Review

International Journal of Distributed

Sensor Networks
2022, Vol. 18(3)
Artificial intelligence Internet of Ó The Author(s) 2022
DOI: 10.1177/15501477211062835
Things: A new paradigm of distributed journals.sagepub.com/home/dsn

sensor networks

Kah Phooi Seng1,* , Li Minn Ang2 and Ericmoore Ngharamike3

Abstract
The advances and convergence in sensor, information processing, and communication technologies have shaped the
Internet of Things of today. The rapid increase of data and service requirements brings new challenges for Internet of
Thing. Emerging technologies and intelligent techniques can play a compelling role in prompting the development of intel-
ligent architectures and services in Internet of Things to form the artificial intelligence Internet of Things. In this article,
we give an introduction and review recent developments of artificial intelligence Internet of Things, the various artificial
intelligence Internet of Things computational frameworks and highlight the challenges and opportunities for effective
deployment of artificial intelligence Internet of Things technology to address complex problems for various applications.
This article surveys the recent developments and discusses the convergence of artificial intelligence and Internet of
Things from four aspects: (1) architectures, techniques, and hardware platforms for artificial intelligence Internet of
Things; (2) sensors, devices, and energy approaches for artificial intelligence Internet of Things; (3) communication and
networking for artificial intelligence Internet of Things; and (4) applications for artificial intelligence Internet of Things.
The article also discusses the combination of smart sensors, edge computing, and software-defined networks as enabling
technologies for the artificial intelligence Internet of Things.

Keywords
Internet of Things, artificial intelligence, deep learning, embedded analytics, distributed sensor networks

Date received: 18 April 2021; accepted: 26 October 2021

Handling Editor: Peio Lopez Iturri

Introduction intelligence (AI) technology have resulted in more

AIoT applications to emerge. These AIoT applications
In recent years, we have witnessed the development have the characteristics of being computationally
and proliferation of billions of mobile and Internet of
Things (IoT) devices. This rapid increase in IoT sen-
1
sors, things, objects, and resources puts a demanding School of AI and Advanced Computing, Xian Jiaotong Liverpool
strain on existing IoT and distributed sensor network University, Suzhou, China
2
School of Engineering and Science, University of Sunshine Coast,
infrastructures. Emerging technologies and intelligent
Sunshine Coast, QLD, Australia
techniques can play a compelling role in addressing 3
Computer Science Department, Federal University, Oye-Ekiti, Nigeria
these new requirements. The result is a new interdisci- *
This author is now affiliated to School of Computer Science,
plinary field and paradigm termed as the artificial intel- Queensland University of Technology, Brisbane, Australia
ligence Internet of Things (AIoT).1 The AIoT is
Corresponding author:
beginning to receive a significant amount of interest Kah Phooi Seng, School of AI and Advanced Computing, Xian Jiaotong
from the research communities and industries. The Liverpool University, Suzhou, China.
widespread acceptance and penetration of artificial Email: kahphooi.seng@qut.edu.au

Creative Commons CC BY: This article is distributed under the terms of the Creative Commons Attribution 4.0 License
(https://creativecommons.org/licenses/by/4.0/) which permits any use, reproduction and distribution of the work
without further permission provided the original work is attributed as specified on the SAGE and Open Access pages
(https://us.sagepub.com/en-us/nam/open-access-at-sage).
2 International Journal of Distributed Sensor Networks

demanding to execute machine-learning or deep learn-

ing algorithms and meet the real-time processing con-
straints. However, IoT devices are resource constrained
(computation, storage, and communications), and this
increases the design and deployment challenges to ful-
fill the quality-of-service (QoS) requirements for the
AIoT and its related applications.
In this article, we analyze the convergence of AI
throughout the IoT architecture to form the AIoT with
a focus on four aspects: (1) architectures, techniques,
and hardware platforms for AIoT; (2) sensors, devices,
and energy approaches for AIoT; (3) communication
and networking for AIoT; and (4) applications for
AIoT. The first aspect is related to the combination of
AI and edge computing as key enabling technologies
for the AIoT. The other remaining three aspects seeks Figure 1. AIoT architecture and layers.
to apply and embed AI and machine-learning tech-
niques toward the design and implementation of the In the application layer, the potential for the AIoT is
various IoT layers and structure. Figure 1 shows the very high. It can be applied to make urban environments
overview concept of the AIoT and the general layer and vehicles smart. In smart homes, home appliances
structure for an AIoT which consists of three layers: (1) such as smart TVs, lights, thermostats, refrigerator, cof-
sensing and device layer; (2) communication and net- fee makers, and so on are equipped with sensors and
work layer; and (3) application layer. The AIoT aims intelligence to learn a user’s habits. The intelligence devel-
to embed AI and machine-learning techniques into the oped can be used to offer automated home support for
sensing, communication, and application layers to daily task, minimize energy usage, and so on. Many
achieve high-performing IoT infrastructures. organizations apply AIoT to assist in managing energy,
In the sensing and device layer, the AIoT paradigm lighting, and access in the office building. Sensors placed
can take advantage of recently developed edge comput- in the office can detect the number of staff present per
ing architectures2 and machine-learning approaches, time and make necessary adjustment to lighting and tem-
such as active learning (AL),3 transfer learning (TL),4 perature to enhance energy efficiency. Sensors also use
and federated learning (FL).5 AL techniques can deal face recognition to determine what personnel to grant
with the time-varying and unpredictable data over the access into certain areas of the office environment.7 The
IoT network. TL utilizes pre-trained models developed introduction of AIoT and edge computing for intelligent
at the edge servers to deliver accurate results. FL can driving allows the vehicle to offload tasks to the server
provide the necessary privacy at the edge server for the closer to the vehicle side, creating a new paradigm for
processing information. In the communication and net- task offloading and resource allocation.
work layer, the AIoT paradigm can take advantage of The survey work has used relevant papers which
newly emerging communication technologies and net- were searched from databases and other sources includ-
works, such as software-defined networking (SDN) and ing IEEE Xplore, ScienceDirect, Scopus, and Google
5G/6G cellular communications. Scholar. We gave priority to papers which have been
A critical factor for the AIoT paradigm is the secu- published more recently (from 2015 onwards). In per-
rity within the IoT network. The heterogeneity of IoT forming the search, we used keywords, such as
devices and massive scale of IoT networks make it ‘‘Artificial Intelligence,’’ ‘‘Internet of Things,’’ and
highly vulnerable to attack in several ways, as the likeli- ‘‘Distributed Sensor Networks’’ using a combination of
hood of detecting an IoT device that has not been AND/OR operators to find the papers which were rele-
properly secured or that has a weak security measure is vant toward the area of AIoT. The studies also utilized
very high. With the deployment of IoT on 5G networks two major contexts which were the problem context
which provide high speed, low latency, and large band- and the solution context to locate the various studies
width, it is easier, in the case of network security which focused on techniques, methods, and applica-
breach, for hackers to steal and download information, tions for AIoT.
such as customer information and personal informa- The remainder of the paper is organized into five
tion, at a faster rate.6 Traditional security schemes are sections. The sections are mostly structured to follow
not efficient in solving security issues of IoT network the different layers in the AIoT as shown in Figure 1.
especially emerging security threats. AI tools have been Section ‘‘The convergence of architectures, techniques,
integrated with IoT (AIoT) to offer more effective and and platforms for AIoT’’ gives discussions on the con-
efficient security solutions for IoT networks. vergence of architectures, techniques, and hardware
Seng et al. 3

Table 1. Summary and discussions for AIoT—a new paradigm of distributed sensor networks.

Area Sub-area Focus of work

Convergence of Edge, fog, and MEC architectures for AIoT Resource-efficient AIoT edge computing
architectures, techniques and (ThriftyEdge),8 AIoT and UAV-mounted MEC,9
platforms for AIoT cooperative edge computing in AIoT devices
(ICE),10,11 acceleration of CNNs on AIoT edge
devices (AWARE-CNN),12 accelerating AIoT
inference,13–15 other works16,17
Novel machine-learning and training Deep learning for AIoT,18,19 transfer learning for
techniques for AIoT AIoT (FATL),20 federated learning for AIoT,20–22
game theory for AIoT scheduling 23
Convergence of sensors, Scalable sensing, computation, and AIoT sensing and deep reinforcement learning
devices, and energy management for AIoT (DRL),24 AIoT and compressive sensing,25 AIoT and
approaches for AIoT decentralized blockchain management26
Wireless power transfer (WPT) and AIoT and energy-harvesting devices,27 AIoT and
energy-harvesting approaches for AIoT WPT on-device learning (WPEG)28
Convergence of Frameworks for AI-enabled IoT networks AIoT and contention-based random access,29 AIoT
communication and over-the-air computation (AirComp),30 AIoT
and networking for AIoT and WLAN,31 AIoT and non-orthogonal multiple
access (NOMA),32 AIoT and multi-hop ad hoc
network33
AI-enabled IoT and cellular networks AIoT and 5G cellular technologies,34 AIoT and 6G
cellular technologies35
Software-defined networking (SDN) for SDN-AIoT and channel assignment,36 SDN-AIoT and
AIoT QoE,37 SDN-AIoT intrusion detection system
(NIDS),38,39 SDN-AIoT and endogenous trusted
network (ETN)40
AIoT and MEC data streams AIoT and MEC caching,41 AIoT and DRL MEC
resource sharing,42 AIoT and MEC VR streaming43
Convergence of Energy, smart grids, and AIoT AIoT and energy prediction,44,45 AIoT and microgrid
applications for AIoT energy dispatch,46 AIoT and energy edge
management,47 AIoT and blockchain architecture for
natural gas48
Industry, smart buildings, and AIoT AIoT and smart building management,49 AIoT and
thermal modeling,50 AIoT and fire evacuation51
Vehicles, smart transportation, and AIoT AIoT and blockchain for intelligent driving edge
systems,52 AIoT and Safety-as-a Services
infrastructure53
Biomedical, smart health, and AIoT AIoT and assisted health systems,54 AIoT and stress
classification system55
Environmental, smart agriculture, and AIoT for disease detection (RiceTalk),56 AIoT for
AIoT environmental monitoring,57 AIoT for farms
(PigTalk)58
Education, smart entertainment, and AIoT AIoT and intelligent learning environments,59–61
AIoT for smart entertainment62,63
Robotics and computer vision in AIoT AIoT drones for smart surveillance,64 AIoT robots
for disaster management,65 AIoT virtual emotion
detection system (CONTVERB),66 AIoT for
intelligent parking management67
Security in AIoT AIoT security for smart cities,68 AIoT and intrusion
detection,69,70 AIoT and trust recommendation,71
other works on security in AIoT72–77

platforms for AIoT. The discussions here include edge, networking for the AIoT. Section ‘‘The convergence of
fog, MEC, and cloud architectures for AIoT deploy- applications for AIoT’’ gives discussions on the conver-
ment. Section ‘‘The convergence of sensors, devices, gence of various applications for the AIoT. Section
and energy approaches for AIoT’’ gives discussions on ‘‘Conclusion’’ concludes the article with some conclud-
the convergence of sensors, devices, and energy ing comments and remarks. Table 1 shows a summary
approaches for the AIoT. Section ‘‘The convergence of and discussions for the AIoT—a new paradigm of dis-
communication and networking for AIoT’’ gives dis- tributed sensor networks, to also serve as a concise
cussions on the convergence of communication and summary for the article.
4 International Journal of Distributed Sensor Networks

Figure 2. Task graph and flow chart for ThriftyEdge:8 (a) task graph for topological sorting and (b) flow chart for computation
offloading.

