutional Neural Networks (CNN) works on
possible for every device in the planet to be uniquely the principal of translational invariance. It implies that the
identifiable[2]. value of the output function is unchanged as the input function
The scope of Fog is not only limited to resolving the translates. In other words, if the input translates, the CNN will
bottlenecks faced by Cloud. This paradigm extends far still be able to detect which class the input belongs. CNN
beyond that. It can form core component in emerging IoT, majorly comprises of two parts: feature extractor and a
embedded AI and 5G among many others. A considerable trainable classifier such as SVM. The feature extractor reads
advantage of Fog Computing also lies in the fact that it and extracts the features and characteristics from the raw data.
possesses pool of resources that can be utilized as per the It comprises of multiple similar stages and layers.
user requirements. Moreover, being near to the end users Classification is performed by the second part: classifier.
promotes greater cognition and swift data analysis. All these Another popular class of ANN is Recurrent Neural Networks
features enhance efficiency. Additionally, bandwidth (RNN). These neural networks have additionally employed
conservation is ensured, as data does not travel to a central the quality of recurrence due to which they are especially
node or server[3] . important in data analytics and streaming. Deep neural
On a detailed note, the Fog architecture is denser and more networks take a step ahead. They utilize stacks of multi-layer
versatile compared to Cloud. Fog computing embeds IoT neurons to efficiently represent complex relationships among
and business insights gathered from nodes in real-time. It data elements [2]. Swift, parallel and distributed computations
provides distributed and embedded intelligence. This results are key features of Neural Networks [22].
in scalability and heterogeneity in Fog[10]. Fog servers are Figure 4 below captures the neural networks briefly.
dense collection of sensors and are geographically located at
closest proximity to the users. These features enhance lower
latency, higher mobility and lower deployment cost.

2.4 Machine Learning Associated Algorithms

Below are some of the widely used machine learning
algorithms.

A.Decision Trees and Random Forests

INPUT (In) PROCESSING (Pn) OUTPUT (On)
These algorithms belong to the class of algorithms that follow
ranking. The ranking method is implemented through
Figure 4. Multilayer Perceptron Neuron Network [22]
tree-like structure. Nodes constitute decision points and
leaves constitute predicted outputs. The leaf topology is based C. Support Vector Machines (SVMs)
on specific, predefined attribute. A major advantage of these These machines work under the principle of boundary
algorithm is their understandability and readability [2]. There consideration.
are two sub branches: classification and regression. These supervised learning models segment data by drawing
A group of random trees forms the forest. These random trees
a hyperplane with maximum margin between data points of
are developed through bagging. Bagging process divides the
each class. These machines are scalable and can be cascaded
dataset of X samples into y smaller, unique datasets. Each of to work on multilevel classification. Due to their flexibility;
these y datasets trains a different decision tree. these SVMs have been extended to solve regression problems
The heuristic followed here is as follows: ‘‘combination of the as well [2].Moreover, there are notable applications such as
resulting models should be more robust to overfitting [2].” Long-Bone X-rays[26].
The output produced by each tree is combined together by a
variety of rules. D. Bayesian Networks
These networks utilize conditional probability to model the
B. Artificial Neural Networks system behavior [22]. These networks apply statistical
Neural networks are one of the most efficient machine techniques to describe probabilityrelationship among
learning algorithms that aim to simulate human brain and different elements. Bayesian networks indicate the probability
underlying communication network [7]. It works by of occurrence of an event given that the prerequisite event has
modelling nervous system in living beings. The functional
already taken place [22]. The method works on directed
unit of the nervous system is a single neuron. The nervous
acyclic graphs. Nodes represent the random variables and
system comprises of multiple layers of interconnected directed vertices indicate the influences [2].
neurons. These pseudo-neurons assist in conducting tasks Notable variations of Bayesian Networks include Naïve
such as data mining, pattern recognition and classification [7]. Bayes and Hidden Markov Model. This machine learning
The input layer moves to the next layer using a non-weighted algorithm is ideal in modelling uncertain domains [2]. It can
combination to generate transformation values. The process be used in supervised and unsupervised learning domains
continues until the output layer is reached. These algorithms successfully [2].

