future internet

Article
TinyML Algorithms for Big Data Management in Large-Scale
IoT Systems
Aristeidis Karras 1, * , Anastasios Giannaros 1 , Christos Karras 1, * , Leonidas Theodorakopoulos 2 ,
Constantinos S. Mammassis 3 , George A. Krimpas 1 and Spyros Sioutas 1

1 Computer Engineering and Informatics Department, University of Patras, 26504 Patras, Greece;
[email protected] (A.G.); [email protected] (G.A.K.); [email protected] (S.S.)
2 Department of Management Science and Technology, University of Patras, 26334 Patras, Greece;
[email protected]
3 Department of Industrial Management and Technology, University of Piraeus, 18534 Piraeus, Greece;
[email protected]
* Correspondence: [email protected] (A.K.); [email protected] (C.K.)

Abstract: In the context of the Internet of Things (IoT), Tiny Machine Learning (TinyML) and Big
Data, enhanced by Edge Artificial Intelligence, are essential for effectively managing the extensive
data produced by numerous connected devices. Our study introduces a set of TinyML algorithms de-
signed and developed to improve Big Data management in large-scale IoT systems. These algorithms,
named TinyCleanEDF, EdgeClusterML, CompressEdgeML, CacheEdgeML, and TinyHybridSenseQ,
operate together to enhance data processing, storage, and quality control in IoT networks, utilizing
the capabilities of Edge AI. In particular, TinyCleanEDF applies federated learning for Edge-based
data cleaning and anomaly detection. EdgeClusterML combines reinforcement learning with self-
organizing maps for effective data clustering. CompressEdgeML uses neural networks for adaptive
data compression. CacheEdgeML employs predictive analytics for smart data caching, and TinyHy-
bridSenseQ concentrates on data quality evaluation and hybrid storage strategies. Our experimental
evaluation of the proposed techniques includes executing all the algorithms in various numbers
of Raspberry Pi devices ranging from one to ten. The experimental results are promising as we
outperform similar methods across various evaluation metrics. Ultimately, we anticipate that the
Citation: Karras, A.; Giannaros, A.; proposed algorithms offer a comprehensive and efficient approach to managing the complexities of
Karras, C.; Theodorakopoulos, L.; IoT, Big Data, and Edge AI.
Mammassis, C.S.; Krimpas, G.A.;
Sioutas, S. TinyML Algorithms for Big Keywords: TinyML; Edge AI; IoT; IoT data engineering; IoT Big Data management; IoT systems
Data Management in Large-Scale IoT
Systems. Future Internet 2024, 16, 42.
https://doi.org/10.3390/fi16020042

Academic Editor: Iwona Grobelna

1. Introduction
The rapid evolution of the Internet of Things (IoT) has significantly influenced vari-
Received: 6 December 2023
ous sectors, including supply chains, healthcare, and energy systems. This technology’s
Revised: 22 January 2024
expansive data generation presents challenges in managing and interpreting vast volumes
Accepted: 23 January 2024
of information. Tiny Machine Learning (TinyML), combining the efficiency of embedded
Published: 25 January 2024
systems with advanced machine learning techniques, stands out as a critical solution for
facilitating effective, localized data processing within the IoT framework.
The research underscores the growing impact of IoT technologies like TinyML across
Copyright: © 2024 by the authors. multiple industries. In supply chain management, IoT’s role is instrumental in enhancing
Licensee MDPI, Basel, Switzerland. inventory management, marking a crucial shift towards Industry 4.0 technologies and
This article is an open access article their potential to add value to businesses [1]. The healthcare sector benefits from IoT in
distributed under the terms and enhancing services such as remote patient monitoring and real-time data analytics, albeit
conditions of the Creative Commons facing challenges in handling the growth in medical data [2]. TinyML, in particular, is
Attribution (CC BY) license (https:// recognized for initiating a new era of IoT and autonomous system applications, attributed
creativecommons.org/licenses/by/ to its power efficiency, enhanced privacy, and reduced latency in data processing [3].
4.0/).