The convergence of architectures, optimal offloading strategies to the UAV MEC servers
techniques, and platforms for AIoT based on a game-theoretic model. The pure Nash equili-
brium (PNE) strategies were determined following the
This section discusses the convergence of architectures, theory of submodular games. Their experimental results
techniques, and platforms for AIoT from two aspects: demonstrated the operational characteristics and per-
(1) edge, fog, and mobile-edge computing (MEC) archi- formance of the models. Gong et al.10 proposed another
tectures for AIoT and (2) novel machine-learning and approach to perform cooperative edge computing in
training techniques for AIoT. A summary of the repre- AIoT devices termed as ICE (intelligent cooperative
sentative studies discussed in this section can be found edge). In their approach, the AI computations are re-
in Table 1. designed from the cloud to be distributed and operate
on edge devices. Their distribution approach utilized
lightweight pipelines for cloud compression and edge
Edge, fog, and MEC architectures for AIoT reconstruction. The authors’ prototype and their eva-
Chen et al.8 proposed an approach for resource- luation showed that their approach could enable a use-
efficient edge computing in AIoT termed as ful combination of AI and edge computing.
ThriftyEdge. In this work, they proposed a device- Sanchez et al.12 proposed an approach for the real-
centric approach and a resource-efficient edge comput- time acceleration of convolutional neural networks
ing scheme to minimize the cloud resource utilization (CNNs) on edge devices termed as AWARE-CNN.
and be able to satisfy its QoS requirement. Their Their approach utilized reconfigurability features of
approach is based on a novel mechanism which consists field-programmable gate array (FPGA) devices to con-
of an efficient topology sorting–based task graph parti- struct application-specific AIoT architectures to guar-
tion algorithm. An optimal virtual machine selection antee the latency-aware execution for IoT data. Figure
method is also proposed to minimize the IoT device 3 shows the architecture of AWARE-CNN from a
edge resource occupancy and meet its QoS require- high-level view to its microarchitecture. Figure 3(a)
ment. The virtual machine types are ranked and then shows an overview of AWARE-CNN which creates a
selected based on the ranking to utilize the graph parti- translation from the high-level logical description of
tion algorithm. Figure 2(a) shows the task graph for CNNs to an application-specific dataflow execution
topological sorting and partitioning, and Figure 2(b) engine. Figure 3(b) shows the logical execution of cross-
shows the flow chart of computation offloading profile. layer parallelism in AWARE-CNN. Figure 3(c) shows
Fragkos et al.9 proposed an AI-MEC framework by the microarchitecture where each layer is mapped to a
exploiting the computing capabilities of a unmanned pipeline stage containing a buffer-sized tile.
aerial vehicle (UAV)-mounted MEC server for the IoT A current limitation of DL systems in IoT environ-
applications. In this work, the authors formulated ments is the efficient assignment of compute tasks.
Seng et al. 5

Figure 3. AWARE-CNN acceleration on edge devices:12 (a) overview of AWARE-CNN, (b) cross-layer parallelism in
AWARE-CNN, and (c) microarchitecture in AWARE-CNN.

work, the authors used a mixed-integer nonlinear pro-

gramming approach and divided the problem into two
subproblems: (1) a computing resource allocation
(CRA) problem with fixed partitions and (2) a DNN
partition deployment (DPD) problem. The authors
then proposed to solve the CRA problem based on a
Markov approximation algorithm and proposed a low-
complexity DPD algorithm to solve the DPD problem
to achieve a near-optimal solution. The computational
complexity for cooperative multiaccess edge computing
can increase significantly when there is large number of
users which leads to delays in decision-making. Shi
et al.11 proposed an approach for deep reinforcement
learning (DRL) for the task placement problem to
Figure 4. AAIoT inference computation on edge devices.13 enable servers to reduce the average service delay for
cooperative multiaccess edge computing. Their
Zhou et al.13 proposed an approach for allocating the approach utilized a mean-field game-guided DRL
inference computation among the IoT devices termed approach to address the task placement problem.
as AAIoT (accelerating artificial intelligence Internet of Liu et al.16 considered another challenge for edge
Things). Figure 4 shows the proposed AAIoT computing in mobile wireless sensor networks (WSNs)
approach. In their approach, the first device collects with disjoint connectivity. In their work, the authors
the data, performs the computation for the first layer, proposed a rendezvous selection strategy for data col-
and delivers the results to the next device. This process lection in disjoint WSNs with mobile edge nodes and
is carried on for the different devices, and the final optimize the path length and network connectivity.
result is communicated back to the first device. The Their approach utilized the ant colony optimization
operations of each layer depend on the results of the (ACO) algorithm to give two advantages: (1) simplify-
previous layer. ing the path construction and computational cost and
He et al.14 proposed an approach to reduce the end- (2) reducing the search space and increasing the conver-
to-end inference delay for DL implementation in multi- gence speed. Yang et al.15 proposed an approach for
ple deep neural network (DNN) partitions. In this energy-efficient execution of deep learning inferences
6 International Journal of Distributed Sensor Networks

Figure 6. Edge-based AIoT (EBI-PAI) architecture.17

Figure 5. System model of edge inference for DNNs.15
during incremental deployment. They formulated the
tasks for deployment on mobile edge devices. Figure 5 deployment problem, proved its complexity, and designed
shows the system model of edge inference for DNNs. heuristic algorithms with near-optimal performance
In this work, the authors proposed to minimize the (greedy minimum dominating set algorithms (GDSAs)
sum of computation and transmission power consump- and greedy cover for QoE-Aware (GCQA)). Their work
tion under probabilistic QoS constraints. The authors was validated with a case study with real user demands.
developed a reweighted power minimization approach
by iteratively solving a regularized reweighted l1.
Yang et al.17 proposed an Edge-Based IoT Platform Novel machine-learning and training techniques for
for AI (EBI-PAI) by integrating SDN and serverless AIoT
technology into QoE-aware edge computing to address A desirable goal is to accelerate the training process of
the issues of resource scheduling and service definition machine and deep learning approaches on the AIoT
in AIoT. Figure 6 shows the system architecture for the network. Liu et al.18 proposed a hierarchical training
EBI-PAI which consists of three layers: (1) control framework termed as HierTrain which has the capabil-
layer; (2) mapping layer; and (3) edge layer. The ity to efficiently deploy DNN training tasks on the
authors investigated the QoE-aware server deployment mobile-edge-cloud computing (MECC) model. Figure 7
problem to optimize the architecture performances shows the system overview of the HierTrain framework

Figure 7. HierTrain AIoT training and deployment over MECC.18

Seng et al. 7

Figure 8. FATL architecture for IIoT applications.20

which consists of three stages: (1) profiling stage— IoT. In this work, the authors proposed an approach
profiling the execution time for different layers on the termed as FADE (Federated Deep Reinforcement
device, edge and cloud; (2) optimization stage—selecting Edge). The FADE approach enabled base stations to
the optimal partition model and determining the train- learn a shared predictive model by utilizing the first-
ing samples for the edge device, edge server, and cloud round training parameters as the initial input for the
server; and (3) hierarchical training stage—sending the training process. Their experimental results showed
partitioned samples to the edge and cloud servers. Their that their approach gave a reduction in performance
work was implemented and deployed on a hardware loss and average delay, and improved the cache hit rate
prototype over the three levels. Their results showed efficiency. Zhou et al.22 proposed an approach for
that HierTrain could achieve an increase of 6.9 times coordinating edge and cloud architectures for the opti-
performance speedup compared to the cloud-based mization and cost efficiency of FL. In this work, the
approach. authors utilized Lyapunov optimization theory and
Foukalas and Tziouvaras20 proposed a combined developed an optimization framework for load balan-
FATL (federated active transfer learning) model for cing, data scheduling, and accuracy tuning for the
industrial IoT (IIoT) applications. Figure 8 shows the dynamically arriving training data samples. The
architecture of the proposed FATL for IIoT. Their authors showed that their approach had the advantages
proposed architecture had the objective to increase the for reducing both the computation and communication
learning speed and reduce the amount of labeled data cost.
during learning iterations for training on the IIoT Chiang et al.19 proposed an approach for co-task
devices. The authors utilized various machine-learning processing in multiaccess edge computing systems. The
techniques to achieve these benefits. TL using a pre- authors proposed an approach termed as DDL (deep
trained artificial neural network (ANN) and FL were dual learning) where the learner updates primal and
used for the edge IIoT architecture and scheduling for dual variables based on randomly perturbed samples.
multiple devices, whereas AL was deployed on the IIoT Other than machine-learning approaches, algorithms
end devices. Their experimental and simulation results from other fields, such as game theory, could also pro-
showed that the proposed FATL architecture gave high vide useful algorithms for IoT architectures. IoT fog
accuracy and allowed scaling for a number of IIoT nodes have the characteristics of having low resource
devices. capabilities. An important challenge is to match the
Another usage of FL was proposed by Wang et al.21 IoT services to fog node resources while guaranteeing
toward the problem for cooperative edge caching for the minimal delay for IoT services and efficient resource
8 International Journal of Distributed Sensor Networks