2615
Urooj Yousuf Khan et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 2611– 2622

E. Evolutionary Computations and Genetic Algorithms G. Clustering

This family of algorithms for machine learning revolve
around the principles of biological evolution. The guiding This family of algorithms involve partitioning and grouping
principal is ‘‘survival of the fittest [2]”. The incremental elements. The input data set is divided into smaller datasets
learning approach here involves trial-and-error guided by the called clusters. The goal is usually to maximize the
fitness function. The fitness function measures the goodness similarities among the objects within the same cluster and to
or reliability of the output. The fitness function involves minimize the distance between the center of the cluster and
working on the principles of natural selection, mutation and individual object. Moreover, the classification of data also
cross-overs [2]. This algorithm is applicable in supervised, involves finding the difference among various clusters. A
unsupervised and reinforcement learning [2]. popular algorithm of this classis K-Means clustering where k
denotes the number of clusters [2].
F. Bees-Swarm
This algorithm is also biologically inspired. It imitates two H. Self-organizing Map (SOM)
major behaviors in a bee’s life: marriage and food foraging
[24]. The algorithm begins by a random search of the space to This algorithm is best suited for topographical distribution.
form initial population. As with any other genetic algorithm, This method is largely applied to clustering problems. SOM
this initial population is evaluated by calling the cost function. implies using finite data models to represent data distribution.
After the primary filtration, queen is selected as the fittest A grid is used to represent the arrangement of nodes. The
among N bees. The following ‘D’ fittest bees are chosen as the adjacent nodes in the grid represent similar models whereas
drones while the remaining ‘W’ bees constitute the worker less similar models are located far away from each other [2].
group. The values of ‘N’,’D’ and ‘W’ are predefined by the Q learning: It is an incremental, reinforcement algorithm. The
user. After the selection, the two important events, namely agents learn how to act optimally in controlled Markovian
marriage ad food foraging begin. Marriage or reproduction environments. It involves techniques for dynamic
follows typical steps of genetic algorithm namely mutation programming thereby utilizing limited computational space.
and crossover to propagate the desired qualities to the next By successively improving the quality of specified actions,
generation. Workers carry out the task of food foraging. It the algorithm reaches the desired results [2].
involves finding the food source and recruiting several other
workers to collect it. More workers will be recruited if the I.KNN
location is a desirable spot (flowers). Moreover; the best
worker is selected for every region to be kept alive for the next This specialized, lazy algorithm works by finding the nearest
population. The cost function is used to select the best worker possible class for elements. It works by finding the Euclidian
[9]. The entire algorithm for Bees swarm is given in Figure 5 distance between two points. One of these points is the source
below. According to the figure, the life cycle of the bees is or the element that needs to be classified. The second point is
iterative. It majorly comprises of bees’ selection for queen and selected from the given dataset and previous classifications.
food search. The genetic algorithm clearly displays the major The algorithm calculates the Euclidian distance of the given
processes of the life cycle. The cross over and mutation are data points to predict where an unknown point must lie [22].
representing the reproductive cycle. The food search is carried
out by the worker bees. The entire life cycle and behavior of
the bees is clearly depicted in Figure 5. 3. MACHINE LEARNINGIN FOG COMPUTING

Fog is the essential player in the next generation of

latency-free network. Introducing intelligence in the Fog
environment can bring immense scalability to the network.
Introducing intelligence in Fog implies integration of
learning capabilities into the Fog network[27]. Context
awareness, real-time response and resource optimization are
just few of the benefits. Two broad categories of this
intelligence could be device-driven approach and
human-driven approach [28].
Device-driven Approach: This approach implies equipping
devices with smarter and more robust devices. It could
include computing, sensing, storage and communication.
This approach is more dominant in Fog networks where
sensors are varied and their data is diverse.
Human-driven Approach: This approach implies embedding
human intelligence into machines, thereby improving
performance. The key idea is to train the network through
human behavior. Considerable works have been done in this
regard. It involves scheduling the household appliances or
Figure 5. Bees Life Algorithm pseudo-code [24]
enhancing the quality of life [28].
2616
Urooj Yousuf Khan et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 2611– 2622

