Full Text
1. Introduction
The Internet of Things has changed the scale and character of digital
information production. Sensors embedded in buildings, factories,
vehicles, agricultural equipment, medical devices and public
infrastructure continuously observe physical conditions and convert them
into data. These devices support monitoring, automation, prediction and
remote control. Their value often depends on the ability to process
information quickly enough to influence events while those events are
still occurring. A connected industrial machine, for example, may
require a response to vibration or temperature changes within a short
operational window. A vehicle safety system cannot always wait for a
remote cloud service to interpret sensor readings and return a
command.
Cloud computing has provided an important foundation for IoT by
offering elastic storage, centralized analytics and access to large
computational resources. However, cloud-centric processing is not
optimal for every IoT workload. Sending all sensor data to distant data
centres can consume network capacity, increase response time and expose
sensitive information during transmission and storage. Connectivity may
also be unreliable in rural areas, underground facilities, moving
vehicles and emergency situations. As the number of connected endpoints
grows, a model in which every event is transmitted to a centralized
platform can become costly and operationally fragile.
Edge computing addresses these limitations by placing computation and
storage near the sources or consumers of data. The term edge can refer
to a device, gateway, base station, local server, industrial controller
or micro-data centre, depending on the architecture. The defining idea
is not that all processing must occur on a sensor. Rather, suitable
tasks are moved from centralized cloud infrastructure to one or more
intermediate or local computing locations. Satyanarayanan (2017)
described edge computing as an architectural response to the limitations
of cloud computing for emerging applications that require low latency,
mobility and context awareness.
Several related concepts have developed around this idea. Fog
computing emphasizes a distributed layer between cloud data centres and
smart end devices. Cisco's early description of fog computing presented
a highly virtualized platform that brings computation, storage and
networking closer to IoT devices (Bonomi et al., 2012). Mobile edge
computing, later commonly referred to as multi-access edge computing,
places cloud-like capabilities within or near mobile access networks.
Cloudlets describe resource-rich computers located close to mobile
users, allowing applications to offload demanding computation to nearby
infrastructure (Satyanarayanan et al., 2009). These approaches overlap
but differ in their historical origins, network assumptions and
deployment environments.
The relationship between edge computing and IoT is particularly
important because IoT creates distributed data sources while edge
computing provides a way to process those data locally or regionally.
Together they support responsive applications in smart manufacturing,
healthcare, transport, energy, agriculture and urban management. Yet
distributing computation also distributes risk. Edge nodes may be
physically accessible to attackers, managed by different organizations,
connected through heterogeneous networks and updated less consistently
than centralized servers. Security, interoperability and lifecycle
management are therefore central architectural concerns, not secondary
additions.
This article reviews edge computing as an enabling architecture for
IoT and examines the implications for information systems and
engineering. It addresses five questions: (1) What architectural
concepts and technologies underpin edge-IoT systems? (2) How should
workloads be distributed among devices, edge nodes and cloud platforms?
(3) Which applications benefit most from edge processing? (4) What
security, privacy, operational and economic challenges arise? (5) What
design principles can guide reliable deployments? The review focuses on
literature available by 31 December 2022, framing the discussion around
developments leading into 2023.
2. Review Methodology
An integrative literature review was adopted to synthesize research
across computer networking, distributed systems, mobile computing,
embedded systems, cybersecurity and IoT applications. Edge computing is
not a single technology or standardized architecture. It includes
multiple deployment patterns, resource models and service environments.
An integrative approach allows foundational architectural papers to be
considered alongside surveys, experimental studies, standards-oriented
work and application research.
The temporal scope was restricted to publications available on or
before 31 December 2022. Search concepts included edge computing, fog
computing, cloudlets, mobile edge computing, multi-access edge
computing, IoT architecture, computation offloading, edge intelligence,
federated learning, edge security, resource orchestration and
distributed data processing. The review prioritized peer-reviewed
journal articles and conference proceedings, seminal technical papers
and recognized standards or institutional publications with identifiable
publication information.
The literature was grouped into six themes: conceptual foundations
and architecture; workload placement and resource management; networking
and enabling technologies; application domains; security and privacy;
and implementation challenges and future research. Within each theme,
the review compares the assumptions and trade-offs of alternative
approaches. Particular attention was given to the distinction between
edge, fog and cloud computing, because inconsistent terminology can lead
to inaccurate system comparisons.
The article is a conceptual synthesis rather than a systematic
meta-analysis. The reviewed studies use different hardware, network
conditions, application workloads and performance measures, so pooled
numerical estimates would be misleading. The review does not claim to
exhaust the rapidly expanding literature. Instead, it selects
influential sources that establish the main architectural principles,
technical challenges and research directions relevant to edge-enabled
IoT.
Cloud computing centralizes substantial computing and storage
capacity in large data centres that users access through networks. Its
strengths include elasticity, economies of scale, centralized
administration and access to specialized hardware. The National
Institute of Standards and Technology defines cloud computing through
characteristics such as on-demand self-service, broad network access,
resource pooling, rapid elasticity and measured service (Mell &
Grance, 2011). These characteristics remain valuable for IoT
applications that require long-term storage, large-scale analytics,
model training and coordination across many locations.
Edge computing shifts selected functions toward the network edge. Shi
et al. (2016) identified edge computing as a way to support applications
with low latency, location awareness and context sensitivity. The edge
is not necessarily a single physical boundary. In a cellular network it
may be located near a radio access network; in a factory it may be an
industrial computer connected to controllers; in a home it may be a
gateway or local hub. The placement depends on the application, network
topology, ownership and available resources.