Future Internet 2024, 16, 42. https://doi.org/10.3390/fi16020042 https://www.mdpi.com/journal/futureinternet

Future Internet 2024, 16, 42 2 of 29

The application of machine learning algorithms in embedded systems, including

TinyML, is being extensively researched. Studies focus on analyzing performance and
developing intelligent systems like solar collectors for thermal energy storage [4,5]. More-
over, integrating multiple blockchain technologies [6,7] and decentralized authentication
systems [8] has been proposed as a viable strategy for secure and efficient IoT data manage-
ment, addressing the digital challenges in the Internet of Everything era [9].
From an industrial perspective, effectively managing TinyML at scale, particularly with
regard to the inherent hardware and software constraints of IoT devices, remains a focal area
of ongoing research. Proposals suggest frameworks that utilize Semantic Web technologies
for the integrated management of TinyML models and IoT devices [10]. Additionally,
the advancement and implementation of sophisticated intelligent-anomaly-based intrusion
detection systems within IoT networks, utilizing advanced machine learning models,
have demonstrated significant efficacy in precisely detecting and proactively mitigating
malicious cyber activities [11]. This approach is crucial for enhancing network security and
ensuring the reliability of IoT systems in various industrial applications.
The impact of TinyML extends to many areas, including wearable technology [12–16],
smart cities [17–22], smart homes [23–25], smart agriculture [26–31], climatic change, en-
vironment protection, green AI sustainable applications [32–38], and automobiles [39,40].
Overcoming TinyML’s challenges, especially in hardware, is key and can be advanced
through creating an open-source community. This initiative would support system develop-
ment, enhance learning capabilities on small-scale hardware, and help unify software across
different platforms [41]. TinyML is set to not only meet current technological demands but
also shape the future of smart, efficient data processing.
However, deploying TinyML on microcontrollers presents certain challenges. These
include the selection of appropriate programming languages, limited support for various
development boards, often overlooked preprocessing steps, the choice of suitable sensors,
and a scarcity of labeled data for training purposes [42]. Overcoming these obstacles is
essential for the development of TinyML systems that are both efficient and effective.
Tiny Machine Learning (TinyML) offers significant progress in machine learning, fo-
cusing on low-resource embedded devices [43]. Despite its extensive potential for enabling
machine learning in compact formats, TinyML encounters several challenges, including
the absence of standardized benchmarks, limited development board support, program-
ming language restrictions, preprocessing oversights, sensor selection issues, and a lack
of sufficient labeled data [42,44]. In regions like Africa, where the adoption of AI and em-
bedded systems is still in its infancy, TinyML could address issues related to connectivity,
energy, and costs [45]. Moreover, its ability to facilitate independent operation from cloud
computing opens new areas for energy-efficient, secure, and private data processing [46],
introducing a new era of novel localized applications in various fields.
In the vast ecosystem of Big Data, TinyML stands out as a significant opportunity.
Given the immense volume, velocity, and variety of data in modern systems, traditional
Big Data management methods often reach their limits. TinyML presents an innovative
solution, functioning more than simply as a tool for data processing but also as an effective
technique for data management. By enabling devices to conduct initial data processing
and improvement at the network’s edge, TinyML significantly reduces the amount of
data requiring centralized processing. The subsequent selective transmission and storage
of critical data enhance both storage and processing efficiency. Moreover, this approach
facilitates quick and immediate analytics, thereby enhancing the overall value gained from
Big Data.
The aim of this study focuses on the field of IoT data management by introducing
and comparing a suite of specialized TinyML algorithms. Each algorithm, namely Tiny-
CleanEDF (Algorithm 1), EdgeClusterML (Algorithm 2), CompressEdgeML (Algorithm 3),
CacheEdgeML (Algorithm 4), and TinyHybridSenseQ (Algorithm 5), has been designed
to address specific challenges inherent in IoT systems. TinyML algorithms are essential
in this context as they enable efficient processing and intelligent decisionmaking directly
Future Internet 2024, 16, 42 3 of 29

on IoT devices, reducing the reliance on central systems and minimizing data transmis-
sion needs. By incorporating advanced techniques such as federated learning, anomaly
detection, adaptive data compression, strategic caching, and detailed data quality assess-
ment, these algorithms jointly enhance the overall efficiency, security, and reliability of
data management within IoT networks. The comparative analysis provided in this study
underscores the distinct functionalities and advantages of each algorithm, highlighting the
necessity and versatility of TinyML in handling data in the increasingly complex landscape
of IoT systems.
The remainder of this study is organized as follows: Section 2 provides comprehensive
background on TinyML, highlighting its emergence as a pivotal tool in managing Big
Data within IoT environments and discussing its application in large-scale IoT systems
and embedded devices. Section 3 outlines our methodology, covering our approach’s
advantages, framework design, hardware setup, and dataset configuration for TinyML
evaluation. Section 3.7 details the proposed algorithms we developed to utilize TinyML in
IoT contexts. The experimental results of the proposed algorithms are thoroughly presented
in Section 4, demonstrating the practical implications and effectiveness of our approach.
Finally, the study concludes in Section 5, summarizing the key findings and discussing
future research directions, emphasizing TinyML’s impact on IoT data management.