utilization. Arisdakessian et al.23 proposed a multicri- Scalable sensing, computation, and management for
teria intelligent IoT-fog scheduling approach using AIoT
game theory. Their approach proposed to design the
The scalability for IoT sensing, computation, and man-
preference list functions for the IoT and fog layers for
agement is an important issue to be addressed due to
ranking based on the criteria for latency and resource
the geographically and rapidly increasing number of
utilization.
IoT devices. The advantages of the AIoT paradigm
The evolvement of edge, fog, and MEC architectures
where the sensing and computation can be offloaded to
will play an important role in future research for the
the IoT devices is expected to offer users with higher
development of the AIoT to fully realize its potential
quality of experience (QoE). However, the current com-
and advantages for real-world deployments and appli-
putation offloading with QoE approaches using
cations. The following gives some aspects and main
machine-learning and deep learning approaches may
interests for researching and future trends on the con-
have issues of instability and slow convergence. Lu
vergence of architectures, techniques, and platforms for
et al.24 proposed an approach for offloading computa-
AIoT:
tion by utilizing DRL. The authors proposed an algo-
rithm based on double-dueling deterministic policy
 A critical requirement would be the development
gradients (D3PG) to address the instability and conver-
of platforms and architectures that will support
gence issues for DRL while offering users better QoE.
and coordinate the usage of distributed machine-
Their QoE approach has the capability to incorporate
learning algorithms and techniques throughout
multiple factors and elements into the model: (1) com-
the AIoT ecosystem. Some illustrative examples
putational and transmission latencies; (2) computa-
are the work by Teerapittayanon et al.,78 who
tional and transmission energy consumption; and (3)
proposed an approach for implementing distrib-
success rate for tasks.
uted DNNs over cloud and edge devices and the
Compressive sensing (CS) techniques gives a poten-
work by Savaglio et al.79 who proposed an
tial way to address the energy efficiency issue in AIoT
approach for distributed data mining on edge
sensor nodes. However, this approach is limited by the
devices;
overhead of signal reconstruction constraints in the
 A second trend and focus would be the develop-
sensor node and the analysis in the remote server.
ment of AIoT architectures, techniques, and
Wang et al.25 proposed a compressive sensing architec-
platforms which are robust and secure to hostile,
ture utilizing deep learning for implantable neural
adversarial, and imposter attacks. An interesting
decoding. The objective of their work was to improve
research direction would be to utilize blockchain
the wireless transmission efficiency and reduce the
technologies to protect against cloned or coun-
overheads. The authors proposed a two-stage classifica-
terfeit edge nodes or devices and demand proof
tion procedure: (1) coarse-grained screening module
of authenticity;80
and (2) fine-grained analysis module. The screening
 A third focus would be the development of tech-
module performed the front-end classification task and
niques and mechanisms for privacy-preserving
transmits the compressed data to a remote server for
data collection and mining within the AIoT. This
the fine-grained analysis.
is specifically important for AIoT containing het-
Xu et al.26 proposed an approach utilizing a
erogeneous nodes and devices.
blockchain-enabled edge computing platform termed
as Edgence for the management of decentralized IoT
The convergence of sensors, devices, and applications. Figure 9(a) shows the decentralized
Edgence architecture which is deployed on edge clouds
energy approaches for AIoT
of MEC. The Edgence platform consists of several
The convergence of IoT and AI for the AIoT needs to master nodes. A masternode contains a full node of the
be supported by different types of sensors, devices, and blockchain and a collateral. In this work, the authors
sensing approaches based on AI and other approaches. proposed using a three-stage validation process: (1)
A review of AI-based sensors and deployment of AI- script validation; (2) smart contract validation; and (3)
based sensors for the next-generation IoT applications master node validation. Figure 9(b) shows the decen-
can be found in Mukhopadhyay et al.’s study81 This tralized AI training under the management of Edgence
section discusses the convergence of sensors, devices, which used feed-propagation and back propagation for
and sensing approaches for the AIoT from two critical updating the AI models. The initial layers are trained
aspects: (1) scalable sensing, computation, and manage- using the data sets of mobile users at the edge clouds,
ment for AIoT and (2) wireless power transfer (WPT) whereas the latter layers are trained at the remote cloud
and energy-harvesting approaches for AIoT. center.
Seng et al. 9

Figure 9. Decentralized Edgence AI architecture and training:26 (a) decentralized Edgence platform architecture and
(b) decentralized Edgence AI models training.

WPT and energy-harvesting approaches for AIoT hybrid-AC (MD-Hybrid-AC)—constructs a centralized

There are several challenges which remain to be critic to output server selections which considers the
addressed to enable the practical deployments of WPT continuous action policies of all devices.
technologies in AIoT, such as ensuring security require- Wu et al.28 proposed a novel approach which inte-
ments to overcome selfish and untrusted devices and grates WPT and battery-constrained on-device learning
providing monetization incentives for charging services. techniques termed as WPEG (Wirelessly Powered Edge
Zhang et al.27 proposed an approach for offloading intelliGence). Figure 10(a) shows the framework of the
time-varying computation task loads and requirements proposed WPEG nodes. Each node has the responsibility
for deployment in MEC and energy-harvesting IoT to manage the edge devices and WPT nodes within
devices. The authors approach utilized hybrid-DRL its coverage. In this work, the authors also proposed a
techniques consisting of: (1) hybrid decision-based permissioned blockchain approach to establish a secure
actor-critic learning (Hybrid-AC)—an improved actor– point-to-point (P2P) energy trading for edge intelligence.
critic architecture where the actor outputs continuous Figure 10(b) shows the permissioned blockchain
actions corresponding to every server, and the critic structure in the WPEG. The authors further proposed a
evaluates the continuous actions and outputs the dis- two-stage Stackelberg game model to achieve the optimi-
crete action of server selection and (2) multidevice zation for transmission power and incentive strategies.
10 International Journal of Distributed Sensor Networks

Figure 10. Framework and blockchain structure for WPEG:28 (a) framework of proposed WPEG nodes and (b) permissioned
blockchain structure in WPEG.

The following gives some aspects and main interests  The need for highly efficient energy manage-
for researching and future trends on the convergence of ment, harvesting, and WPT techniques remains
sensors, devices, and energy approaches for AIoT: another critical requirement for the continuing
development of the AIoT. Ijemaru et al.83
 A potential solution for scalable sensing to cover discuss the utilization of mobile collectors for
large-scale areas and environments, such as opportunistic IoT sensing with WPT. Security
smart cities would be to exploit mobile, opportu- aspects for WPT are going to become more
nistic, and/or vehicular crowdsensing technolo- significant and are opportunities for more
gies and/or the internet of vehicles (IoV) for research. Encryption techniques for WPT are
deployment of the AIoT sensors. Ang et al.82 required so that only the authorized receptors
discuss the applications, architecture, and chal- can acquire the transmitted energy and
lenges for IoV sensing and deployment in smart unauthorized receptors are prevented from
cities; illegally accessing the transmitted energy.
Seng et al. 11

Figure 11. Network architecture showing AI in the IoT.84

The convergence of communication and Frameworks for AI-enabled IoT networks

networking for AIoT Song et al.29 proposed an approach for centralized and
There are several advantages and challenges to con- distributed AI-enabled IoT networks. In this work, the
verge AI in the network layer for routing, scheduling, authors discussed the challenges of contention-based
and enhancing the QoS. Ijaz et al.84 discussed some random access and proposed some AI approaches in
challenges of AI in wireless networks for IoT (as centralized IoT networks. The authors proposed to uti-
shown in Figure 11). Their article focused on the lize DRL techniques for the online training process of
communication aspects for IoT and proposed a gen- the deep Q-network (DQN). Two neural networks
eric network architecture showing a few key areas of (NNs) are used, namely an evaluated NN and a target
AI in the IoT: (1) software-defined network; (2) net- NN for the DRL. For distributed IoT networks, the
work function virtualization (NFV); (3) multiaccess authors pointed out some key technical challenges of
edge computing; and (4) content delivery network spectrum access and spectrum sensing. The work also
(CDN). They discussed some challenges for the above proposed a DQN-based spectrum access strategy for
areas. These included challenges for latency-critical intelligent spectrum access. Yang et al.30 proposed an
IoT, challenges in routing and network traffic con- over-the-air computation approach termed as AirComp
trol, challenges in caching, and so on. This section for a communication-efficient federated machine-
discusses the convergence of AIoT in communication learning architecture. Figure 12 shows the AirComp
and networking from four aspects: (1) frameworks architecture and communication process of federated
for AI-enabled IoT networks; (2) AI-enabled IoT and machine learning in intelligent IoT. The global and
cellular networks; (3) SDN for AIoT; and (4) AIoT local models are iteratively updated through a number
and MEC computing. of communication rounds until the consensus of the
12 International Journal of Distributed Sensor Networks

Figure 12. AirComp architecture and communication in AIoT.30

global model is reached. As shown in the figure, three

steps are performed for each learning round: (1) device
selection; (2) local model upload; and (3) global model
download. The AirComp approach achieved this by
using the superposition property of a multiaccess chan-
nel to achieve fast local aggregation for federated
machine learning.
Ramı́rez et al.31 proposed a WLAN and AIoT archi-
tecture for resource allocation and sharing among the
devices in the IoT. The design of this architecture is for
the IoT-WLAN. Figure 13 shows the AIoT-WLAN
architecture. The proposal is based on an IoT network
with centralized management architecture and AI con-
trol agent. Their proposed approach utilizes three con-
trol agents for the different environments. The AI
control agents are located inside the smart things, in the
centralized gateway administration, and in the cloud
Figure 13. AIoT-WLAN architecture.31
platform. The AI agents are responsible for learning the
machine-to-machine (M2M) relations. Each smart thing
connects to the network using the MQTT protocol. used a model-driven deep learning technique called
Non-orthogonal multiple access (NOMA) schemes offline-trained LSTM-based network to perform the
have a high potential for massive machine-type commu- multi-user detection. Their work showed that the pro-
nication (mMTC) framework because it allows multiple posed approach could outperform conventional detec-
users and devices to share resource blocks. Emir et al.32 tors (even with perfect channel state information (CSI).
proposed a deep learning–aided multi-user detection Their experimental results showed that the same perfor-
approach termed DeepMuD for uplink grant-free mance was obtained for increasing number of users
NOMA IoT networks to utilize the massive machine- with a 10-dB decrease in power consumption.
type communication. Figure 14 shows the proposed Kwon et al.33 proposed a distributed approach to
DeepMuD model. In the proposed model, the authors design a multi-hop ad hoc IoT network as shown in
Seng et al. 13