Fog computing is often used to describe a hierarchical or distributed
computing layer between cloud services and end devices. Bonomi et al.
(2012) introduced fog computing as a platform for IoT and emphasized
distribution, proximity, location awareness and support for
heterogeneous devices. In practice, fog nodes may include gateways,
routers, switches, access points and local servers. Fog and edge are
sometimes used interchangeably, although fog is often associated with a
broader continuum of distributed nodes rather than only the immediate
network edge.
Cloudlets were proposed as trusted, resource-rich computers close to
mobile users. Satyanarayanan et al. (2009) presented cloudlets as a way
to overcome the limitations of mobile devices by offloading computation
to nearby infrastructure. A cloudlet may be located in a building,
campus, public venue or enterprise environment. The concept is
especially relevant to interactive applications that require more
compute capacity than a mobile device can provide but cannot tolerate
the delay of a distant data centre.
Mobile edge computing emerged from the need to place cloud
capabilities within mobile networks. The European Telecommunications
Standards Institute (ETSI) developed a mobile edge computing
architecture and later broadened the concept to multi-access edge
computing, reflecting its applicability beyond cellular networks. MEC
can provide application hosting, radio network information and
low-latency access to services near users. It is relevant to connected
vehicles, augmented reality, industrial wireless systems and
location-aware applications.
These paradigms can be understood as points on a distributed
computing continuum. At the device layer, sensors and embedded
controllers perform acquisition and immediate control. At the local
edge, gateways and industrial servers aggregate data, run inference and
coordinate nearby devices. At the regional edge, operator or enterprise
infrastructure hosts applications shared across a wider area. The cloud
provides centralized orchestration, large-scale analytics, model
training and long-term storage. Workloads may move between these layers
as conditions change.
The continuum perspective avoids the false choice between edge and
cloud. Many systems require both. A local controller may detect an
unsafe machine condition and stop equipment immediately, while the cloud
later aggregates historical data to identify patterns across multiple
factories. A vehicle may process sensor data locally for collision
avoidance while uploading selected events for fleet analysis. The
architecture should allocate each task to the location that best
satisfies its latency, bandwidth, privacy, availability, energy and cost
requirements.
4. Edge-IoT Reference
Architecture
An edge-IoT architecture can be represented as a set of cooperating
layers. The device and perception layer contains sensors, actuators,
cameras, meters, wearables and embedded controllers. These components
observe the physical environment and may perform filtering, event
detection or simple control. Device constraints include limited energy,
memory, processing power and physical protection. The communication
layer connects devices to gateways and edge infrastructure through
technologies such as Ethernet, Wi-Fi, cellular networks, low-power
wide-area networks, Bluetooth and industrial fieldbuses.
The gateway and aggregation layer provides protocol translation,
device onboarding, local buffering and traffic management. Gateways may
normalize data from heterogeneous devices and provide a point for
enforcing security policies. They can also perform preliminary
analytics, such as threshold detection, deduplication and compression.
Gateway design is important because poorly secured gateways can become a
central point of compromise for many connected devices.
The edge computing layer hosts applications and services close to
data sources. It may include containerized workloads, virtual machines,
message brokers, databases, inference engines and local orchestration
agents. Edge nodes are often resource constrained compared with cloud
data centres, and their capacity may vary significantly across sites.
Applications should therefore be designed for graceful degradation,
resource limits and intermittent connectivity.
The cloud and platform layer supports fleet-wide management,
long-term data retention, large-scale analytics, model training,
software distribution and coordination across edge locations. It can
provide centralized identity services, policy management and
observability. Cloud services should not be assumed to be continuously
reachable. Edge nodes need secure local operating modes and well-defined
synchronization behaviour when the network becomes available again.
The application and service layer exposes capabilities to users and
business processes. Applications may include predictive maintenance
dashboards, traffic management, remote monitoring, energy optimization
and automated inspection. The management and security plane cuts across
all layers, providing identity, authentication, authorization,
encryption, logging, configuration, software updates, vulnerability
management and incident response. Security should be designed as a
cross-layer function because threats can originate from devices,
networks, edge hosts, cloud services or the interfaces between them.
Figure 1. Conceptual edge-IoT architecture
Physical devices and sensors
↓
Connectivity and protocol adaptation
↓
Gateways and local aggregation
↓
Edge compute: filtering, inference, control and local storage
↕
Regional edge: shared services and coordination
↕
Cloud platform: fleet management, model training and long-term
analytics
↓
Applications, operators and business processes
Security, identity, monitoring and governance operate across every
layer.
The architecture should support both north-south and east-west
communication. North-south traffic moves between devices, edge
infrastructure and cloud services. East-west traffic occurs among peer
devices or edge nodes. In some applications, peer coordination reduces
dependence on centralized services. However, distributed communication
increases the need for consistent identity, policy and trust management.
The system must define which nodes may exchange data, which node is
authoritative for a particular decision and how conflicting states are
reconciled.
Interoperability is a major design requirement. IoT ecosystems
contain devices from different manufacturers, using different protocols,
data formats and security capabilities. Edge platforms may need to
translate between industrial protocols, IP networks and cloud APIs.
Standardized data models and interfaces can reduce integration costs,
but standards adoption is uneven. Organizations should document protocol
support, device lifecycle expectations and the implications of
vendor-specific extensions before selecting an edge platform.
5.
Workload Placement, Offloading and Resource Orchestration
Workload placement is the process of deciding where a computation
should execute. A task may run on the device that generated the data, on
a nearby gateway, on a regional edge server or in a cloud data centre.