2. Background and Related Work

IoT systems frequently generate vast quantities of data, posing substantial manage-
ment and analysis challenges. Researchers have introduced several frameworks and archi-
tectures to address these challenges in IoT Big Data management and knowledge extraction.
One such proposal is the Cognitive-Oriented IoT Big Data Framework (COIB Framework),
as outlined in Mishra’s works [47,48]. This framework encompasses an implementation
architecture, layers for IoT Big Data, and a structure for organizing data. An alternative
method involves employing a Big-Data-enhanced system, adhering to a data lake archi-
tecture [49]. Key features of this system include a multi-threaded parallel approach for
data ingestion, strategies for storing both raw and processed IoT data, a distributed cache
layer, and a unified SQL-based interface for exploring IoT data. Furthermore, blockchain
technologies have been investigated for their potential to maintain continuous integrity in
IoT Big Data management [50]. This involves five integrity protocols implemented across
three stages of IoT operations.

2.1. Big Data Challenges, Internet of Things, and TinyML

The rapid expansion of the Internet of Things (IoT) marks a significant shift in the
digital landscape, marked by an extensive network of devices and sensors continuously
collecting and sending data. This fusion of environments rich in data greatly increases
the challenges related to Big Data, particularly concerning its large volume, high speed,
and varied complexity.

The Big Data Dilemma in IoT

As IoT systems evolve, they inherently generate data that challenge conventional
processing and storage infrastructures. Key challenges arising from this scenario include:
• Storage Capacity and Scalability: Traditional storage systems grapple with the ever-
growing influx of data from IoT sources, necessitating the development of more
scalable and adaptive solutions.
• Data Processing and Analysis: The heterogeneity of IoT data requires sophisticated
adaptable algorithms and infrastructures to derive meaningful insights efficiently.
• Data Transfer and Network Load: Ensuring efficient and timely data transmission
across a myriad of devices without overburdening the network infrastructure remains
a paramount concern.
• Data Integrity and Security: As data become increasingly decentralized across devices,
ensuring their authenticity and safeguarding them from potential threats are critical.
Future Internet 2024, 16, 42 4 of 29

These challenges, which highlight the wider complexities that IoT introduces to Big
Data management, are outlined in Table 1.

Table 1. Overview of Big Data challenges in IoT.

Challenge Description
The increase in interconnected devices leads to unprecedented
Data Volume data generation, surpassing the capacity of conventional storage
and processing systems.
Continuous data generation in IoT necessitates real-time analysis
Data Velocity
and response, stressing the need for prompt processing solutions.
Diverse data sources in IoT range from structured to unstructured
Data Variety
formats, posing integration and analytical complexities.
The accuracy, authenticity, and reliability of data from varied
Data Veracity
devices present significant challenges in data verification.
Consolidating data from heterogeneous sources while preserving
Data Integration
integrity and context remains complex.
Increased interconnectivity broadens the risk of cyber attacks,
Security
necessitating robust security protocols.
Balancing the protection of sensitive data within extensive
Privacy
datasets, while maintaining utility, is crucial.
Processing or transmission delays can affect the timeliness and
Latency
relevance of insights, impacting decisionmaking.

2.2. TinyML
Tiny Machine Learning (TinyML) has emerged as a growing field in machine learning,
characterized by its application in highly constrained Internet of Things (IoT) devices
such as microcontrollers (MCUs) [51]. This technology facilitates the use of deep learning
models across a multitude of IoT devices, thereby broadening the range of potential
applications and enabling ubiquitous computational intelligence. The implementation of
TinyML is challenging, primarily due to the limited memory resources of these devices and
the necessity for simultaneous algorithm and system stack design. Attracting substantial
interest in both research and development areas, numerous studies have been conducted,
focusing on the challenges, applications, and advantages of TinyML [52,53].
An essential goal of TinyML is to bring machine learning capabilities to battery-
powered intelligent devices, allowing them to locally process data without necessitating
cloud connectivity. This ability to operate independently from cloud services not only
enhances functionality but also provides a more cost-effective solution for IoT applica-
tions [3,54–56]. The academic community has thoroughly examined TinyML, with sys-
tematic reviews, surveys, and research papers delving into aspects such as its hardware
requirements, frameworks, datasets, use cases, algorithms/models, and broader applica-
tions. Notably, the development of specialized TinyML frameworks and libraries, coupled
with its integration with networking technologies, has been explored to facilitate its deploy-
ment in various sectors, including healthcare, smart agriculture, environmental monitoring,
and anomaly detection. One practical application of TinyML is in the development of soft
sensors for economical vehicular emission monitoring, showcasing its real-world applica-
bility [57]. In essence, TinyML marks a significant progression in the domain of machine
learning, enabling the execution of machine learning tasks on resource-constrained IoT
devices and microcontrollers, thus laying the groundwork for an expansive ecosystem
surrounding this technology.