The input from the edge devices moves to the data centers for 3.4 Optimization
immediate storage and capturing. The core of data analysis
relies heavily on Machine learning and its numerous Enabling intelligence in nodes can be beneficial in reducing
algorithms. To embed thinking into a machine points to the design complexity of a given scenario. Optimal Resource
traditional and modern approaches of machine learning. utilization prevents conflicts. It provides an acceptable
These may include brute force and greedy K-nearest tradeoff between accuracy and time complexity [12],[28].
neighbor, neural networks, support vector machines or
nature-inspired genetic algorithms. Moreover, advanced 4. APPLICATIONS OF M ACHINE LEARNING ALGORITHMS IN
hardware such as GPU or embedded platforms can locally FOG COMPUTING
perform data-analytics. Cloud is then specialized for specific
tasks such as data training for huge datasets[20]. Few major areas where highest impact is expected are as
follows:
Advantages: 4.1 Health
A considerable point here to understand many potential
benefits that come along this architecture. Efficient energy The impact of machine learning and fog computing is huge
consumption is one such benefit. Fog devices are in the health sector. Various applications are being proposed
resource-constrained and applications such as image and in this regard. One such application is IoTPulse. This project
voice processing are energy intensive. Catering the predicts the probability of a person being a drug addict. It
requirements of these devices over the edge of the network utilizes machine-learning algorithms deployed in Fog
could mean task offloading, thereby conserving the precious Networks [22].
bandwidth. Scheduling job sequence in this scenario is still a Another study revolves around the detection of a cardiac
challenge. arrest using the wearable sensors[4]. The introduction of an
Notable advantages are given below[28]: efficient machine learning algorithm namely ROC analysis
has helped in identifying most significant clinical parameters
3.1 Reduced Latency to get cardiac arrest [4].

Introducing Fog devices concludes that traffic on the 4.2 Smart Grid
network would be entertained locally. Some IoT devices
such as real-time health monitoring system are time and An application of Fog computing and machine learning lies
resource intensive. If these Fog nodes are embedded with in smart grids and energy management. A grid is an
intelligent decision-making systems the capability of the arrangement of various networks ranging from micro grids to
system and performance could increase many folds[28]. international levels. Its applications include power
generation and distribution systems, control systems for
3.2 Bandwidth Conservation balancing and energy production. Due to highly dynamic and
volatile nature of this application, an efficient fault tolerant
With localized preprocessing and storage of data near the mechanism is also needed. The inclusion of self-learning in
edge of the network, huge volume of data can be addressed this arrangement would enhance the error-detection and
locally. This essentially frees up the constrained bandwidth correction capabilities of the system[2].
since the amount of user data has increased exponentially
over the years. Smart machine learning algorithms can 4.3 Security
provide the much-needed solution[28].
Offloading the computational tasks adds another strength to The Cloud services come with the issues of latency and
the Network. It provides means for preserving the battery life connectivity. Moreover, due to performance requirements,
in pervasive devices. Specific machine learning algorithm the real-time implementation of effective security policy is
namely Neuro-fuzzy logic using Neural networks are highly difficult to achieve. In this perspective, the Fog devices can
applicable in this perspective[5]. play a major role in offloading the traffic, conserving
bandwidth and enforcing security. The Fog architecture,
3.3 Improved Network Availability boasting horizontal connectivity has dual effects: it reduces
storage space and latency issues. Fog nodes are efficient at
Whenever a device moves from one region to another in Fog handling widespread and massive information. The detection
or Cloud environment, it experiences disruption and of attacks can be further controlled using the Machine
connectivity glitch. Intelligent machine learning algorithms learning algorithms. Moreover, the highly parallel nature of
should enable self-organization and co-operation to ensure the Fog eases the mathematical modelling and
smooth and uninterrupted services. calculations[3],[13].
Security and Privacy: Self-organizing and learning also
provides robust protection against malicious attacks. It is 4.4 Data Aggregation
because network is able to identify and analyze data locally
before propagating it to the network. The synthesis, analysis and study of data aggregation
produced an array of innovative benefits. But the issues
related to data piracy and privacy associated with sensitive,
2617
Urooj Yousuf Khan et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 2611– 2622