Figure 14. Non-orthogonal multiple access (NOMA) AIoT.32

Figure 15. Distributed AIoT multi-hop ad hoc network.33

Figure 15. In this work, the authors formulated a communication channels. Wang et al.34 proposed an
decision-making process based on an MDP framework architecture for 5G intelligent Internet of Things (5G I-
which considers the trade-off between network IoT) which consist of three major components: (1) pro-
throughput and individual transmission power cessing center; (2) object processor; and (3) sensing
consumption. A DRL algorithm using a double deep regions. Figure 16(a) shows the architecture of the 5G
Q-network (DDQN) was deployed for decision-making I-IoT, and Figure 16(b) shows the details of the pro-
at the relay nodes. Each node owns a DDQN to build cessing center component. The processing center in the
its own policy e.g. transmission range, etc. The learning cloud includes the intelligent computing module and
process for their system enables individual relay nodes execution module to process data automatically with
to make decisions using a minimal amount of intelligent algorithms. DL algorithms are used to
information. design the intelligent computing module.
Zhang et al.35 discussed their model of the AIoT with
6G cellular technologies. Figure 17 shows their AIoT
AI-enabled IoT and cellular networks model with 6G technologies which includes three
The development of cellular network technologies such aspects: (1) mobile ultra-broadband—terabits per sec-
as 5G and 6G provides the potential to deploy complex ond wireless transmissions; (2) super IoT—supplemen-
sensors in the AIoT and to optimize the ted by symbiotic radio and satellite-assisted IoT
14 International Journal of Distributed Sensor Networks

Figure 16. 5G I-IoT architecture and processing center:34 (a) architecture overview and (b) processing center component.

communications to support a wider coverage and larger data. Tang et al.36 proposed an approach for SDN-IoT
number of connected IoT devices; and (3) AI— which exploits partially overlapping channel assignment
advanced machine-learning focused on deep learning (POCA) in the SDN-IoT. In this work, the authors first
and reinforcement learning. investigated and showed that the conventional fixed
POCA algorithms are not viable for the highly dynamic
large-scale SDN-IoT. Figure 18 shows the wireless
Software-defined networking for AIoT SDN-IoT with a heterogeneous structure and the prob-
In recent years, software-defined networks has emerged lem with the fixed POCA. The authors proposed a DL-
as a potential solution to provide cost-effective and flex- based intelligent POCA for the wireless SDN-IoT
ible IoT services. The SDN-IoT can optimize network where the IoT data traffic dynamically changes. Their
resources which may be underutilized to deliver the IoT approach utilized two DL learning strategies for the
Seng et al. 15

Figure 17. AIoT model with 6G cellular technologies.35

prediction of IoT traffic load and to adaptively assign SeArch NIDS architecture, and Figure 20(b) shows the
the POCs. NIDS node graph. The architecture consists of three
QoE for different services is an important factor for layers: (1) edge-IDS; (2) f-IDS; and (3) cloud-IDS.
the AIoT. He et al.37 proposed QoE models to evaluate These IDSs utilized DL and ML algorithms. For detec-
the qualities of the IoT for both network and users. In tion engines in the SeArch architecture, machine-
this work, the authors focused on optimization of cache learning algorithms were used (e.g. SVM for Edge-IDS,
capacity among content-centric computing nodes and SOM for Fog-IDS, and Stacked autoencoder for
the transmission rates under a constrained total net- Cloud-IDS).
work cost. Figure 19 shows the cache resource alloca- Li et al.39 proposed an intrusion detection approach
tion for content-centric IoT. The authors formulated by utilizing an AI SDN-IoT architecture. Figure 21
the QoE as a cache resource allocation problem under shows the proposed AI SDN-IoT architecture. The
the different transmission rate to acquire the best QoE. authors utilized different types of AI algorithms to per-
Their approach used DRL to adaptively improve the form the feature selection and flow classification tasks.
QoE. These algorithms were (1) improved Bat algorithm with
Wu and Zhang38 proposed a collaborative and intel- swarm division and differential mutation and (2)
ligent network-based intrusion detection system improved random forest. The improved Bat algorithm
(NIDS) architecture termed as SeArch for SDN-based with swarm division and differential mutation was used
cloud IoT networks. Figure 20(a) shows the proposed to select the typical features. The algorithm splits the
16 International Journal of Distributed Sensor Networks

Figure 18. Wireless SDN-IoT.36

Figure 19. Cache resource allocation for content centric IoT.37

swarm into several subgroups. Each subgroup is then approach is used to achieve trust between resource
used to learn among the different populations requesters and resource providers in a distributed
efficiently. network.
Guo et al.40 proposed an approach for an endogen-
ous trusted network (ETN) to support virtual network
operations, resource allocation, and provide shared ser- AIoT and MEC data streams
vices. In this work, the authors utilized blockchain MEC is a promising technology to improve QoE in
technology together with SDN and NFV to achieve the AIoT networks. Jiang et al.41 proposed an approach
trusted sharing of network resources. Figure 22 shows termed as Crowd-MECS which used a crowdsourcing
an overview of the proposed ETN framework. The framework for mobile edge caching and sharing.
ETN consists of five layers: (1) terminal device layer; Figure 23 shows the Crowd-MECS architecture. The
(2) network layer; (3) blockchain layer; (4) platform two edge devices (in green) which denote a mobile
layer; and (5) application layer. Their blockchain edge device and a fixed edge device store the cache
Seng et al. 17

Figure 20. SeArch NIDS architecture and node graph:38 (a) SeArch NIDS architecture and (b) NIDS node graph.

contents. These cached contents are shared with the to learn the optimal viewport rendering offloading and
red edge devices. Incentives are offered to the green transmit power control policies and proposed an actor–
edge devices for the content caching. critic (A3C)-based joint optimization algorithm.
Other works for AIoT and MEC can be found in Jiang The following gives some aspects and main interests
et al.42 and Du et al.’s studies.43 Jiang et al.42 proposed a for researching and future trends on the convergence of
resource-scheduling approach based on DRL for large- communication and networking for AIoT:
scale MEC networks. In this work, the authors proposed
a DRL approach to minimize the sum of weighted task  The future will see the development of large-
latency for all IoT users by optimization of the task off- scale (e.g. citywide) IoT and AIoT networks. In
loading decision, transmission power, and resource alloca- these scenarios, interoperability is an important
tion in the large-scale MEC system. Du et al.43 proposed a consideration in view of the rapid advances in
DRL approach for optimization of the energy consump- new technologies such as 5G and 6G. The
tion for immersive VR video streaming over terahertz challenges for interconnection of multiple IoT net-
wireless networks. In this work, the authors utilized DRL works which are owned by different organizations/
18 International Journal of Distributed Sensor Networks

stakeholders and/or developed using different com-

munication networks/protocols and heterogeneous
technologies remain an important issue to be satis-
factorily resolved. SDN and NFV are interesting
research directions and solutions to resolve the
interoperability issues in IoT and AIoT networks;
 There is a growing importance to build trust into
the communication networks for IoT. A growing
trend in this regard is the replacement of centra-
lized IoT trust architectures with decentralized
architectures. An interesting research direction
would be to utilize distributed ledger technolo-
gies (DLTs) for trust development and protec-
tion in IoT networks. Danzi et al.85 give a
discussion of the challenges for the integration
of IoT networks and DLT.

The convergence of applications for AIoT

The convergence of AI and IoT enable the utilization of
AI tools to introduce intelligence and decision-making
Figure 21. AI-based SD-IoT architecture.39 capability into diverse IoT applications and systems.

Figure 22. Endogenous trusted network (ETN) AIoT.40

Seng et al. 19

Figure 23. Crowd-MECS architecture.41

This convergence is helping to solve multiple IoT and household appliances. The authors utilized differ-
problems that have slowed its global adoption. The ent sensors, such as piezoelectric sensor, which gener-
AIoT focuses on providing fast services, helping to ates energy from stress caused by human body weight,
unleash the potentials of utilize data in a better and body heat to electric converter which generates energy
faster way.86 For example, in transportation applica- from heat generated by movement of human body, and
tions, vehicles can be equipped with camera, GPS, solar panel which generates energy from the sunlight.
radars, and sonar sensors to gather and analyze data The authors built and validated the data collected from
to determine the condition of roads, pedestrian beha- the sensors with two AI models (ANN and ANFIS) to
vior, and so on and make decisions in response to perform the prediction of generated power output. The
such situations. Transportation companies have authors showed that their system could produce accu-
deployed AIoT to monitor and manage their fleets. rate results in predicting the power generated from
Data gathered by the fleets is analyzed to manage fuel renewable resources.
cost, track maintenance, and identify driver’s unsafe Lei et al.46 proposed a DRL-based energy dispatch
behaviors. This section discusses the convergence of approach for IoT microgrids (MGs). The authors for-
AIoT applications from eight aspects: (1) energy, mulated and solved a finite-horizon partial observable
smart grids, and AIoT; (2) industry, smart buildings, Markov decision process (POMDP) model by learning
and AIoT; (3) vehicles, smart transportation, and from past data to identify uncertainties in future renew-
AIoT; (4) education, smart entertainment, and AIoT; able power generation and electricity consumption,
(5) biomedical, smart health, and AIoT; (6) environ- which is the focus of the authors. To address the
mental, smart agriculture, and AIoT; (7) robotics and instability challenges of DRL algorithms and the dis-
computer vision in AIoT; and (8) security in AIoT. A tinctive features of finite-horizon models, the authors
summary of the representative studies which are dis- designed two novel DRL algorithms. The first is a
cussed in this section can be found in Table 1. finite-horizon deep deterministic policy gradient (FH-
DDPG) for finite-horizon MDP based on DDPG. The
authors conducted experiments applying the idea of
Energy, smart grids, and AIoT time-dependent actor to the proposed algorithm against
Researchers have deployed the AIoT in multiple ways the baseline DDPG. The results showed that the pro-
to help energy producers optimize their equipment for posed FH-DDPG performed better than the baseline
better service delivery and to help users efficiently man- DDPG in providing stability to energy management
age energy to minimize cost.87 Puri et al.44 developed problem. The second algorithm is a finite-horizon
an AIoT system that generates electrical energy from recurrent deterministic policy gradient (FH-RDPG) to
different sensors which can be used in industrial areas solve the partial observable problem. This algorithm
20 International Journal of Distributed Sensor Networks