2.2.1. TinyML as a Novel Facilitator in IoT Big Data Management

Within this challenging landscape, TinyML emerges as an innovative intersection
between machine learning and embedded systems. Specifically tailored for resource-
constrained devices, it presents several avenues for mitigating Big Data challenges:
Future Internet 2024, 16, 42 5 of 29

• Localized On-Device Processing: TinyML facilitates local data processing, markedly

reducing the need for continuous data transfers, thus optimizing network bandwidth
and improving system responsiveness.
• Intelligent Data Streamlining: With the ability to perform preliminary on-device
analysis, TinyML enables IoT systems to discern and selectively transmit pivotal data,
ensuring efficient utilization of storage resources.
• Adaptive Learning Mechanisms: IoT devices embedded with TinyML can continu-
ously refine their data processing algorithms, fostering adaptability to dynamic data
patterns and environmental changes.
• Reinforced Security Protocols: By integrating real-time anomaly detection at the
device level, TinyML significantly enhances the security framework, providing an
early detection system for potential data breaches or threats.
The complex challenges and problems associated with Big Data in the Internet of
Things (IoT) paradigm are diverse and complex, encompassing numerous aspects such as
data management, processing, unstructured data analytics, visualization, interoperability,
data semantics, scalability, data fusion, integration, quality, and discovery [58]. These
issues are closely related to the growing trend of “big data” within cloud computing
environments and the progressive development of IoT technologies, exerting a significant
impact on various industries, including but not limited to the power sector, smart cities,
and large-scale petrochemical plants [58–60]. Additionally, the realm of IoT architectures is
not immune to pressing security and privacy threats, making them salient challenges that
require immediate and effective addressal [61]. The efficiency and completeness of IoT Big
Data, coupled with security concerns, have emerged as critical areas of focus in the realm
of research and development [62]. Furthermore, the potential integration of blockchain
technology is being explored as a solution to ensure continued integrity in IoT Big Data
management, particularly in addressing concerns related to data correctness, resource
sharing, and the generation and verification of service-level agreements (SLA) [50].
In the IoT Big Data landscape, as delineated in Table 1, key challenges include man-
aging the vast volume of data from numerous devices, necessitating advanced storage
and processing systems. Rapid data generation requires real-time analysis and response,
highlighting the importance of data velocity. The variety of data, both structured and
unstructured, from diverse sources, complicates integration and analysis. Ensuring data
veracity, or accuracy and trustworthiness, is increasingly challenging. Integrating vari-
ous data sources while maintaining integrity is vital. Security and privacy concerns are
paramount due to heightened interconnectivity, necessitating robust protocols. Lastly,
minimizing latency to avoid obsolete insights is crucial in IoT Big Data management.
Table 2 highlights how TinyML addresses key challenges in Big Data and IoT. It
offers solutions to data overload by facilitating on-device data filtering and summarization,
significantly reducing the amount of data that needs to be transmitted to central systems.
This approach is pivotal for real-time processing needs, where localized TinyML models
enable instant data analysis, ensuring timely insights without the dependency on external
servers. Such capability is crucial in scenarios with limited or no connectivity, maintaining
device functionality.
Additionally, TinyML greatly enhances energy efficiency by optimizing models for
specific tasks, thereby conserving resources and extending battery life. This technology also
bolsters security and privacy; by processing data locally, it minimizes the risks associated
with data transmission and ensures that sensitive information remains within the user’s
control. Furthermore, TinyML contributes to the longevity of devices by reducing the strain
on their components through local processing, potentially extending their operational
lifespan. These enhancements demonstrate TinyML’s significant role in improving the
efficiency, security, and sustainability of IoT systems.
Future Internet 2024, 16, 42 6 of 29

Table 2. TinyML solutions for Big Data and IoT challenges.

Challenge TinyML Solution

Data Overload On-device preprocessing to reduce data transmission.
Real-time Processing Edge-deployed models for swift decisionmaking.
Connectivity Issues Local processing for uninterrupted operation.
Energy Efficiency Optimized tasks to conserve energy and extend device life.
Security Local data processing to minimize transmission risks.
Privacy In situ data processing to enhance privacy.
Device Longevity Reduced transmission strain to extend device life.