unhandled user data is a key concern[6].Intrusion Detection operations[27].The tables 5-11 below signify the application
systems using Fog nodes are widely under development and of each machine learning algorithm.
research. Intrusion detection systems form the keystone of
security systems. The resource constrained features of the Table 5: Random Forest[2]
Fog require specialized machine learning algorithms[7].
Domain Functional Category Task
4.5 CPU Job Scheduling
Smart Grid Forecasting/Prediction/Regre demand side
Efficient scheduling of the jobs that need to be catered by the
ssion load
Fog environment is a research question. An experiment
forecasting/Pri
conducted showed that machine learning algorithms enhance
ce forecasting
CPU scheduling problem. This is an ideal requirement in the
Anomaly/Fault Detection Power record
Fog networks. It enhances the operational capacity of the
faults
network and upgrades its performance. A significant
algorithm here is Bee swarm algorithm that captures the Transport Pattern Behavior/Even
essence of Bees life. This particular algorithm finds the Recognition/Clustering t Recognition
optimal tradeoff between CPU execution time and allocated Manufacturing Anomaly/Fault Detection Tooling wear/
memory. An enhanced version of this algorithm can be Errors
utilized in network optimization [5]. detection
.
4.6 Smart Cities
Table 6:ANN[2]
Smart city is also termed as “sustainable city” or
self-dependent city. It implies a real-world scenario of Smart Grid Forecasting/Prediction/ Electrical Power
enabling a city to be smart enough to identify its own needs. Regression prediction, load
Geo-distribution is a major characteristic for the big data forecasting
generated by smart cities. This huge data can only be handled Transport Pattern Recognition/ Behavior/Event
using a massive collection of nodes such as in Fog. The data Clustering Recognition
analysis for such varied and diverse data can be achieved Forecasting/Prediction/ traffic flow features
affectively by using an effective machine learning Regression road side CO and
algorithm[18]. NO2
concentrations
4.7 Connected Vehicles estimation
travel time prediction
Fog computing along with machine learning is also Classification obstacle detection and
incorporated in the smart vehicles. Efficient and effective recognition
machine learning algorithms to minimize handover Image Processing
optimization is possible [17]. Vehicular Fog simulator and Manufacturing Forecasting/Prediction/ Energy Consumption/
Recurrent neural network can be used to minimize the Regression/ Process
latency and handover delay [17]. optimization parameters
optimization
4.8 Network Connectivity and Detection

The next generation of Internet revolves around highest Table 7:SVM[2]

internet connectivity and broadband i.e. 5G. These networks
are extremely dynamic and require great bandwidth. Due to Forecasting/ Price Prediction
Smart Grid
the heterogeneous nature of these platforms, many Prediction/Regression Electrical Power
challenges for network planning, deployment and operations
prediction,
are expected. Concurrently, the huge volumes of data being
Anomaly/Fault Non Technical
generated are a big question for scalability. Such challenges
Detection Loss detection
require dynamic resource provisioning and intelligent
Blackout Warning
approaches. Fog offers great exploration of these resources.
The dynamic and complex nature of future networks call for Power Line
an intelligent and flexible structure[26]. This structure is Attacks
clearly embedded in the Fog. Numerous machine-learning Transport Classification Unintentional
algorithms implemented here is include linear programming vehicle lane
(LP), ant colony optimization (ACO) and genetic algorithm departure
(GA) [12]. Natural language processing techniques are being prediction
used for enhancing the efficiency of wireless network Obstacles
classification

2618
Urooj Yousuf Khan et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 2611– 2622

Pattern Recognition/ Behavior/Event optimal machine learning algorithm to facilitate mobility