was evaluated against RDPG. The result showed that solve the problem of load forecasting in energy man-
the FH-RDPG algorithm performed better than base- agement. Their proposed framework is composed of
line RDPG algorithm in a POMDP domain. edge devices connected to a cloud server in an IoT net-
Liu et al.47 proposed an IoT-based system for man- work. The IoT network is connected to smart grids to
aging energy on edge computing devices using DRL. effectively maintain energy demand and supply activi-
Their proposed architecture for the system consists of ties. To deal with several parameters contained in
three major components: (1) energy device; (2) energy energy data set, the authors applied techniques to pre-
edge server; and (3) energy cloud server. The energy process the data before initial training. Their experi-
device is any entity in the network that requires and mental results showed that the approach was able to
can supply energy. The energy edge server is a server predict energy consumption with high accuracy.
deployed either at the network gateway or base station.
The energy cloud server is connected to the energy
management central controller with the responsibility Industry, smart buildings, and AIoT
of providing real-time analysis, maintaining record of
AIoT has been adopted in industries and manufactur-
devices connected to the network and proffering com-
ing to help foresee challenges and deal with them before
putational support to energy edge servers. The authors
they cause breakdown of machinery and operations. It
designed two DRL approaches of energy-scheduling
enables M2M sensors and other industry smart devices
scheme with two phases—offline DNN and online
to receive data, analyze it, and take decision to optimize
dynamic deep Q-learning phase. The DRL scheme is
deployed at the energy edge server and energy cloud operation, avoid unplanned costly downtime, plan
server. In the first approach, the task of scheduling logistics and supply chain, respond to cyber threats,
energy is transferred from the devices to the edge server and manage staff safety. Many researches have been
which utilizes the DRL approach to find the best sche- done to better utilize AIoT in this domain.88
duling solution for devices. In the second approach, the An automatic real-time online method to evaluate
edge server, in order to minimize computation cost, the reliability and trustworthiness of cyber physical sys-
relinquishes the task of training the DNN to the cloud tems (CPS) in an IIoT network was proposed by Lv
server and then uses the deep Q-learning operation, et al.49 The evaluation framework and online rank algo-
making use of the estimated Q-value from the server. rithm was designed using machine learning to achieve
The results of their experiments showed that the DRL the online assessment and analysis in real time. The
scheme outperformed the traditional cloud scheduling result of the assessment could be utilized in a timely
scheme. manner to take preventive measures to address any sys-
Miao et al.48 proposed a blockchain and AI-based tem faults or malicious attacks. To obtain a broad relia-
architecture for the natural gas IIoT to address the sup- bility model assessment strategy, the authors analyzed
ply chain failure of existing centralized supply architec- quintessential CPS control model that utilizes a spatio-
tures. The authors premised that numerous supply temporal correlation model. Based on the assessment
requests from centralized energy supply sites could be strategy, they developed a smart control strategy and a
overwhelming and cause the indicators such pressure, trusted smart predictive model. The proposed model
temperature, and natural gas load to exceed safety lim- from the simulation results was very effective in guiding
its. They stated that problems such as obsolete infra- the CPS defenders to select applicable defense resource
structure, unreliable transaction, variable prices, input depending on the type of CPS threat and defense
inaccurate gas data, and unsecured user information domain.
are associated with centralized supply architectures. To Zhang et al.50 presented a learning framework that
address these problems, the authors proposed an archi- recognizes the thermal model of thermal zones auto-
tecture with three dimensions—infrastructure, data, matically, to control the heating, ventilation, and air-
and value. They implemented a blockchain solution to conditioning (HVAC) systems, in a smart building for
provide trusted transaction, distributed storage, authen- the comfort of the residents, under different environ-
ticity, and reliability of stored data and data generated mental conditions, in order to lessen energy consump-
during energy generation and transmission. They fur- tion. The proposed framework learns the thermal
ther proposed an LSTM-based model to predict natural model from the temperature reading of the smart ther-
gas load and dispatch and a transaction model that uti- mostat. The authors installed an IoT platform in a
lizes natural gas value and supply–demand interaction. real-world building and used the data collected to vali-
Their simulation experiments validated the good effec- date the learning framework. Their experiment results
tiveness of the models. showed that the learned model can be delivered in min-
Han et al.45 proposed a deep learning–based frame- utes if the learning process is done both in a cloud
work to predict future energy consumption in smart infrastructure and on edge device for a single thermal
residential homes and industries. Their aim was to zone. The results also showed that the learned model
Seng et al. 21

produces reliable evaluation of thermal comfort in trust measures are then used to evaluate trust of every
intelligent control implementation. data item generated from the network. An integer lin-
Jiang51 proposed an intelligent dynamic evacuation ear programming (ILP) model is formulated by the
path solving model in the case of fire outbreak in public authors to identify optimal data with reduced effect of
buildings using AI technology. The author discussed unauthorized nodes for decision-making. Experimental
that in order to bring people to safety in a fire outbreak results of the proposed scheme against greedy methods
situation, there should be a means of determining the show that the proposed scheme outperformed existing
shortest effective route to evacuate people depending methods.
on the verified fire situation, internal structure of the
building, and fire product influences, such as tempera-
ture, smoke, and carbon monoxide concentration. The Biomedical, smart health, and AIoT
model was built based on appropriate fire emergency The AIoT has been employed to manage and process
evacuation practices. The model was implemented using large amounts of data generated in the internet of
ant colony algorithm, and an AI-based mobile terminal medical things (IoMT).90 Yang et al.54 proposed an
system was developed for large public buildings to assist intelligent architecture for processing visual data from
in guiding people to safe exit routes. IoT-assisted health systems. Their architecture consists
of three modules: (1) end processing module; (2) edge
control module; and (3) cloud management module.
Vehicles, smart transportation, and AIoT The intelligence at the end side was generated from the
AI algorithms and techniques which have been inte- analysis of human, machine, and sensor characteristics.
grated into IoT (AIoT) have found important applica- The intelligence in the edge and cloud sides was deter-
tions in vehicles and transportation. An example of mined by an intelligent measurement model proposed by
such application is in self-driving or autonomous vehi- the authors. Their experimental results showed that the
cles. Future self-driving cars and vehicles will be proposed method could outperform existing methods.
embedded with several types of sensing instruments Mustafa et al.55 proposed an AI-based IoT system
(e.g. cameras, lidar, and radar) and will generate huge for detecting and classifying stress. In their approach, a
amounts of data (e.g. 100 GB per second).89 A signifi- wearable device was equipped with various sensors
cant challenge to be addressed is the real-time and (skin conductance, electrocardiograph (ECG), and skin
secure processing of these sensor data to enable fast temperature sensors) to measure the physiological char-
response to complex scenarios such as obstacle avoid- acteristics. The physiological data collected were trans-
ance and velocity adaptation. The AIoT deployed at ferred through the user’s mobile to the cloud. An AI
the network edge together with federated machine algorithm was used to analyze the data to ascertain the
learning and secure trust models offer potential solu- level of stress. The user is alerted of the predicted stress
tions. Xiao et al.52 proposed an approach based on level through a mobile phone. In a situation of a severe
using blockchain for intelligent driving edge systems. stress level, the user’s physician would be alerted for
The approach utilized a double auction mechanism to necessary action. Their system achieved a 97.6% per-
optimize the satisfaction of users and service providers formance accuracy binary classification with regards to
for edge computing. Their experimental results showed real-time sensor data.
that the approach could give better performance for
resource utilization.
Dass et al.53 proposed a trust evaluation scheme Environmental, smart agriculture, and AIoT
called T-Safe to evaluate the trustworthiness of data Many researches utilizing AIoT technology have been
generated from sensing nodes in an IoT-based intelli- focused toward smart agriculture, food processing, and
gent transportation system. Safety-related information optimization of environmental conditions to increase
are provided to end users by safety-as-a-services infra- food production. Chen et al.56 proposed an AIoT-based
structure based on data generated by the sensor nodes scheme that detects rice blast disease called RiceTalk.
by a method termed as decision virtualization. The The scheme is based on soil cultivation IoT platform.
authors premised that the accuracy and efficiency of Agriculture IoT sensors are used to gather data, which
such information depends on the trustworthiness, secu- the AI mechanism trains and analyses automatically in
rity, and privacy of the sensing nodes and the commu- real time. Existing studies made use of hyperspectral
nication medium through which data are transmitted. image or non-image data to determine plant diseases
The authors designed a trust evaluation model to which needs human efforts to capture the photographs
address this problem. They utilized direct and indirect and data for analysis. The novel capability of RiceTalk
trust mechanism on every node in order to carry out is that the AI model is treated and managed as an IoT
trust measures update at regular time intervals. The device which significantly minimizes the cost of
22 International Journal of Distributed Sensor Networks