2.2.2. Characteristics of Large-Scale IoT Systems

The characteristics of large-scale IoT systems and the enhancements introduced by
TinyML are effectively outlined in Table 3. In these systems, a distributed topology with
devices spread across various locations results in data decentralization and increased la-
tency; TinyML tackles this by facilitating edge computation, enabling local data processing
to reduce latency and provide real-time insights. The voluminous data streams gener-
ated continuously can burden storage and transmission channels, but TinyML assists by
prioritizing, compressing, and filtering data at the device level, managing storage needs
and reducing data transmission demands.

Table 3. Characteristics of large-scale IoT systems and enhancements with TinyML.

IoT System Characteristic Implication Enhancement with TinyML

Numerous devices scattered across different locations TinyML facilitates edge computation, reducing latency
Distributed Topology
lead to data decentralization and increased latency. and ensuring real-time insights.
On-device TinyML prioritizes, compresses, and filters
Continuous data generation can overwhelm storage and
Voluminous Data Streams data, managing storage and reducing transmission
transmission channels.
needs.
Variety in device types introduces inconsistency in data TinyML standardizes data processing at source,
Diverse Device Landscape
formats and communication protocols. ensuring unified data representation across devices.
Power and Resource Devices, especially battery-operated ones, have limited TinyML models maximize computational efficiency,
Constraints computational resources. conserving device resources.
Delays in data processing can hinder time-sensitive TinyML ensures rapid on-device processing for
Real-time Processing Needs
applications. immediate responses to data changes.

The varied landscape of IoT devices, each with different data formats and commu-
nication protocols, is harmonized by TinyML, which standardizes data processing and
extraction at the source, ensuring consistent data representation across diverse device types.
Power and resource constraints, especially in battery-operated devices, pose significant chal-
lenges in IoT systems. TinyML models are designed for optimal computational efficiency,
performing tasks effectively without draining device resources. Finally, in applications
that require real-time processing, such as health monitoring or predictive maintenance,
delays in processing can be critical. TinyML enables rapid on-device processing, allowing
immediate responses to changing data patterns, thus enhancing the overall functionality
and effectiveness of large-scale IoT systems.

2.2.3. Applications of TinyML on Embedded Devices

Table 4 illustrates various applications of TinyML and machine learning in embedded
devices across different sectors. In predictive maintenance, TinyML models analyze real-
time sensor data from machinery, enabling early detection of potential failures and reducing
maintenance costs. This technology is also pivotal in health monitoring, where wearable
devices equipped with TinyML offer continuous health tracking, instantly analyzing critical
health metrics while ensuring user privacy.
In agriculture, TinyML enhances efficiency by adjusting operations based on real-time
environmental data, leading to optimal resource usage and increased yield. Voice and
Future Internet 2024, 16, 42 7 of 29

face recognition technologies in embedded devices benefit from TinyML through faster
localized processing, enhancing reliability and privacy. TinyML also plays a crucial role
in energy management within smart grids and home automation, optimizing energy use
for cost and environmental benefits. In urban development, it contributes to traffic flow
optimization by analyzing real-time vehicle and pedestrian movements, improving urban
mobility. These examples showcase TinyML’s significant impact in enhancing operational
efficiency, user experience, and sustainable practices across various industries.

Table 4. Applications of TinyML and machine learning on embedded devices.

Use Case Description

Real-time analysis of sensor data for early fault detection in machinery, reducing downtime and
Predictive Maintenance
maintenance costs.
Continuous health monitoring with wearables for vital signs and anomaly detection, enhancing
Health Monitoring
preventative healthcare.
Smart Agriculture Adaptive agriculture practices based on sensor data, optimizing resource use for better crop yield.
Voice Recognition Local processing of voice commands for quicker privacy-focused responses.
Face Recognition Low-latency facial recognition for secure access control and personalization.
Anomaly Detection Immediate detection of irregular patterns in industrial and environmental data for proactive response.
Gesture Control Touch-free device control via gesture recognition, improving user interaction and accessibility.
Energy Management Intelligent energy use in smart grids and homes based on usage patterns and predictive analytics.
Real-time traffic analysis for dynamic routing and light sequencing, enhancing urban traffic
Traffic Flow Optimization
management.
Environmental Monitoring Continuous monitoring of environmental conditions, with real-time adjustments and alerts.
Smart Retail Analysis of customer behavior for tailored retail experiences and store management.

Table 5 presents a detailed overview of TinyML applications across a range of fields.