Clustering Recognition and computational offloading in a specific application
Anomaly/Fault Mechanism presents potential research area[29].
Detection Failure Considering the heterogenous nature of multiple devices,
Forecasting/Prediction/ Travel time the choice of optimal machine learning algorithm to
Regression prediction facilitate mobility and computational offloading in a
Manufacturing Forecasting/Prediction/ Machine specific application presents potential research area [29].
Regression Maintenance Similar research gap exits in fog resource allocation in
Design / Configuration Feature Design; blockchain networks [30]. An optimal machine learning
Production algorithm that maximizes resource allocation and the
Processing arrival pattern of requests is an unanswered question [30].
Anomaly/Fault Quality Control Another unexplored paradigm is the handling of vast
Detection amount of data in a Wi-Fi environment both in delayed and
Smart Home Pattern Recognition/ Activity non-delayed transmissions [31]. The optimization of
Clustering recognition Genetic algorithm to best suit varied architecture for
differentiated data is still to be answered [31].
Table 8: Bayesian Network[2] The debate of optimal resource allocation and management
Transport Classification Event and behavior for energy conservation is still open [32]. The quest for
detect ideal design that promises best resource allocation for
energy conservation in contrary to existing designs is still to
Smart Grid Anomaly/Fault Detection Non technical losses
conclude[32].
and fault
detection Security and access control to meet different stakeholders is
a research area in itself. The overwhelming amount of data
Manufacturing Anomaly/Fault Detection Fault detection in the
and its variety, its storage and access methods and the
production line
choice of optimal machine learning algorithm to ensure
Forecasting/Prediction/Re Tool wear
optimal performance are still unanswered [33].The fine
gression prediction/Energy
grained control provided by the Fog network and best-fit
consumption
machine learning algorithm are major design concerns
prediction
[33].
Another research area is the load balancing in a dynamic
Table 9: Evolutionary Computing[2] Fog environment [34]. The challenge of load balancing
Smart Grid Optimization/ Short Term load utilizing Q-learning with greedy algorithm can be enhanced
Forecasting/Prediction forecasting further [34]. The optimization techniques to achieve lower
processing delay and less machine dependency can be
Table 10: Q learning based algorithm[2] studied further[34].
Smart Grid Optimizatio Aided Optimal Customer The Fog to Cloud communication is another research gap.
n Decisions for an The design techniques to minimize latency and speedup the
Interactive Smart Grid Fog-Cloud communication are being researched [35].This
Transport Optimizatio the road latent cost is the major architectural design for Internet of Things
n [35].The ideal machine learning algorithm for data caching
at fog nodes would reduce network latency. Machine
Table11: KNN[2] learning techniques to predict future needs are majorly
Smart Grid Optimization/ Short Term load forecasting unanswered[35].
Forecasting/Prediction As previously stated, one way to view Fog Networks is to
envision it as a Cyber-physical system. There are numerous
applications of this visualization [36].They include smart
5. FUTURE RESEARCH buildings, power grids and connected vehicles. The design
concerns for individual realization are largely developed
Machine learning algorithms play a major role in enhancing through machine learning models. Yet the optimal choice of a
the performance and efficiency of the system. The paper specific machine learning algorithm based on a particular
discussed various applications of Fog computing using scenario is still unanswered.The design aspects such as energy
machine learning. The various algorithms for machine efficiency, computational offloading and reduced latency are
learning were also discussed. Enabling intelligence in fog increasingly important research areas [36].
computing using machine learning offers huge ground for Scalability and sustainability in a Fog architecture is also
research and implementation[28] . one among many other design concerns. Due to horizontal,
An important research question is to address device wide-spread,layered architecture of Fog nodes, network
heterogeneity. authentication and resource provisioning are an open challenge
An example of computational offloading utilizing machine [37].The common model that caters device heterogeneity is
learning further covers research gap [29]. Considering the lacking.The maximum number of devices that can be
heterogenous nature of multiple devices, the choice of connected to a Fog Nodedepends on aspecific scenario.The
2619
Urooj Yousuf Khan et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 2611– 2622

Quality of Service (QoS) can be seriously compromised if the Pakistan, 2018 .