The decision depends on several variables: processing demand, memory
requirements, response-time deadline, network bandwidth, data
sensitivity, energy availability, reliability and operating cost. A task
with a strict latency requirement may need local execution, while a
batch analytics job may be more efficient in the cloud.
Computation offloading allows a resource-constrained device to
transfer some processing to a more capable node. The offloading decision
is not always beneficial. Transmitting input data and receiving results
consume time and energy. If the network is congested or unstable, local
execution may be faster. If the edge node is overloaded, the task may
experience queuing delay. A useful offloading policy estimates total
execution time, including communication, queueing and computation,
rather than comparing processor speeds alone.
Mobile edge environments add mobility and handover to the problem. A
user or device may move between access points or cells while an
application is running. The edge service may need to migrate state or
redirect requests to a new location. Mobility management is especially
important for connected vehicles, augmented reality and industrial
assets moving across large sites. Service continuity depends on the
ability to maintain session state, preserve security context and avoid
unnecessary application interruption.
Resource orchestration coordinates CPU, memory, storage, network and
accelerator resources across multiple edge nodes. Container platforms
can simplify deployment and portability, but edge environments may not
have the capacity or reliability of cloud clusters. Orchestration
systems must account for intermittent connectivity, heterogeneous
processor architectures, limited power and local operational
constraints. Lightweight agents and hierarchical control may be more
suitable than a single centralized controller for large deployments.
Task scheduling may use static rules, heuristics, optimization models
or machine learning. Static policies are easier to audit and can be
effective when workloads are predictable. Optimization approaches can
model latency, energy and resource constraints, but may be
computationally expensive or depend on accurate estimates.
Learning-based schedulers can adapt to changing conditions, but their
behaviour may be difficult to validate in safety-critical systems.
Hybrid approaches can combine hard constraints with adaptive
decisions.
Quality of service should be specified in terms that reflect the
application. Relevant measures include end-to-end latency, jitter,
throughput, packet loss, availability, deadline-miss ratio, energy
consumption and service recovery time. Average latency alone may conceal
dangerous tail behaviour. For an industrial control task, the 99th
percentile or worst-case response time may be more important than the
mean. The service objective should be selected with domain experts and
tested under realistic peak loads and network failures.
Data placement is closely related to workload placement. Some
information can be processed locally and discarded after an event has
been identified; other information may need to be retained for audit,
safety or regulatory reasons. Data may be aggregated before
transmission, reducing bandwidth while preserving the information needed
for analysis. However, aggregation can remove details required for later
investigation. Data lifecycle policies should therefore define what is
retained locally, what is sent to the cloud, how long it is stored and
how deletion is verified.
The most appropriate architecture is often adaptive. A system may
execute inference locally during a network outage, synchronize summaries
when connectivity returns and use cloud resources for periodic model
updates. A hospital may prioritize local operation for monitoring
devices while using centralized infrastructure for population-level
analytics. An industrial facility may retain immediate safety control
within certified local systems and use edge analytics for optimization.
The allocation should reflect the consequences of failure, not only
computational efficiency.
6. Enabling Technologies
IoT connectivity technologies provide the communication foundation
for edge systems. Wi-Fi and Ethernet are common in buildings and
industrial facilities, while cellular networks provide wide-area
mobility and managed quality-of-service capabilities. Low-power
wide-area technologies support devices that transmit small amounts of
data over long distances with low energy consumption. Bluetooth and
other short-range technologies connect wearables, sensors and personal
devices. No single connectivity option is ideal for all IoT workloads;
selection depends on range, throughput, latency, energy, mobility,
spectrum and deployment cost.
The emergence of 5G networks has increased interest in edge computing
because of the potential for high data rates, lower latency and network
slicing. 5G does not automatically guarantee end-to-end low latency.
Performance depends on radio conditions, core network design,
application placement, congestion, backhaul and device capabilities.
Edge infrastructure near the radio access network can reduce the
distance to computation, but the application must be engineered to use
that proximity effectively.
Virtualization and containerization allow applications to be packaged
and deployed across heterogeneous infrastructure. Virtual machines
provide strong isolation and support different operating systems, but
they can have higher resource overhead. Containers are lightweight and
support rapid deployment, though their security depends on kernel
isolation, configuration and runtime controls. At the edge, deployment
systems should account for limited storage, intermittent connectivity
and the need to roll back failed updates.
Message-oriented middleware supports asynchronous communication
between devices and applications. Protocols such as MQTT and AMQP are
used in IoT environments to transport messages between producers and
consumers. CoAP is designed for constrained devices and networks. The
selection of a protocol should consider reliability, security, message
size, delivery semantics, device capability and interoperability. A
broker located at the edge can buffer messages and reduce the need for
every device to maintain a direct cloud connection.
Time-series databases and stream-processing systems are important
because many IoT workloads involve continuous measurements. Edge
analytics may perform windowing, aggregation, anomaly detection and
event correlation. Stream processing can reduce the volume of data sent
to the cloud by transmitting events or summaries rather than every raw
observation. Yet systems must preserve the ability to inspect raw data
when required for diagnosis, model retraining or regulatory
evidence.
Specialized hardware can improve edge performance. Graphics
processing units, field-programmable gate arrays, application-specific
integrated circuits and neural processing units can accelerate image
analysis, signal processing and machine learning inference. The benefit
must be balanced against cost, power, software support and hardware
availability. Heterogeneous edge deployments create portability
challenges because models and applications may need to be optimized for
different processor architectures.