It includes concise descriptions of each application and corresponding academic refer-
ences. The table illustrates the versatility of TinyML, from implementing CNN models
on microcontrollers for material damage analysis to its use in environmental monitoring.
Each example not only provides a clear application scenario but also cites relevant studies,
showcasing TinyML’s extensive impact in practical situations. This presentation highlights
TinyML’s role in enhancing the capabilities of embedded devices in various industries.

Table 5. Various applications of TinyML in embedded devices.

Application Description Ref.

Concrete Materials Damage Lightweight CNN on MCU for damage recognition in concrete materials, showing TinyML’s
[63]
Classification potential in structural health.
TinyML for predictive maintenance in hydraulic systems, improving service quality, performance,
Predictive Maintenance [64]
and sustainability.
TinyML for efficient keyword detection in voice-enabled devices, reducing processing costs and
Keyword Spotting [65]
enhancing privacy.
ML hardware accelerators for real-time analysis of time-series data in IoT, optimizing neural
Time-Series Analysis [66]
networks for on-device processing.
Asset Activity Monitoring TinyML for continuous monitoring of tool usage, identifying usage patterns and potential misuses. [67]
TinyML for monitoring environmental factors like air quality, contributing to smart systems for
Environmental Monitoring [68]
sustainability.

2.3. TinyML Algorithms

Table 6 provides a structured overview of various TinyML algorithms and their
specific applications in different domains. It categorizes these algorithms into areas such as
predictive maintenance, data compression, tool usage monitoring, and more, illustrating
the range of TinyML’s applicability. Each entry in the table is linked to a corresponding
reference, offering a direct connection to the source material. This format effectively
showcases the diversity of TinyML’s real-world applications, highlighting its potential to
transform various sectors through intelligent on-device data processing and analysis.
Future Internet 2024, 16, 42 8 of 29

Table 6. Classification of TinyML algorithms by application areas.

Application Area TinyML Algorithm Reference

Predictive Maintenance RUL prediction using LSTMs and CNNs [69]

Data Compression Tiny Anomaly Compressor (TAC) [70]
Tool Usage Monitoring TinyML for handheld power tool monitoring [67]
On-Device Training Neural network training for dense networks [71]
IoT Compression Evolving TinyML compression algorithm [72]
Safety-Critical Applications Software-implemented hardware fault tolerance [73]
Model Optimization Unified DNAS for compressible models [74]

Table 7 presents an organized summary of cutting-edge research in the field of Tiny

Machine Learning (TinyML). This table methodically categorizes various studies into
distinct focus areas, covering a broad spectrum from optimizing deep neural networks
on microcontrollers to applying federated meta-learning techniques in environments with
limited resources. This structured presentation not only underscores the multifaceted nature
of TinyML research but also highlights its significant role in advancing the functionalities
of embedded devices for a wide array of applications.

Table 7. Classification and focus of recent TinyML algorithm studies.

Study Focus Key Contributions Reference

Deep Neural Network Reduced Precision Optimization for DNN on-device learning in
[75]
Optimization MCUs.
Unsupervised Online Adaptive TinyML algorithm for driver behavior analysis in
[76]
Learning automotive IoT.
TinyML algorithm for anomaly detection in Industry 4.0 using
Anomaly Detection [77]
extreme values theory.
Low Precision Empirical study on quantization techniques for TinyML
[78]
Quantization efficiency.
Development of RAMAN, a re-configurable and sparse tinyML
Sparse tinyML Accelerator [79]
accelerator for edge inference.
TinyReptile: federated meta-learning algorithm for TinyML on
Federated Meta-Learning [80]
MCUs.

2.4. Data Management Techniques Utilizing TinyML in IoT Systems

In the field of Tiny Machine Learning (TinyML), data management techniques are
essential for handling machine learning models on devices with limited resources, such
as those in IoT networks. One method involves augmenting thing descriptions (TD) with
semantic modeling to provide comprehensive information about applications on devices,
facilitating the efficient management of both TinyML models and IoT devices on a large
scale [81]. Additionally, employing TinyML for training devices can lead to the creation
of a decentralized and adaptive software ecosystem, enhancing both performance and
efficiency. This approach has been effectively implemented in the development of a smart
edge computing robot (SECR) [82]. Such methodologies are increasingly important in
sectors like supply chain management, where they play a crucial role in predicting product
quality parameters and extending the shelf life of perishable goods, including fresh fruits in
modified atmospheres [83]. Moreover, the growing complexity in communication systems,
spurred by diverse emerging technologies, underscores the need for AI and ML techniques
in the analysis, design, and operation of advanced communication networks [84].
In the context of IoT systems, data management techniques incorporating TinyML
focus on effective data handling, ensuring privacy and security, and leveraging machine
learning for insightful data analysis. One strategy employs distributed key management
for securing IoT wireless sensor networks, utilizing the principles of elliptic curve cryptog-
raphy [85]. Another method involves applying data-driven machine learning frameworks
to enhance the accuracy of vessel trajectory records in maritime IoT systems [86]. Power
Future Internet 2024, 16, 42 9 of 29