number of connected devices exceed the threshold.Moreover, 4. P. . M. Kumar and U. D. Gandhi, "A novel three-tier
Fog nodes are acting as local data processing units.Hence the Internet of Things architecture with machine learning
need for change the functionality as per the scenario needs to algorithm for early detection of heart diseases,"
be embedded in the Fog nodes.Thus,Fog nodes virtualization is Computers & Electrical Engineering, vol. Volume 65, pp.
another open research question[37]. Pages 222-235, 2018.
Task scheduling in a CPS is also an important research
direction [38] .Considering Fog as an expected framework for 5. F. Haouari, . R. Faraj and J. M. AlJa'am, "Fog
CPS, task scheduling to fulfil QoS is an unanswered question. Computing Potentials, Applications, and Challenges,"
An implementation of task scheduling utilizing moth’s in International Conference on Computer and
behavior offers a promising solution[38]. The future direction Applications (ICCA), Beirut, 2018.
points to integration of task-scheduling with load 6. D. Kimovski, H. Ijaz, N. Saurabh and R. Prodan,
balancing[38]. "Adaptive Nature-Inspired Fog Architecture," in IEEE
The optimal design for load balancing and task scheduling is 2nd International Conference on Fog and Edge
greatly applicable in Vehicle Adhoc Networks (VANETs). Computing (ICFEC), Washington, DC, 2018.
The load balancing approach focuses on finding the optimal 7. X. An, X. Zhou, X. Lü, F. Lin and L. Yang, "Sample
path from the source to the destination, better network Selected Extreme Learning Machine Based Intrusion
partitions and higher transmission rates. Intelligent decisions Detection in Fog Computing and MEC," Wireless
can be made for task scheduling if machine learning Communications and Mobile Computing, vol. vol. 2018,
algorithms are properly utilized. In such a scenario, the pp. 1-10, 2018.
consumer satisfaction and data privacy take the central stage.
8. A. A. Alli and M. . M. Alam, "SecOFF-FCIoT: Machine
Efficient buffer design, scalability and security are sizable
learning based secure offloading in Fog-Cloud of
research gaps[39].
things for smart city applications," Internet of Things,
vol. Volume 7, September 2019 2019.
6.CONCLUSION 9. M. Lavassani, S. Forsström, U. Jennehag and T. Zhang,
"Combining Fog Computing with Sensor Mote
Machine Learning for Industrial IoT," Sensors, Vols.
The applications of machine learning algorithms in Fog vol. 18, no. 5, no. vol. 18, no. 5, p. 1532, May 2018.
computing are numerous. there exist major applications where
10. M. Yang, T. Zhu, . B. Liu, . Y. Xiang and W. Zhou,
this fusion can be game changing. This design paradigm can
"Machine Learning Differential Privacy With
be effective in designing the optimal solution for Internet of
Things. Deploying traditional machine learning techniques to Multifunctional Aggregation in a Fog Computing
make fog nodes intelligent can greatly enhance the IoT Architecture," IEEE Access, vol. Volume: 6, pp. 17119 -
17129, 20 March 2018.
networks [40].Optimal resource provisioning, security, task
scheduling and many other important design parameters can 11. D. Wu, S. Liu, L. Zhang, J. Terpenny, R. X. Gao, T.
be addressed by integrating machine learning with Fog Kurfess and J. A. Guzzo, "A fog-computing based
devices [40].Minimal computational latency, real-time framework for process monitoring and prognsis in
analysis and addressing device heterogeneity in Fog Networks cyber-manufacturing," Journal of Manufacturing
can be addressed using machine learning algorithms[41]. Systems, Vols. Volume 43,, no. Part 1,, pp. Pages 25-34,,
2017.
12. Z. Tang, . X. Zhou, F. Zhang, W. Jia and W. Zhao,
"Migration Modeling and Learning Algorithms for
REFERENCES Containers in Fog Computing," IEEE Transactions on
Services Computing, Vols. vol. 12, no. 05, no. vol. 12, no.
05, pp. pp. 712-725, 2019.
1. U. Y. Khan and T. R. Soomro, "Envisioning Internetnet
of Things using Fog Computing," International Journal 13. M. Moh and . R. Raju, "Machine Learning Techniques
of Advanced Computer Science and Applications, vol. 9, for Security of Internet of Things (IoT) and Fog
no. 1, pp. 441-448, 2018. Computing Systems," in 2018 International Conference
on High Performance Computing & Simulation (HPCS),
2. X. Fei, N. Shah, N. Verba, K.-M. Chao, V.
Orleans, 2018.
Sanchez-Anguix, J. Lewandowski, A. James and Z.
Usman, "CPS data streams analytics based on machine 14. I. Alrashdi, A. Alqazzaz, E. Aloufi, R. Alharthi, M. Zohdy
learning for Cloud and Fog Computing: A survey," and H. Ming, "AD-IoT: Anomaly Detection of IoT
Future Generation Computer Systems, vol. Volume 90, Cyberattacks in Smart City Using Machine
pp. 435-450, 2019. Learning," in IEEE 9th Annual Computing and
Communication Workshop and Conference (CCWC)IEEE
3. U. Y. K. a. T. R. Soomro, "Applications of IoT: Mobile
9th Annual Computing and Communication Workshop
Edge Computing Perspectives," in 12th International
and Conference (CCWC), Las Vegas, NV, USA,, 2019.
Conference on Mathematics, Actuarial Science,
Computer Science and Statistics (MACS), Karachi, 15. Z. Shen, . K. Yokota, J. Jin, . A. Tagami and T. Higashino,