Artificial intelligence at the edge, often called edge intelligence,
combines local data processing with machine learning. It includes local
inference, distributed training, model partitioning and collaborative
learning. Mao et al. (2017) surveyed mobile edge computing and
identified the relationship between edge infrastructure and
computation-intensive applications. Later work has examined how machine
learning can support resource management, anomaly detection and adaptive
network operation. The location of model training and inference should
be chosen according to data volume, latency, privacy and computational
capacity.
Federated learning is one approach to collaborative model training
without directly centralizing all raw training data. McMahan et al.
(2017) introduced a communication-efficient method for training deep
networks from decentralized data. In an edge-IoT context, devices or
edge nodes can train local model updates and share selected parameters
with a coordinating server. Federated learning can reduce raw data
transfer, but it does not automatically guarantee privacy. Model updates
may leak information, participants may be compromised and
non-identically distributed data can affect model convergence and
fairness.
7. Application Domains
Smart cities combine sensors, cameras, connected transport systems,
environmental monitors, smart lighting, public utilities and citizen
services. Edge computing can support traffic signal coordination, video
analytics, public safety monitoring, waste management and environmental
sensing. Processing video near cameras can reduce the volume of raw
footage transmitted to centralized data centres and can enable faster
local alerts. City-wide applications still require coordination across
districts, departments and service providers. Governance must address
surveillance, retention, public accountability and the risk of unequal
monitoring across communities.
Healthcare IoT includes wearable devices, bedside monitors, connected
imaging equipment, remote patient monitoring and smart hospital
infrastructure. Edge processing can support local alerting and reduce
dependence on continuous wide-area connectivity. A wearable may detect a
potentially abnormal pattern and send a compact alert to a clinician,
while raw measurements are retained according to clinical requirements.
Healthcare deployments must prioritize safety, data integrity,
confidentiality and clear responsibility for responding to alerts. Edge
analytics should be validated for the patient population and operating
conditions in which they are used.
Smart manufacturing uses industrial sensors, programmable logic
controllers, robots, machine vision and production management systems.
Edge computing can support predictive maintenance, quality inspection,
process optimization and coordination among machines. Local processing
can reduce latency and preserve operation when external connectivity is
disrupted. However, industrial systems often include legacy equipment
and operational technology that cannot be updated easily. Edge
integration should preserve safety controls and avoid creating new
pathways between enterprise IT networks and critical control
systems.
Transportation and connected vehicles generate data from cameras,
radar, lidar, engine sensors, positioning systems and
vehicle-to-everything communication. Edge computing can support object
detection, traffic coordination, fleet monitoring and local decision
support. Vehicle safety functions require strict timing and dependable
operation. A vehicle should not rely on a distant cloud service for a
time-critical safety response. Cloud infrastructure can support
fleet-wide learning and route optimization, while local systems retain
responsibility for immediate control.
Agriculture increasingly uses soil sensors, weather stations, drones,
irrigation controllers, livestock monitoring and satellite data. Edge
processing can help farmers act on local conditions even when broadband
connectivity is limited. A field gateway may aggregate soil moisture
readings and control irrigation according to local thresholds, while
selected summaries are transmitted for seasonal analysis. Agricultural
systems should be designed for harsh environmental conditions, low
maintenance availability and intermittent power. Cost and usability are
especially important for small farms.
Energy systems use smart meters, distributed generation, battery
storage, grid sensors and building management systems. Edge computing
can support local load balancing, fault detection, demand response and
coordination of distributed energy resources. Local decisions may
improve resilience when communication with central systems is delayed.
However, energy infrastructure is critical and requires strong
authentication, secure updates, reliable time synchronization and
rigorous testing. An erroneous control action can have consequences
beyond a single device.
Retail and logistics use connected shelves, inventory sensors,
cameras, point-of-sale systems, vehicle trackers and warehouse
automation. Edge processing can support real-time stock monitoring,
local personalization, robotic navigation and loss prevention.
Processing data locally may reduce latency and limit unnecessary
transfer of customer information. Retail applications also raise privacy
concerns, particularly when video analytics or behavioural tracking are
used. Organizations should define transparent purposes, retention
periods and access controls.
Environmental monitoring can use distributed sensors to measure air
quality, water quality, temperature, noise and hazardous conditions.
Edge nodes can identify threshold events, compress observations and
continue local monitoring during network outages. Such systems are
valuable in remote or resource-constrained settings. Data quality
remains a major concern because sensor calibration, placement, weather
and maintenance affect measurement accuracy. Edge analytics should
retain metadata about calibration and device status so that
decision-makers can assess the reliability of observations.
8. Security and Privacy
Challenges
Edge-IoT systems expand the attack surface because they contain many
devices, communication links, gateways, edge servers, cloud interfaces
and management services. A weakness in one component may provide a route
to other components. The number of endpoints makes asset inventory,
patching and configuration management difficult. Some devices operate
for many years and may use outdated software or proprietary protocols.
Security planning must therefore begin with the device lifecycle and
extend through decommissioning and secure disposal.
Device identity and authentication are foundational. Each device
should have a unique identity and should not rely on shared default
passwords or unauthenticated network access. Mutual authentication can
help ensure that devices communicate with legitimate gateways and
services. Credentials need secure provisioning, rotation and revocation.
Hardware-backed keys and secure elements can improve protection, but
cost and device capability may limit their availability. Organizations
should document how identity is established and what happens when a
device is lost, compromised or transferred.
Communication security requires encryption, integrity protection and
appropriate key management. Data may travel across wireless links,
public networks and organizational boundaries. Encryption in transit
protects against some interception threats, but it does not protect
compromised endpoints or prevent misuse by authorized services.