managing the platform to deliver real-time training and Node Red, NEMA GUI Builder and Embedded
prediction. Their experimental results showed that Wizard for developing GUI for the IoT platform. Their
RiceTalk could give a prediction accuracy of rice blast framework is designed to provide students the opportu-
at 89.4%. nity to effectively interact during teaching and learning
Nurminen and Malhi57 designed an AIoT-based sys- activities. IoT wearable devices used in educational
tem to monitor and analyze the measurement of humid- application can utilize the framework for enhanced vir-
ity, temperature, moisture, and luminosity to support tualization. Their experiments showed that the frame-
clients for optimizing the conditions of the environment work enabled IoT wearables to provide better learning
for houseplants. In this work, the authors proposed a experiences to users.
solution to provide open interfaces so as to build an Liu et al.60 applied AI analysis and IoT to determine
Open Houseplant Database for intelligent agents. Their student’s concentration levels in individualized learning
approach also used an open messaging interface (O- environments and to enhance learning quality using an
MI) node for data collection from sensor and a smart automated IoT control model. In this work, the authors
device for data analysis and prediction. The users were employed the abilities of image recognition and brain-
provided with an online user interfaces with standard wave signal analysis technologies, and environmental
data virtualization using dialogues and graph. Their quality detection to capture data from students learning
experimental results showed the system showed great different courses under different environmental condi-
potential to support in a smart manner, the cultivation, tions to determine their concentration levels. The data
and management of herbs and houseplants. collected were passed through an ANN computation.
Chen et al.58 proposed an AIoT-based platform to Afterwards, the results were applied to IoT control
detect and mitigate piglet crushing in pig farms termed devices to adjust the learning conditions. Their experi-
as PigTalk. The PigTalk approach was deployed in a mental results showed that learners concentration
farrowing house made up of different farrowing cages. improved as learning conditions improved.
Smart microphones were planted on top of the cage to Zhao et al.61 designed an AI collection of self-built
detect piglet screams and rotating Internet (IP) cameras learning resources. The aim of the work was to popular-
to monitor the cages. Floor vibration and water drop ize the teaching and learning of AI in China. To reduce
actuators were also used to alert a sow to stand up to certain learning factors that discouraged teachers and
detect when a piglet scream. In their approach, raw students from participating in AI teaching activities, the
audio data (the piglet scream) are preprocessed by an authors developed an open hardware resource called
audio clip transform method to detect piglet screaming. STEAM Suite and a graphical drag-block program-
When a sow lies on a piglet, the data (piglet scream) ming language called December. The STEAM Suite
are received by PigTalk through the microphone. provides several types of motors, plates, tires, beams,
PigTalk processed the data in real time and automati- gears, shafts, batteries, and other basic components.
cally activated the actuators (floor vibration, water This enables students to try and build different shapes
drop, etc) or the heat light to force the sow to stand up. of mechanical structures which is expected to motivate
In contrast, existing systems would require a farm them to develop interest in AI as they cultivate strong
worker to locate the particular cage and push up the ability in using the components. The authors verified
sow to avoid crushing. The PigTalk achieved its opera- that due to the practical oriented nature of the learning
tion using two IoT devices: (1) DataBank and (2) ML- resources, the level of interest in teaching and learning
device. The DataBank device provided the improved of AI increased.
generation of spectrograms of vocalizations received The application of AIoT in the entertainment sector
from farrowing cages, whereas the ML_device executed has offered some revolution in many ways. The enor-
the CNN machine-learning model. Their experimental mous data generated can be analyzed and used to
results showed that crushed piglets could be saved by improve user experiences and make targeted business
PigTalk within 0.05 s with a success rate of 99%. decisions. Researchers are utilizing AIoT to develop
smart systems to empower businesses in this sector for
better services delivery.62 De Lima et al.63 proposed an
Education, smart entertainment, and AIoT adaptive method to generate personalized interactive
Several authors have proposed to use AIoT in the edu- storytelling using users’ preferences. The authors inves-
cation field to make teaching and learning more effec- tigated the application of personality modeling and
tive and rewarding. Wangoo and Reddy59 proposed an generated personalized account of experience with
intelligent learning environment framework for IoT respect to users’ personality using the Big Five factors.
wearable devices to enhance interactive learning envir- The proposed method was evaluated in an online inter-
onments. The authors made use of GUI tools such as active storytelling platform. Their experimental results
Seng et al. 23

showed that the proposed approach could effectively achieved 95% accuracy. Li et al.92 proposed an
identify users’ preferences for story episodes with approach for defect detection in large-scale photovol-
91.9% average accuracy and enhanced the experience taic plants. Their approach utilized UAVs and edge
and satisfaction of users. computing to perform the defect detection. In this
work, the authors developed an approach that com-
bines deep and TL with data augmentation for deploy-
Robotics and computer vision in AIoT ment on resource-constraint edge devices. Techniques
The AIoT has enjoyed tremendous application in were also used in the network to reduce the parameters
robotics and computer vision applications. Robots are and the size of the model. Some other examples of
empowered with sensors and AI algorithms to capture AIoT for robotics and computer vision using various
and learn from newer data to develop intelligence. This techniques can be found in the literature. Velasco-
ability of the robot has motivated its use in manufac- Montero et al.93 proposed a methodology to predict
turing, health care, and so on to perform tasks meant the performance of CNNs on vision-based AIoT
for human experts faster and with lower cost.91 In devices. Chiu et al.94 proposed a distributed learning
smart cities, AIoT-enabled drones are used for several approach for AIoT video-based service platforms.
surveillance purposes such as real-time traffic monitor-
ing. Traffic data are transmitted and analyzed and used
to make decisions on the best way to reduce congestion Security in AIoT
by automatically adjusting the speed limits and timing The security in the AIoT is a critical factor. Research in
of the traffic lights.64 this field shows that with AIoT, IoT devices are enabled
Lee and Chien65 designed an AIoT architecture to to learn from data and make swift decisions when an
coordinate surface, underwater, aerial ground robots abnormal behavior is detected in the network in real-
to respond to disaster situations where it is impractical time.95 Chakrabarty and Engels68 proposed a frame-
to send personnel. The robots are connected to an IoT work for the IoT enabled smart city using AI to miti-
network and used to collect data from a disaster site. gate a range of present and future cyber threats. The
The data are transmitted from the field workstations massive adoption of IoT in smart cities provides a wide
through the IoT network to the cloud where it is used surface of attack for a heterogeneous, large and com-
to train a deep learning model. After training and veri- plex smart city system. The authors proposed frame-
fication, the model is transmitted back to the robots work focused on the security of the IoT communication
through the field workstations in order to continue on- protocols. An AI engine was strategically positioned at
site object classification and would enable them make different stages (devices and big data center) within the
response decision as they unceasingly confirm identifi- framework to learn the normal network behavior and
cation with the environment. Kim et al.66 proposed an to monitor the security and health of the network. It is
approach termed as CONTVERB (continuous virtual also able to acquire, store, and analyze big data gener-
emotion detection system) for an IoT environment. IoT ated by IoT devices. Their framework showed domi-
devices are equipped with wireless signal which they nance in providing better effectiveness compared to
transmit to a person within a signal range and also existing smart city frameworks.
receive the reflection of the signal. The IoT devices pro- A self-adaptive and parallel processing intrusion
cesses the reflected signal using a set of heart beat seg- detection system for an SDN network was proposed by
mentation and respiration procedure to obtain no less Suresh and Madhavu.69 The design of the AI-based
than four different kinds of human emotion, such as IDS was done using the self-adaptive energy BAT algo-
sadness, joy, pleasure, and anger. Simulations and rithm. In the initial stage of their design, the software
implementations of the proposed system showed its layer performs analysis of arriving traffic packets and
effectiveness. to perform feature selection. The classification of the
Ke et al.67 proposed an intelligent edge computing packets is performed by the system in the second stage,
approach for the detection of parking space occupancy and if any attack is confirmed, the system takes appro-
in a smart city. Edge AI and IoT are employed so as to priate control measures and makes decisions regarding
break the computation load. The volume of data to be network restrictions, such as resource allocation, rout-
transmitted was designed to be low to handle band- ing, traffic management, and packet management. The
width problem associated with real-time processing of authors performed training using the KDD CUP
video data. Tensorflow Lite was used to implement an 99 data set, and testing was performed using data from
SSD-Mobilenet detector on the IoT devices and was a real-time IoT platform. The AI-based IDS result
trained with the MIO-TCD data set. A tracking algo- showed a reduction in energy overhead with regards to
rithm was deployed on the server end to track vehicles computation time to discover important features, and a
in parking facilities. The system was tested in a real- fast response time to intrusion compared to swarm
world environment for 3 months and the system intelligence-based BAT algorithm which lacks parallel
24 International Journal of Distributed Sensor Networks

processing and self-adaptive abilities. Eskandari et al.70 Declaration of conflicting interests

proposed an intelligent anomaly-based IDS termed as The author(s) declared no potential conflicts of interest with
Passban. The particular quality of the system is its respect to the research, authorship, and/or publication of this
direct deployment on low-cost IoT devices and its article.
platform-independent ability. The aim of the authors
was to ensure the protection of data very close to the
Funding
IoT data sources, so that data generated from several
IoT source can be analyzed to detect anomaly. The The author(s) received no financial support for the research,
authorship, and/or publication of this article.
IDS training phase was carried out on a normal net-
work flow. After the training, the learned model is
placed into the gateway’s internal memory where it is ORCID iDs
used to detect attacks from arriving network traffic. Kah Phooi Seng https://orcid.org/ 0000-0002-8071-9044
The system was evaluated against common cyber Li Minn Ang https://orcid.org/0000-0002-2402-7529
threats (SYN flood attacks, port scanning and brute Ericmoore Ngharamike https://orcid.org/0000-0002-3572-
force attacks) and demonstrated its efficiency. 9956
Trust recommendation is another area of impor-
tance for IoT security. Xu et al.71 proposed an References
approach termed as TTSVD for trust recommendation
1. Shi F, Ning H, Huangfu W, et al. Recent progress on the
in AI-enabled IoT systems which incorporates trust
convergence of the Internet of Things and artificial intel-
and rating information. In this work, the authors used ligence. IEEE Netw 2020; 34(5): 8–15.
a dual truster model (TrusterSVD) and trustee model 2. Yang L, Chen X, Perlaza SM, et al. Special issue on
(TrusteeSVD) based on a rating-only recommendation artificial-intelligence-powered edge computing for Inter-
model. The influence of the two models were combined net of Things. IEEE IoT J 2020; 7(10): 9224–9226.
by the weighted average to produce a hybrid model 3. Qian S, Sengupta S and Hansen LK. Active learning
called TT-SVD. Some other examples of AIoT security solution on distributed edge computing, 2019, https://
using various techniques can be found in the literature. arxiv.org/abs/1906.10718
Li et al.72 proposed an approach for resource optimiza- 4. Shao L, Zhu F and Li X. Transfer learning for visual
tion in blockchain IoT based on dueling DQN. Wang categorization: a survey. IEEE T Neu Netw Learn Syst
et al.73 proposed another approach using blockchains 2015; 26(5): 1019–1034.
5. Hao M, Li H, Lao X, et al. Efficient and privacy-
to address the issues for decentralized multiparty learn-
enhanced federated learning for industrial artificial intel-
ing in AIoT. Hu et al.74 and Lim et al.75 proposed ligence. IEEE T Ind Inform 2019; 16: 6532–6542.
using FL toward the problems for privacy and incen- 6. Tawalbeh LA, Muheidat F, Tawalbeh M, et al. IoT Pri-
tive mechanism design, respectively. Jia et al.76 pro- vacy and security: challenges and solutions. Appl Sci
posed an approach termed as FlowGuard to defend 2020; 10(12): 4102.
against IoT DDoS attacks. Libri et al.77 proposed an 7. Sodhro AH, Gurtov A, Zahid N, et al. Towards conver-
approach termed as pAElla for real-time malware gence of AI and IoT for energy efficient communication
detection in edge IoT devices. in smart homes. IEEE IoT J 2020; 8: 9664–9671.
8. Chen X, Shi Q, Yang L, et al. Thrifty edge: resource-
efficient edge computing for intelligent IoT applications.
Conclusion IEEE Netw 2018; 32(1): 61–65.
9. Fragkos G, Tsiropoulou EE and Papavassiliou S. Artifi-
This paper has reviewed and discussed the convergence cial intelligence enabled distributed edge computing for
of the AIoT, where the advantages of intelligent Internet of Things applications. In: Proceedings of the
machine-learning algorithms are integrated into 2020 16th international conference on distributed comput-
resource-constrained IoT sensors and devices to enable ing in sensor systems (DCOSS), Marina del Rey, CA,
large-scale and complex sensor deployments for IoT 25–27 May 2020, pp.450–457. New York: IEEE.
infrastructures. The paper has discussed the AIoT from 10. Gong C, Lin F, Gong X, et al. Intelligent cooperative
several aspects and layers (sensors and devices, commu- edge computing in Internet of Things. IEEE IoT J 2020;
nication, and networking and AIoT applications), dis- 7(10): 9372–9382.
11. Shi D, Gao H, Wang L, et al. Mean field game guided
cussed the combination of the AIoT with different
deep reinforcement learning for task placement in coop-
emerging technologies (e.g. edge, fog, and MEC com- erative multiaccess edge computing. IEEE IoT J 2020;
puting, SDN, 5G/6G cellular networks, etc.), and given 7(10): 9330–9340.
insights into the various challenges and issues which 12. Sanchez J, Sawant A, Neff C, et al. AWARE-CNN:
remain to be resolved to enable practical deployments automated workflow for application-aware real-time
of the AIoT into increasingly varied and complex edge acceleration of CNNs. IEEE IoT J 2020; 7(10):
environments. 9318–9329.
Seng et al. 25