2620
Urooj Yousuf Khan et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 2611– 2622

"In-network Self-Learning Algorithms for BEMS Amani Al-Ghraibah et al., International Journal of
Through a Collaborative Fog Platform," in IEEE 32nd Advanced Trends in Computer Science and Engineering,
International Conference on Advanced Information vol. 10, no. 3, pp. 1494-1500, 12 6 2021.
Networking and Applications (AINA),, Krakow, 2018. 27. T. . E. Bogale, X. Wang and L. B. Le, "Machine
16. . D. Rahbari and . M. Nickray, "Task offloading in Intelligence Techniques for Next-Generation
mobile fog computing by classification and regression Context-Aware Wireless Networks," ITU Special Issue
tree.," Peer-to-Peer Networking and Applications, vol. N.1 The impact of Artificial Intelligence (AI) on
13, p. 104–122, 2019. communication networks and services, 12 Jan 2018.
17. S. Memon and M. Maheswaran, "Using machine 28. Q. . D. La, M. V. Ngo, T. Q. Dinh, T. Q.S. Quek and H.
learning for handover optimization in vehicular fog Shin, "Enabling intelligence in fog computing to
computing," in Proceedings of the 34th ACM/SIGAPP achieve energy and latency reduction," Digital
Symposium on Applied Computing, Limassol Cyprus, Communications and Networks, Vols. Volume 5, Issue 1,
April 2019. no. Volume 5, Issue 1, pp. 3-9, 2019.
18. B. Tang , Z. Chen, G. Hefferman, P. Shuyi , T. Wei, H. He 29. S. Zhou, . S. Y, Z. Jiang and Z. Niu, ""Exploiting Moving
and Q. Yang, "Incorporating Intelligence in Fog Intelligence: Delay-Optimized Computation
Computing for Big Data Analysis in Smart Cities," Offloading in Vehicular Fog Networks," IEEE
IEEE Transactions on Industrial Informatics, Vols. vol. Communications Magazine, Vols. vol. 57, no. 5, no. vol.
13, no. 5, no. vol. 13, no. 5, pp. pp. 2140-2150, Oct. 2017. 57, no. 5, pp. pp. 49-55, May 2019.
19. Y. Tu, Y. Ruan, S. Wagle, C. G. Brinton and C. 30. N. C. Luong, Y. Jiao, P. Wang, D. Niyato, D. I. Kim and
Joe-Wong, "Network-Aware Optimization of Z. Han, "A Machine-Learning-Based Auction for
Distributed Learning for Fog Computing," in IEEE Resource Trading in Fog Computing," IEEE
INFOCOM 2020 - IEEE Conference on Computer Communications Magazine, Vols. vol. 58, no. 3, no. vol.
Communications, Toronto, ON, Canada, 2020. 58, no. 3, pp. pp. 82-88, March 2020.
20. K. H. Abdulkareem, . M. . A. Mohammed, S. 31. George-Eduard Zaharia, T.-A.-I. Şoşea, R.-I. Ciobanu and
Gunasekaran, A. Mutlag, . A. M. Salama and N. Salih Ali, C. Dobre, "Machine learning-Based traffic offloading
""A Review of Fog Computing and Machine Learning: in fog networks," Simulation Modelling Practice and
Concepts, Applications, Challenges, and Open Theory,, vol. Volume 101, 2020,.
Issues,," IEEE Access, vol. vol. 7, no. vol. 7, pp. pp. 32. W. Shafik, , . s. M. Matinkhah and S. Mamman,
153123-153140, 2019. "Network Resource Management Drives Machine
21. J. C. Guevara, R. . d. S. Torres and N. L. d. Fonseca, "On Learning: A Survey and Future Research Direction,"
the classification of fog computing applications: A Journal of Communications Technology, Electronics and
machine learning perspective," Journal of Network and Computer Science, vol. Issue 30, no. Issue 30, 2020.