Sensitive information may also require encryption at rest and access
controls within edge storage. Security mechanisms should be selected
with attention to constrained-device performance and the consequences of
certificate or key expiration.
Physical security is a distinctive edge challenge. Edge nodes may be
installed in roadside cabinets, factories, utility substations, vehicles
or public buildings. Attackers may gain physical access, remove storage,
replace hardware or connect debugging interfaces. Tamper-resistant
enclosures, secure boot, disk encryption, hardware roots of trust and
physical monitoring can reduce risk. Yet no single measure is
sufficient. Sensitive deployments should consider the possibility that
an edge node will be captured and should limit the damage that a
compromised node can cause.
Software supply-chain security is critical because edge platforms
depend on operating systems, container images, libraries, firmware,
orchestration agents and vendor services. Vulnerabilities may enter
through third-party components or insecure update channels. Secure
software updates should authenticate packages, verify integrity, support
rollback and record deployment status. Organizations need an inventory
of software components and a process for responding to newly disclosed
vulnerabilities. Unsupported devices should be isolated, replaced or
subject to documented compensating controls.
Network segmentation can limit lateral movement between IoT devices,
edge services and enterprise systems. Devices should be grouped
according to function and trust level, and communication should be
restricted to required services. A compromised camera should not
automatically provide access to industrial controllers or sensitive
business databases. Zero-trust principles can strengthen this model by
requiring continuous verification of identity and authorization rather
than assuming that a device is trustworthy because it is inside a
network perimeter.
Privacy risks arise because IoT devices may collect location, health,
behavioural, biometric, household and operational data. Edge processing
can reduce the need to transmit raw data to remote services, but local
processing does not automatically make a system privacy-preserving. Data
can still be accessed by unauthorized users, retained too long or
combined with other datasets. Privacy-by-design requires purpose
limitation, data minimization, retention controls, transparency and
appropriate legal grounds for processing.
Federated learning is sometimes proposed as a privacy-enhancing
approach because raw data can remain at local nodes. However, model
updates may reveal information about training examples, and compromised
participants can poison the global model. Secure aggregation,
differential privacy, robust aggregation and participant authentication
can reduce some threats, but each introduces trade-offs in accuracy,
communication and computation. Kairouz et al. (2021) reviewed the
opportunities and challenges of federated learning and emphasized that
privacy, systems heterogeneity and statistical heterogeneity remain
central research problems.
Denial-of-service attacks can disrupt edge services by exhausting
bandwidth, CPU, memory or storage. IoT botnets can use insecure devices
to generate large volumes of traffic. Edge systems should include rate
limiting, anomaly detection, resource quotas, traffic filtering and
recovery procedures. Local safety functions should remain isolated from
non-essential services so that a failure in a user-facing analytics
application does not disable critical control. Redundancy and graceful
degradation are important for applications that must continue operating
during partial failures.
The security of machine learning models is another concern.
Adversarial inputs may cause misclassification, while data poisoning can
influence training. Model extraction and inference attacks may reveal
proprietary models or sensitive training information. Edge devices are
often exposed to physical manipulation and may lack the capacity for
heavyweight security monitoring. Model integrity checks, trusted
execution where appropriate, robust validation and secure model
distribution should be incorporated into the deployment
architecture.
9.
Operational, Economic and Environmental Challenges
Heterogeneity complicates edge management. Devices vary in processor
architecture, memory, storage, operating system, network capability and
power supply. Edge sites may include hardware from several vendors and
may be managed by different teams. A platform that works well in a cloud
data centre may be too resource intensive for a gateway or industrial
controller. Organizations should define a supported hardware baseline,
test applications on representative devices and avoid assuming that
containerization alone guarantees portability.
Scalability is not simply a matter of adding more nodes. Large edge
deployments require device discovery, provisioning, configuration,
monitoring, software distribution, policy enforcement and fault
management across many locations. Centralized management can become a
bottleneck or single point of failure. Hierarchical management and local
autonomy can improve resilience, but they introduce consistency and
coordination challenges. Systems should define how local decisions are
reconciled with central policy after connectivity is restored.
Reliability depends on more than network availability. Power failure,
hardware degradation, environmental exposure, software defects, storage
exhaustion and clock drift can all affect service. Edge nodes may
operate in locations where physical maintenance is expensive. Remote
diagnostics and automated recovery can reduce operating cost, but they
must be designed securely. Critical systems should have tested fallback
modes and should not rely on an edge node that cannot be repaired or
replaced within an acceptable period.
Economic evaluation should include capital expenditure, connectivity,
power, cooling, hardware replacement, software licensing, security
operations, maintenance and staff training. Edge computing may reduce
cloud data transfer and improve service responsiveness, but it can also
increase the number of managed assets and sites. A
total-cost-of-ownership analysis should compare the full edge-cloud
design with a cloud-only alternative and should include the cost of
failures and service interruptions. The cheapest hardware configuration
may not be the least expensive system over its lifecycle.
Energy efficiency is important because edge deployments may involve
thousands of devices and distributed servers. Local processing can
reduce communication energy, but additional computation also consumes
power. Hardware accelerators, workload scheduling, sleep modes, model
compression and efficient cooling can reduce energy demand. The optimal
design depends on the balance between communication and computation
costs. Pushing every task to the edge is not inherently sustainable;
workloads should be placed where total resource use is lowest while
meeting service requirements.