management also plays a crucial role in IoT systems, particularly those reliant on compact
devices and smart networks. This often includes the adoption of low-power communica-
tion protocols and the integration of autonomous power systems, which are frequently
powered by renewable energy sources [87]. Furthermore, AI-based analytics, processed in
the cloud, are increasingly being utilized for healthcare-related data management, such as
systems designed for managing diabetic patient data [88].
Table 8 illustrates how TinyML is revolutionizing data management techniques in IoT
systems, bringing efficiency and accuracy to various processes. Techniques like predictive
imputation and adaptive data quantization exemplify this transformation. Predictive
imputation, using TinyML, maintains data integrity by filling in missing values based
on historical and neighboring data, thereby ensuring dataset completeness. Adaptive
data quantization, on the other hand, optimizes data storage and transmission. TinyML’s
role here is to analyze current data trends and dynamically adjust quantization levels for
optimal data representation.

Table 8. Advanced data management techniques utilizing TinyML in IoT systems.

Technique Objective Role of TinyML

Maintain data integrity by compensating Uses historical and neighboring data to predict and fill missing
Predictive Imputation
for missing or lost data. values, ensuring dataset completeness.
Analyzes current trends and adjusts quantization levels
Adaptive Data Quantization Optimize data storage and transmission.
dynamically for optimal representation.
Integrate data from multiple sensors for Processes and merges diverse sensor data in real time, enhancing
Sensor Data Fusion
a holistic view. accuracy and context.
Identify unusual data patterns or device Continuously monitors data streams, recognizing and flagging
Anomaly Detection
malfunctions. anomalies for prompt action.
Provide instant access to frequently used Uses predictive analytics to anticipate future data needs, caching
Intelligent Data Caching
or critical data. relevant data.
Group similar data at the edge for Performs lightweight clustering on-device for efficient aggregation
Edge-Based Clustering
efficient analytics. and transmission.
Enhance data to improve machine Augments sensor data in real time, enriching them for better
Real-time Data Augmentation
learning performance. analytics.
Local Data Lifespan Manage the relevance and storage Predicts data utility, retaining or discarding them for effective local
Management duration of data. storage management.
Discard or prioritize data based on the Filters data relevant to the situation, enhancing decisionmaking
Contextual Data Filtering
current context. processes.
Annotate raw data for subsequent Automatically labels data based on learned patterns, aiding in
On-device Data Labeling
processing. categorization and retrieval.

Sensor data fusion, another critical technique, is enhanced by TinyML’s ability to

process and merge data from various sensors in real time, thus providing a more compre-
hensive view and enhancing the accuracy of insights. Anomaly detection is particularly
vital in IoT systems, and TinyML enhances this by continuously monitoring data streams
to quickly identify and act upon unusual patterns or malfunctions. Intelligent data caching,
enabled by TinyML, predicts future data needs, ensuring that frequently used or critical
data are cached for instant access.
Further, TinyML facilitates Edge-based clustering, grouping similar data at the net-
work’s edge to simplify analytics and processing. This on-device clustering leads to more
efficient data aggregation and transmission. Real-time data augmentation and local data
lifespan management are also key areas where TinyML makes a significant impact. TinyML
augments sensor data in real time to enhance machine learning performance while also
predicting the utility of data for effective local storage management.
Contextual data filtering and on-device data labeling are other areas where TinyML
shows its prowess. By using environment-aware models, TinyML filters data relevant
to the current context, thereby enhancing decisionmaking processes. Additionally, it can
automatically label data based on learned patterns, facilitating efficient data categorization
and retrieval. These advanced data management techniques, powered by TinyML, are
pivotal in harnessing the full potential of IoT systems, ensuring that they are not only more
efficient and accurate but also more responsive to real-time demands.
Future Internet 2024, 16, 42 10 of 29