Computer Applications, vol. Volume 159, no. Volume 33 A. S. M. Kayes, R. Kalaria, I. H. S. Sarker, I. H. Sarker,
159, pp. 1-10, 2020. M. S. Islam and P. A. Watters, "A Survey of
22. A. Dhillon, A. Singh, H. Vohra, C. Ellis, B. Varghese and Context-Aware Access Control Mechanisms for Cloud
S. Gil, "IoTPulse: machine learning-based enterprise and Fog Networks: Taxonomy and Open Research
health information system to predict alcohol addiction Issues," Sensors, vol. vol. 20, no. 9, p. 2464, Apr. 2020.
in Punjab (India) using IoT and fog computing," 34. J. Baek, G. Kaddoum, S. Garg, K. Kaur and V. Gravel,
Enterprise Information Systems, pp. 1-34, 21 Sep 2020. "Managing Fog Networks using Reinforcement
23. E. Balevi, and . R. Gitlin, "Unsupervised machine Learning Based Load Balancing Algorithm," in 2019
learning in 5G networks for low latency IEEE Wireless Communications and Networking
communications," in 2017 IEEE 36th International Conference (WCNC), Marrakesh, Morocco, 2019.
Performance Computing and Communications 35. RIYA, . G. NITIN and Q. XIN, Data Caching at Fog
Conference (IPCCC), University of South Florida, 2017. Nodes Under IoTNetworks: Review of Machine
24. A. MelloukImag, S. Batim and S. Zeadally, "Fog Learning, techrxiv.org, 2020.
computing job scheduling optimization based on bees 36. J. Moura and D. Hutchison, "Fog Computing Systems:
swarm," Enterprise Information Systems, Vols. Volume State of the Art, Research Issues and Future Trends,"
12, 2018, no. Volume 12, 2018, 10 Apr 2017. Journal of Network and Computer Applications, vol.
25. A. Abeshu and N. Chilamkurti, "Deep Learning: The Volume 169, no. Volume 169, 2020.
Frontier for Distributed Attack Detection in 37. R. Z. Naeem, S. Bashir, M. . F. Amjad, H. Abbas and H.
Fog-to-Things Computing," IEEE Communications Afzal , "Fog computing in internet of things:Practical
Magazine, Vols. vol. 56, no. 2, no. vol. 56, no. 2, pp. pp. applications and future directions," Peer-to-Peer
169-175, Feb. 2018. Networking and Applications, vol. 12, no. 12, p.
26. A. Al-Ghraibah, M. Algharibeh, W. AlMohtasib and M. 1236–1262, 13 September 2019.
Altayeb, "Classification of Long Bone X-ray Images 38. M. Ghobaei-Arani, A. Souri, F. Safara and N. Monire ,
using New features and Support Vector Machine," "An efficient task scheduling approach using

2621
Urooj Yousuf Khan et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 2611– 2622

moth-flame optimization algorithm for cyber-physical

system applications in fog computing," Trans Emerging
Tel Tech., pp. 1-14, 4 September 2019 4 September 2019.
39. H. A. Khattak, S. U. Islam, I. U. Din and M. Guizani,
"Integrating Fog Computing with VANETs: A
Consumer Perspective," IEEE Communications
Standards Magazine, vol. vol. 3, no. 1, pp. pp. 19-25,
March 2019.
40. F. Hussain, . R. Hussain, S. . A. Hassan and . E. Hossain,
"Machine Learning in IoT Security: Current
Solutions and Future Challenges," IEEE
Communications Surveys & Tutorials, vol. vol. 22, no. no.
3, pp. 1686-1721, thirdquarter 2020.
41. D. V and L. S. R, "Intelligent Deep Reinforcement
Learning Based Resource Allocation in Fog Network,"
in 26th International Conference on High Performance
Computing, Data and Analytics Workshop (HiPCW),,
Hyderabad, India, 2019.

2622