Vendor lock-in can arise when applications depend on proprietary edge
orchestration, device management or cloud interfaces. Interoperability
standards and container-based deployment can improve portability, but
they do not remove dependencies on proprietary hardware, data formats or
management APIs. Procurement should examine data export, workload
portability, lifecycle support and the ability to operate if a vendor
discontinues a service. Open interfaces can reduce switching costs,
although they may require additional integration work.
Skills shortages are another constraint. Edge-IoT deployments require
knowledge of networking, embedded systems, cloud platforms,
cybersecurity, data engineering and the application domain.
Organizations may have separate teams responsible for operational
technology and information technology, with different priorities and
change-control procedures. Successful deployment requires shared
responsibility, clear escalation routes and training that helps teams
understand the complete system rather than only their individual
components.
10. Design
Principles and Proposed Deployment Framework
A practical deployment framework should begin with the application
requirement rather than with a preferred technology. Teams should define
the event to be detected, the decision to be supported, the maximum
acceptable response time, data sensitivity, availability target and
consequence of failure. These requirements determine whether processing
should occur on the device, local edge, regional edge or cloud. A
low-latency safety function may require local execution, while
historical trend analysis may be better suited to cloud
infrastructure.
The second step is workload decomposition. Applications should be
divided into sensing, filtering, inference, control, storage,
visualization and fleet analytics tasks. Each task can then be assigned
to a suitable layer. Data that must remain local can be processed or
summarized at the edge. Large model training and cross-site analytics
may use cloud resources. Critical control loops should be separated from
non-critical analytics and should have defined fallback behaviour.
The third step is security and privacy design. Teams should create a
threat model that identifies assets, trust boundaries, attack paths and
likely adversaries. Device identity, secure boot, encrypted
communication, access control, segmentation, logging and update
mechanisms should be specified before deployment. Privacy requirements
should define what data are collected, why they are needed, where they
are processed, how long they are retained and who can access them.
The fourth step is platform and interoperability selection.
Organizations should evaluate hardware support, operating-system
compatibility, container runtime, orchestration, message protocols,
device management, monitoring and vendor lifecycle commitments. The
platform should be tested under realistic network conditions, including
latency variation, packet loss, intermittent connectivity and node
failure. Benchmarks should use representative workloads rather than
synthetic tests alone.
The fifth step is pilot deployment and operational validation. A
pilot should include normal operating conditions, peak demand, network
outages, security incidents and recovery exercises. Teams should measure
end-to-end latency, availability, energy consumption, bandwidth,
resource utilization, update success and time to recover. Domain
specialists should validate whether the system's decisions are
operationally meaningful and whether alerts are actionable. The pilot
should identify failure modes before the system is scaled.
The sixth step is lifecycle management. Every device and edge node
should have an owner, asset record, software inventory, update schedule,
monitoring policy and retirement process. Model updates should be
versioned and validated before rollout. Configuration changes should be
auditable. Organizations should define how compromised devices are
isolated, how credentials are revoked and how data are securely deleted
when equipment is retired.
Figure 2. Edge-IoT deployment decision process
Define application and service objectives
↓
Decompose workload and data flows
↓
Place tasks across device, edge, regional and cloud layers
↓
Apply security, privacy and interoperability controls
↓
Pilot under realistic load and failure conditions
↓
Measure latency, reliability, cost and energy
↓
Deploy, monitor, update and periodically reassess
Table 1 summarizes common application requirements and likely
placement strategies.
11. Discussion
The review indicates that edge computing is best understood as a
complement to cloud computing within a distributed continuum. The cloud
remains valuable for elastic resources, centralized management,
large-scale training and long-term analytics. Edge infrastructure is
valuable when proximity, local autonomy, privacy or bandwidth
constraints matter. The architectural question is not whether edge or
cloud is superior, but how to divide tasks across layers and how to
maintain consistency, security and service quality across that
division.
A second finding is that edge-IoT value depends on workload
characteristics. Applications with strict latency requirements, high
data volumes or unreliable connectivity may benefit substantially from
local processing. Other applications may gain little from edge
deployment and may incur additional complexity. A careful workload
analysis should compare communication time, processing time, queueing,
energy and operational cost. Organizations should avoid adopting edge
computing as a branding exercise or assuming that every sensor requires
a nearby server.
Third, edge distribution creates a governance challenge. Data and
computation may cross boundaries among device manufacturers, network
operators, cloud providers, enterprises and public agencies.
Responsibility for security updates, data retention and incident
response can become unclear. Service contracts and technical
architecture should identify which party is responsible for each layer
and how evidence is shared during an incident. Without clear ownership,
the distributed nature of edge computing can create gaps in
accountability.
Fourth, security and privacy must be treated as architectural
properties. Edge computing can reduce central data collection, but it
increases the number of physical and logical locations that must be
protected. Device identity, secure update, segmentation, monitoring and
lifecycle management are necessary across the entire system. Federated
learning and local processing can support data minimization, but they
should not be presented as automatic privacy guarantees.
Finally, edge-IoT systems should be evaluated in operational context.
A laboratory demonstration may show low latency under controlled
conditions, but production systems experience variable networks,
hardware faults, environmental stress, cyber threats and organizational
change. Performance measures should include tail latency, recovery time,
energy, update reliability and the ability to operate safely during
disconnection. Long-term field studies are needed to understand
maintenance burden and total cost of ownership.
12. Research Gaps and
Future Directions
One priority is adaptive and context-aware workload placement. Future
systems need to select execution locations as network conditions, device
mobility, workload demand and energy availability change. Such decisions
must be made quickly without consuming excessive resources. Research
should compare optimization, heuristic and learning-based approaches
under realistic workloads and should report not only average performance
but also worst-case behaviour and policy stability.