13. Zhou J, Wang Y, Ota K, et al. AAIoT: accelerating artifi- 31. Ramı́rez PLG, Taha M, Lloret J, et al. An intelligent
cial intelligence in IoT systems. IEEE Wirel Commun Lett algorithm for resource sharing and self-management of
2019; 8(3): 825–828. wireless-IoT-gateway. IEEE Access 2019; 8: 3159–3170.
14. He W, Guo S, Guo S, et al. Joint DNN partition deploy- 32. Emir A, Kara F, Kaya H, et al. DeepMuD: multi-user
ment and resource allocation for delay-sensitive deep learn- detection for uplink grant-free NOMA IoT networks via
ing inference in IoT. IEEE IoT J 2020; 7(10): 9241–9254. deep learning. IEEE Wirel Commun Lett 2021; 10:
15. Yang K, Shi Y, Yu W, et al. Energy-efficient processing 1133–1137.
and robust wireless cooperative transmission for edge 33. Kwon M, Lee J and Park H. Intelligent IoT connectivity:
inference. IEEE IoT J 2020; 7(10): 9456–9470. deep reinforcement learning approach. IEEE Sens J 2019;
16. Liu X, Qiu T, Dai B, et al. Swarm-intelligence-based ren- 20(5): 2782–2791.
dezvous selection via edge computing for mobile sensor 34. Wang D, Chen D, Song B, et al. From IoT to 5G I-IoT:
networks. IEEE IoT J 2020; 7(10): 9471–9480. The next generation IoT-based intelligent algorithms and
17. Yang S, Xu K, Cui L, et al. EBI-PAI: towards an efficient 5G technologies. IEEE Commun Mag 2018; 56(10):
edge-based IoT platform for artificial intelligence. IEEE 114–120.
IoT J 2020; 8: 9580–9593. 35. Zhang L, Liang Y-C and Niyato D. 6G Visions: mobile
18. Liu D, Chen X, Zhou Z, et al. HierTrain: fast hierarchi- ultra-broadband, super Internet-of-Things, and artificial
cal edge AI learning with hybrid parallelism in mobile- intelligence. China Commun 2019; 16(8): 1–14.
edge-cloud computing. IEEE Open J Commun Soc 2020; 36. Tang F, Mao B, Fadlullah ZM, et al. On a novel deep-
1: 634–645. learning-based intelligent partially overlapping channel
19. Chiang Y-H, Chiang T-W, Zhang T, et al. Deep-dual- assignment in SDN-IoT. IEEE Commun Mag 2018;
learning-based cotask processing in multiaccess edge 56(9): 80–86.
computing systems. IEEE IoT J 2020; 7(10): 9383–9398. 37. He X, Wang K, Huang H, et al. Green resource alloca-
20. Foukalas F and Tziouvaras A. Edge AI for industrial IoT tion based on deep reinforcement learning in content-
applications. IEEE Ind Electr Mag 2021; 15: 28–36. centric IoT. IEEE T Emerg Top Comput 2018; 8(3):
21. Wang X, Wang C, Li X, et al. Federated deep reinforce- 781–796.
ment learning for Internet of Things with decentralized 38. Wu Q and Zhang R. Towards smart and reconfigurable
cooperative edge caching. IEEE IoT J 2020; 7(10): environment: intelligent reflecting surface aided wireless
9441–9455. network. IEEE Commun Mag 2020; 58(1): 106–112.
22. Zhou Z, Yang S, Pu L, et al. CEFL: online admission 39. Li J, Zhao Z, Li R, et al. AI-based two-stage intrusion
control, data scheduling, and accuracy tuning for cost- detection for software defined IoT networks. IEEE IoT J
efficient federated learning across edge nodes. IEEE IoT 2018; 6(2): 2093–2102.
J 2020; 7(10): 9341–9356. 40. Guo S, Qi Y, Yu P, et al. When network operation meets
23. Arisdakessian S, Wahab OA, Mourad A, et al. FoG- blockchain: an artificial-intelligence-driven customization
Match: an intelligent multi-criteria IoT-Fog scheduling service for trusted virtual resources of IoT. IEEE Netw
approach using game theory. IEEE/ACM T Netw 2020; 2020; 34(5): 46–53.
28(4): 1779–1789. 41. Jiang C, Gao L, Wang T, et al. Crowd-MECS: a novel
24. Lu H, He X, Du M, et al. Edge QoE: Computation off- crowdsourcing framework for mobile edge caching and
loading with deep reinforcement learning for Internet of sharing. IEEE IoT J 2020; 7(10): 9426–9440.
Things. IEEE IoT J 2020; 7(10): 9255–9265. 42. Jiang F, Wang KZ, Dong L, et al. Stacked autoencoder-
25. Wang A, Song C, Xu X, et al. Selective and compressive based deep reinforcement learning for online resource
sensing for energy-efficient implantable neural decoding. scheduling in large-scale MEC networks. IEEE IoT J
In: Proceedings of the 2015 IEEE biomedical circuits and 2020; 7(10): 9278–9290.
systems conference, Atlanta, GA, 22–24 October 2015, 43. Du J, Yu FR, Lu G, et al. MEC-assisted immersive VR
pp. 1–4. New York: IEEE.
video streaming over terahertz wireless networks: a deep
26. Xu J, Wang S, Zhou A, et al. Edgence: a blockchain-
reinforcement learning approach. IEEE IoT J 2020;
enabled edge-computing platform for intelligent iot-based
7(10): 9517–9529.
dapps. China Commun 2020; 17(4): 78–87.
44. Puri V, Jha S, Kumar R, et al. A hybrid artificial intelli-
27. Zhang J, Du J, Shen Y, et al. Dynamic computation off-
gence and Internet of Things model for generation of
loading with energy harvesting devices: a hybrid-decision-
renewable resource of energy. IEEE Access 2019; 7:
based deep reinforcement learning approach. IEEE IoT J
111181–111191.
2020; 7(10): 9303–9317.
45. Han T, Muhammad K, Hussain T, et al. An efficient deep
28. Wu AKB, Li J, Yang W, et al. Blockchain-based incen-
learning framework for intelligent energy management in
tive energy-knowledge trading in IoT: joint power trans-
IoT networks. IEEE IoT J 2020; 8: 3170–3179.
fer and AI design. IEEE IoT J. Epub ahead of print 15
46. Lei L, Tan Y, Dahlenburg G, et al. Dynamic energy dis-
September 2020. DOI: 10.1109/JIOT.2020.3024246.
patch based on deep reinforcement learning in IoT-driven
29. Song H, Bai J, Yi Y, et al. Artificial intelligence enabled
smart isolated microgrids. IEEE IoT J 2020; 8:
Internet of Things: network architecture and spectrum
7938–7953.
access. IEEE Comput Intel Mag 2020; 15(1): 44–51.
47. Liu Y, Yang C, Jiang L, et al. Intelligent edge computing
30. Yang K, Shi Y, Zhou Y, et al. Federated machine learn-
for IoT-based energy management in smart cities. IEEE
ing for intelligent IoT via reconfigurable intelligent sur-
Netw 2019; 33(2): 111–117.
face. IEEE Netw 2020; 34(5): 16–22.
26 International Journal of Distributed Sensor Networks