A second priority is distributed orchestration at scale. Large
deployments may contain thousands or millions of devices across many
administrative domains. Hierarchical control, decentralized discovery
and policy synchronization require further study. Research should
examine how systems can preserve local autonomy while maintaining
consistent security and configuration. It should also investigate how
edge services recover after partitions and how conflicting data are
reconciled.
A third priority is security for constrained and physically exposed
devices. Many IoT devices cannot support frequent updates or heavyweight
cryptographic operations. Research is needed on lightweight
authentication, secure boot, hardware roots of trust, remote attestation
and secure lifecycle management that can be deployed at realistic cost.
Supply-chain assurance and the long-term support of low-cost devices
also require attention.
A fourth priority concerns privacy-preserving edge intelligence.
Federated learning, split learning, secure aggregation and differential
privacy may reduce raw data movement, but they introduce communication,
computation and accuracy trade-offs. Research should assess privacy
under realistic adversarial assumptions and heterogeneous data
distributions. It should also examine whether privacy guarantees remain
meaningful when multiple edge services and external datasets are
combined.
A fifth priority is interoperability. The edge ecosystem contains
diverse hardware, operating systems, network technologies, data models
and management interfaces. More work is needed on portable application
packaging, semantic interoperability, standardized telemetry and
cross-provider orchestration. Interoperability should be tested in
multi-vendor environments rather than inferred from nominal standards
compliance.
A sixth priority is sustainability. Research should measure the
energy and environmental impact of computation placement across devices,
edge nodes, networks and cloud data centres. The comparison should
include embodied hardware impacts, replacement cycles, cooling and
communication energy. Efficient model architectures, quantization,
event-driven processing and workload consolidation may improve
sustainability, but the effects should be measured across the full
system lifecycle.
A seventh priority is trustworthy edge AI for safety-critical
applications. Systems used in healthcare, transport, energy and
industrial control require robust validation, explainability, fail-safe
behaviour and clear human responsibility. Research should explore how
local AI decisions can be audited, how models can be updated without
disrupting safety and how systems behave when inputs are incomplete or
corrupted. The ability to fall back to deterministic control may be as
important as model accuracy.
Finally, more field evidence is needed from low-resource and rural
environments. Many proposed edge architectures assume reliable
electricity, skilled maintenance and high-capacity connectivity.
Research in developing regions should investigate solar-powered edge
nodes, intermittent backhaul, low-cost hardware, local repair capacity
and community governance. These conditions are important for determining
whether edge computing can support inclusive digital infrastructure
rather than deepen existing technology gaps.
13. Conclusion
Edge computing provides an architectural response to the growing
demands of IoT by moving selected computation, storage and
decision-making closer to devices and users. It can reduce response
time, limit unnecessary data transmission, support local autonomy and
improve the resilience of applications that cannot depend continuously
on remote cloud services. Its greatest value is found in workloads where
proximity, timing, data volume, privacy or connectivity constraints are
significant.
The review has shown that edge computing encompasses related
paradigms including fog computing, cloudlets and mobile or multi-access
edge computing. These approaches differ in deployment assumptions but
share the principle of distributing computing resources beyond
centralized data centres. A practical edge-IoT system combines devices,
connectivity, gateways, local and regional edge infrastructure, cloud
services and applications. Workload placement should be determined by
latency, resource availability, privacy, energy, reliability and cost
rather than by a blanket preference for local or centralized
processing.
The benefits of edge computing are accompanied by substantial
challenges. Heterogeneous hardware, distributed orchestration, physical
exposure, device identity, software supply-chain risks, privacy, energy
use and lifecycle management must be addressed across the system. Edge
intelligence and federated learning offer additional opportunities, but
they do not remove the need for security, validation and governance. The
proposed deployment framework emphasizes requirements analysis, workload
decomposition, security-by-design, interoperability, realistic pilot
testing and continuous lifecycle management.
As IoT continues to expand, edge computing is likely to remain an
important component of digital infrastructure. Its success will depend
on whether organizations can build systems that are not only fast, but
also secure, maintainable, interoperable, energy-conscious and
accountable. A well-designed edge architecture does not simply move
computation outward. It places each task where it can be performed most
responsibly and reliably, while preserving coordination across the wider
computing continuum.
Table 1. Application requirements and indicative workload
placement
| Application requirement |
Likely placement |
Reason and qualification |
| Immediate safety response |
Device or local edge |
Minimizes network dependence; safety logic requires rigorous local
validation |
| Video event detection |
Camera, gateway or nearby edge |
Reduces raw video transfer; selected clips or metadata may be sent
to cloud |
| Fleet-wide predictive analytics |
Cloud or regional edge |
Benefits from cross-site aggregation and larger compute
resources |
| Remote patient monitoring alerts |
Device or local edge with secure clinical integration |
Supports timely alerting; clinical workflows and data governance
remain essential |
| Long-term historical analysis |
Cloud data platform |
Supports large-scale storage and cross-location analysis |
| Intermittent-connectivity operation |
Local edge with later synchronization |
Maintains essential functions during outages; conflict resolution
must be defined |
| Collaborative model training |
Edge participants with coordinating service |
Can limit raw data transfer but requires safeguards for model
updates |
References
Abbas, N., Zhang, Y., Taherkordi, A., & Skeie, T. (2018). Mobile
edge computing: A survey. IEEE Internet of Things Journal, 5(1),
450–465. https://doi.org/10.1109/JIOT.2017.2750180
Atzori, L., Iera, A., & Morabito, G. (2010). The Internet of
Things: A survey. Computer Networks, 54(15), 2787–2805.
https://doi.org/10.1016/j.comnet.2010.05.010
Bonomi, F., Milito, R., Zhu, J., & Addepalli, S. (2012). Fog
computing and its role in the Internet of Things. Proceedings of the
First Edition of the MCC Workshop on Mobile Cloud Computing, 13–16.
https://doi.org/10.1145/2342509.2342513
Buyya, R., & Srirama, S. N. (Eds.). (2019). Fog and edge
computing: Principles and paradigms. Wiley.
https://doi.org/10.1002/9781119525080
Chiang, M., & Zhang, T. (2016). Fog and IoT: An overview of
research opportunities. IEEE Internet of Things Journal, 3(6), 854–864.
https://doi.org/10.1109/JIOT.2016.2584538
ETSI. (2019). Multi-access edge computing (MEC); Framework and
reference architecture (ETSI GS MEC 003 V2.1.1). European
Telecommunications Standards Institute.
https://www.etsi.org/deliver/etsi_gs/MEC/001_099/003/02.01.01_60/gs_MEC003v020101p.pdf
Gupta, H., Vahid Dastjerdi, A., Ghosh, S. K., & Buyya, R. (2017).
iFogSim: A toolkit for modeling and simulation of resource management
techniques in the Internet of Things, edge and fog computing
environments. Software: Practice and Experience, 47(9), 1275–1296.
https://doi.org/10.1002/spe.2509
Shi, W., Cao, J., Zhang, Q., Li, Y., & Xu, L. (2016). Edge
computing: Vision and challenges. IEEE Internet of Things Journal, 3(5),
637–646. https://doi.org/10.1109/JIOT.2016.2579198
Satyanarayanan, M. (2017). The emergence of edge computing. Computer,
50(1), 30–39. https://doi.org/10.1109/MC.2017.9
Satyanarayanan, M., Bahl, P., Caceres, R., & Davies, N. (2009).
The case for VM-based cloudlets in mobile computing. IEEE Pervasive
Computing, 8(4), 14–23. https://doi.org/10.1109/MPRV.2009.82
Mao, Y., You, C., Zhang, J., Huang, K., & Letaief, K. B. (2017).
A survey on mobile edge computing: The communication perspective. IEEE
Communications Surveys & Tutorials, 19(4), 2322–2358.
https://doi.org/10.1109/COMST.2017.2745201
McMahan, B., Moore, E., Ramage, D., Hampson, S., & y Arcas, B. A.
(2017). Communication-efficient learning of deep networks from
decentralized data. Proceedings of the 20th International Conference on
Artificial Intelligence and Statistics, 54, 1273–1282.
https://proceedings.mlr.press/v54/mcmahan17a.html
Kairouz, P., McMahan, H. B., Avent, B., Bellet, A., Bennis, M.,
Bhagoji, A. N., Bonawitz, K., Charles, Z., Cormode, G., Cummings, R.,
D’Oliveira, R. G. L., Eichner, H., El Rouayheb, S., Evans, D., Gardner,
J., Garrett, Z., Gascón, A., Ghazi, B., Gibbons, P. B., ... Zhao, S.
(2021). Advances and open problems in federated learning. Foundations
and Trends in Machine Learning, 14(1–2), 1–210.
https://doi.org/10.1561/2200000083
Shi, W., & Dustdar, S. (2016). The promise of edge computing.
Computer, 49(5), 78–81. https://doi.org/10.1109/MC.2016.145
Stojmenovic, I., & Wen, S. (2014). The fog computing paradigm:
Scenarios and security issues. Proceedings of the 2014 Federated
Conference on Computer Science and Information Systems, 1, 1–8.
https://doi.org/10.15439/2014F503
Yi, S., Li, C., & Li, Q. (2015). A survey of fog computing:
Concepts, applications and issues. Proceedings of the 2015 Workshop on
Mobile Big Data, 37–42. https://doi.org/10.1145/2757084.2757097
Zhang, W., & Zhang, Y. (2017). Multi-tenancy and resource
management in edge computing: A survey. IEEE Access, 5, 17241–17253.
https://doi.org/10.1109/ACCESS.2017.2743258
Zhou, Z., Chen, X., Li, E., Zeng, L., Luo, K., & Zhang, J.
(2019). Edge intelligence: Paving the last mile of artificial
intelligence with edge computing. Proceedings of the IEEE, 107(8),
1738–1762. https://doi.org/10.1109/JPROC.2019.2918951
Roman, R., Lopez, J., & Mambo, M. (2018). Mobile edge computing,
fog et al.: A survey and analysis of security threats and challenges.
Future Generation Computer Systems, 78, 680–698.
https://doi.org/10.1016/j.future.2016.11.009
Botta, A., de Donato, W., Persico, V., & Pescapé, A. (2016).
Integration of cloud computing and Internet of Things: A survey. Future
Generation Computer Systems, 56, 684–700.
https://doi.org/10.1016/j.future.2015.09.021
Al-Fuqaha, A., Guizani, M., Mohammadi, M., Aledhari, M., &
Ayyash, M. (2015). Internet of Things: A survey on enabling
technologies, protocols, and applications. IEEE Communications Surveys
& Tutorials, 17(4), 2347–2376.
https://doi.org/10.1109/COMST.2015.2444095
Varghese, B., & Buyya, R. (2018). Next generation cloud
computing: New trends and research directions. Future Generation
Computer Systems, 79, 849–861.
https://doi.org/10.1016/j.future.2017.09.020