48. Miao Y, Zhou M and Ghoneim A. Blockchain and AI- and Aslekar A (eds) Digital entertainment. Singapore:
based natural gas industrial IoT system: architecture and Palgrave Macmillan, 2021, pp.87–109.
design issues. IEEE Netw 2020; 34(5): 84–90. 63. De Lima ES, Feijó B and Furtado AL. Adaptive story-
49. Lv Z, Han Y, Singh AK, et al. Trustworthiness in indus- telling based on personality and preference modeling.
trial IoT systems based on artificial intelligence. IEEE Entertain Comp 2020; 34: 100342.
Transactions on Industrial Informatics 2020; 17(2)): 64. Dilshad N, Hwang J, Song J, et al. Applications and
1496–1504. challenges in video surveillance via drone: a brief survey.
50. Zhang X, Pipattanasomporn M, Chen T, et al. An IoT- In: Proceedings of the 2020 international conference on
based thermal model learning framework for smart build- information and communication technology convergence
ings. IEEE IoT J 2019; 7(1): 518–527. (ICTC), Jeju, Korea, 21–23 October 2020, pp.728–732.
51. Jiang H. Mobile fire evacuation system for large public New York: IEEE.
buildings based on artificial intelligence and IoT. IEEE 65. Lee MFR and Chien TW. Artificial intelligence and
Access 2019; 7: 64101–64109. Internet of Things for robotic disaster response. In: Pro-
52. Xiao K, Shi W, Gao Z, et al. DAER: a resource preallo- ceedings of the 2020 international conference on advanced
cation algorithm of edge computing server by using robotics and intelligent systems (ARIS), Taipei, Taiwan,
blockchain in intelligent driving. IEEE IoT J 2020; 7(10): 19–21 August 2020, pp.1–6. New York: IEEE.
9291–9302. 66. Kim H, Ben-Othman J, Mokdad L, et al. CONTVERB:
53. Dass P, Misra S and Roy C. T-safe: trustworthy service continuous virtual emotion recognition using replaceable
provisioning for IoT-based intelligent transport systems. barriers for intelligent emotion-based IoT services and
IEEE T Veh Technol 2020; 69(9): 9509–9517. applications. IEEE Netw 2020; 34(5): 269–275.
54. Yang Z, Liang B and Ji W. An intelligent end-edge-cloud 67. Ke R, Zhuang Y, Pu Z, et al. A smart, efficient, and reli-
architecture for visual IoT assisted healthcare systems. able parking surveillance system with edge artificial intel-
IEEE IoT J 2021; 8: 16779–16786. ligence on IoT devices. IEEE T Intel Transp Syst 2020;
55. Mustafa A, Alahmed M, Alhammadi A, et al. Stress 99: 1–13.
detector system using IoT and artificial intelligence. In: 68. Chakrabarty S and Engels DW. Secure smart cities
Proceedings of the 2020 advances in science and engineer- framework using IoT and AI. In: Proceedings of the 2020
ing technology international conferences (ASET), Dubai, IEEE global conference on artificial intelligence and Inter-
United Arab Emirates, 4 February–9 April 2020, pp.1–6. net of Things (GCAIoT), Dubai, United Arab Emirates,
New York: IEEE. 12–16 December 2020, pp.1–6. New York: IEEE.
56. Chen WL, Lin YB, Ng FL, et al. RiceTalk: rice blast 69. Suresh GM and Madhavu ML. AI based intrusion detec-
detection using Internet of Things and artificial intelli- tion system using self-adaptive energy efficient BAT algo-
gence technologies. IEEE IoT J 2019; 7(2): 1001–1010. rithm for software defined IoT networks. In: Proceedings
57. Nurminen A and Malhi A. Green thumb engineering: of the 2020 11th international conference on computing,
artificial intelligence for managing IoT enabled house- communication and networking technologies (ICCCNT),
plants. In: Proceedings of the 2020 IEEE global conference Kharagpur, India, 1–3 July 2020, pp.1–6. New York:
on artificial intelligence and Internet of Things (GCAIoT), IEEE.
Dubai, United Arab Emirates, 12–16 December 2020, 70. Eskandari M, Janjua ZH, Vecchio M, et al. Passban IDS:
pp.1–7. New York: IEEE. an intelligent anomaly-based intrusion detection system
58. Chen WE, Lin YB and Chen LX. PigTalk: an AI-based for IoT edge devices. IEEE IoT J 2020; 7: 6882–6897.
IoT platform for piglet crushing mitigation. IEEE T Ind 71. Xu G, Zhao Y, Jiao L, et al. TT-SVD: an efficient sparse
Inform 2020; 17: 4345–4355. decision making model with two-way trust recommenda-
59. Wangoo DP and Reddy SN. Smart learning environ- tion in the AI enabled IoT systems. IEEE IoT J 2020; 8:
ments framework for educational applications in IoT 9559–9567.
enabled educational ecosystems: a review on AI based 72. Li M, Yu FR, Si P, et al. Resource optimization for
GUI tools for IoT wearables. In: Proceedings of the 2020 delay-tolerant data in blockchain-enabled iot with edge
IEEE 17th India council international conference (INDI- computing: a deep reinforcement learning approach.
CON), New Delhi, India, 10–13 December 2020, pp.1–8. IEEE IoT J 2020; 7(10): 9399–9412.
New York: IEEE. 73. Wang Q, Guo Y, Wang X, et al. Ai at the edge:
60. Liu CY, Yu HY and Tsai IC. Applying AI analysis-based blockchain-empowered secure multiparty learning with
IoT system control to the individualized learning field. In: heterogeneous models. IEEE IoT J 2020; 7(10):
Proceedings of the 2019 IEEE international conference on 9600–9610.
engineering, technology and education (TALE), Yogya- 74. Hu R, Guo Y, Li H, et al. Personalized federated learn-
karta, Indonesia, 10–13 December 2019, pp.1–7. New ing with differential privacy. IEEE IoT J 2020; 7(10):
York: IEEE. 9530–9539.
61. Zhao S, Chen S, Li W, et al. The development of artificial 75. Lim WYB, Xiong Z, Miao C, et al. Vincent poor: ‘‘hier-
intelligence education resources under the background of archical incentive mechanism design for federated
the Internet of Things. In: Proceedings of the 2020 Chinese machine learning in mobile networks.’’IEEE IoT J 2020;
control and decision conference (CCDC), Hefei, China, 7(10): 9575–9588.
22–24 August 2020, pp.2327–2332. New York: IEEE. 76. Jia Y, Zhong F, Alrawais A, et al. Flowguard: an intelli-
62. Hallur GG, Prabhu S and Aslekar A. Entertainment in gent edge defense mechanism against IoT DDoS attacks.
era of AI, big data & IoT. In: Hallur GG, Prabhu S IEEE IoT J 2020; 7(10): 9552–9562.
Seng et al. 27

77. Libri A, Bartolini A and Benini L. pAElla: edge AI-based 86. Mastorakis G, Mavromoustakis CX, Batalla JM, et al.
real-time malware detection in data centers. IEEE IoT J (eds). Convergence of artificial intelligence and the Internet
2020; 7(10): 9589–9599. of Things. Berlin; Springer, 2020.
78. Teerapittayanon S, McDanel B and Kung HT. Distribu- 87. Antonopoulos I, Robu V, Couraud B, et al. Artificial
ted deep neural networks over the cloud, the edge and intelligence and machine learning approaches to energy
end devices. In: Proceedings of the 2017 IEEE 37th inter- demand-side response: a systematic review. Renew Sust
national conference on distributed computing systems Energ Rev 2020; 130: 109899.
(ICDCS), Atlanta, GA, 5–8 June 2017, pp.328–339. 88. Trakadas P, Simoens P, Gkonis P, et al. An artificial
New York: IEEE. intelligence-based collaboration approach in industrial
79. Savaglio C, Gerace P, Di Fatta G, et al. Data mining at IoT manufacturing: key concepts, architectural extensions
the IoT edge. In: Proceedings of the 2019 28th interna- and potential applications. Sensors 2020; 20(19): 5480.
tional conference on computer communication and net- 89. Zhang J and Letaief KB. Mobile edge intelligence and
works (ICCCN), Valencia, 29 July–1 August 2019, pp.1– computing for the internet of vehicles. Proc IEEE 2019;
6. New York: IEEE. 108: 246–261.
80. Guin U, Cui P and Skjellum A. Ensuring proof-of- 90. Sun L, Jiang X, Ren H, et al. Edge-cloud computing and
authenticity of iot edge devices using blockchain tech- artificial intelligence in internet of medical things: archi-
nology. In: Proceedings of the 2018 IEEE international tecture, technology and application. IEEE Access 2020; 8:
conference on Internet of Things (iThings) and IEEE 101079–101092.
green computing and communications (GreenCom) and 91. Hu L, Miao Y, Wu G, et al. iRobot-factory: an intelli-
IEEE cyber, physical and social computing (CPSCom) gent robot factory based on cognitive manufacturing
and IEEE smart data (SmartData), Halifax, NS, 30 and edge computing. Fut Gen Comp Syst 2019; 90:
July–3 August, pp.1042–1049. New York: IEEE. 569–577.
81. Mukhopadhyay SC, Tyagi SKS, Suryadevara NK, et al. 92. Li X, Li W, Yang Q, et al. Edge-computing-enabled
Artificial intelligence-based sensors for next generation unmanned module defect detection and diagnosis system
IoT applications: a review. IEEE Sens J 2021; 21: for large-scale photovoltaic plants. IEEE IoT J 2020;
24920–24932. 7(10): 9651–9663.
82. Ang LM, Seng KP, Ijemaru GK, et al. Deployment of 93. Velasco-Montero D, Femández-Bemi J, Carmona-Gálán
IoV for smart cities: applications, architecture, and chal- R, et al. On the correlation of CNN performance and
lenges. IEEE Access 2018; 7: 6473–6492. hardware metrics for visual inference on a low-cost CPU-
83. Ijemaru GK, Ang KLM and Seng JK. Mobile collectors for based platform. In: Proceedings of the 2019 international
opportunistic Internet of Things in smart city environment conference on systems, signals and image processing (IWS-
with wireless power transfer. Electronics 2021; 10(6): 697. SIP), Osijek, 5–7 June 2019, pp.249–254. New York:
84. Ijaz A, Shahabuddin S, Sauter T, et al. Challenges of AI IEEE.
in wireless networks for IoT. IEEE Ind Electr Mag 2021; 94. Chiu T-C, Shih Y-Y, Pang A-C, et al. Semisupervised
15: 16–29. distributed learning with non-IID data for AIoT service
85. Danzi P, Kalor AE, Sorensen RB, et al. Communication platform. IEEE IoT J 2020; 7(10): 9266–9277.
aspects of the integration of wireless iot devices with dis- 95. Wu H, Han H, Wang X, et al. Research on artificial
tributed ledger technology. IEEE Netw 2020; 34(1): intelligence enhancing Internet of Things security: a
47–53. survey. IEEE Access 2020; 8: 153826–153848.