3. Methodology
In this study, we adopt a structured methodology to investigate the application of
TinyML in handling Big Data challenges within extensive IoT systems. Initially, our
approach involves integrating IoT devices with Raspberry Pi units, which are crucial
for managing the complexities of Big Data characterized by high volume, rapid velocity,
and diverse variety while ensuring accuracy and value extraction.
Subsequently, we concentrate on the technical deployment of TinyML on Raspberry
Pis, focusing on essential tasks such as data cleaning, anomaly detection, and feature
extraction. The effectiveness of these processes is comprehensively evaluated through a
series of tests, ensuring that our approach aligns with the desired outcomes. Moreover, we
introduce a feedback mechanism linked to the central Big Data system, enabling continuous
updates and enhancements to the TinyML models on Raspberry Pis. This methodology is
designed to create an efficient and adaptable system capable of addressing the dynamic
needs of Big Data management in large-scale IoT applications and systems.
Our approach involves deploying these algorithms on Raspberry Pi units, utilizing
their strengths and capabilities in federated learning, anomaly detection, data compression,
caching strategies, and data quality assessment. We systematically evaluate each algo-
rithm’s performance in real-time IoT scenarios, focusing on their efficiency in processing
and managing data. This includes assessing the scalability, responsiveness, and accuracy of
each algorithm in handling the unique data streams generated by IoT devices. By incorpo-
rating these algorithms into our methodology, we aim to provide a comprehensive solution
for Big Data challenges in IoT systems, ensuring robust and efficient data management.

3.1. Advantages of TinyML

• Reduced Latency: Data processing on Raspberry Pi eliminates the lag associated
with transmitting data to a centralized server and then fetching results. This ensures
real-time or near-real-time responses.
• Decreased Bandwidth Consumption: Only crucial or processed data may be sent to
the central server, reducing network load.
• Enhanced Privacy and Security: On-device processing ensures data privacy. Addi-
tionally, Raspberry Pis can be equipped with encryption tools to secure data before
any transmission.
• Energy Efficiency: Although Raspberry Pis consume more energy than simple sensors,
they are far more efficient than transmitting vast amounts of data to a distant server.
• Operational Resilience: Raspberry Pis equipped with TinyML can continue operations
even when there is no network connectivity.
• Scalability and Flexibility: Raspberry Pis can be equipped with a variety of tools and
software, allowing custom solutions for different data types and processing needs.

3.2. Big Data Challenges and Problems Addressed

• Volume: Local processing reduces data volume heading to centralized systems.
• Velocity: Raspberry Pis can handle high-frequency data, making real-time require-
ments attainable.
• Variety: Given their flexibility, Raspberry Pis can be customized to manage a multitude
of data formats and types.
• Veracity: They can ensure data quality, filtering anomalies or errors before transmission.
• Value: On-device processing extracts meaningful insights, ensuring only the most
relevant data are transmitted to central systems.

3.3. Framework Architecture

The proposed architecture outlines a systematic approach for managing Big Data in IoT
environments. At the base is the IoT layer, composed of various devices such as sensors and
wearables, which generate vast amounts of data. These data are directed towards Raspberry
Pi devices, equipped with TinyML capabilities. Within the Raspberry Pi layer, three primary
Future Internet 2024, 16, 42 11 of 29

tasks are undertaken: data cleaning to remove inconsistencies, anomaly detection to identify
unusual patterns, and feature extraction to select relevant data attributes. Once processed,
the refined data are transmitted to the centralized Big Data system via the communication
layer. Notably, the volume of data being transmitted is reduced due to the preliminary
processing at the Raspberry Pi level. At the top layer, the centralized system performs
further storage, analytics, and processing tasks. A feedback mechanism is incorporated,
allowing the centralized system to send updates to the Raspberry Pis, ensuring continuous
optimization. Overall, this architecture presents a structured methodology for efficient data
processing and management in large-scale IoT settings. The illustration of this architecture
is represented in Figure 1.

Model updates or
Centralized Big configuration changes
Data System

Data transmission

Communication
Processing at the Raspberry Pi level
Layer
Anomaly Feature
Data Cleaning
Detection Extraction
Raspberry Pi and
TinyML Layer

IoT Layer

Figure 1. Proposed system architecture.

3.4. Hardware Configuration

Our study’s hardware infrastructure comprises a selection of Raspberry Pi devices
and a variety of sensors, each chosen for its specific role within our IoT framework. The fol-
lowing outlines the key components of our hardware setup:
• Raspberry Pi Devices:
– 10 × Raspberry Pi 4 Model B : These are the workhorses of our setup, deployed for
edge computing and intensive data processing tasks.
– 5 × Raspberry Pi Zero W: These smaller units are used for less demanding tasks,
primarily for collecting sensor data.
• Sensor Array:
– 15 × DHT22 Temperature and Humidity Sensors: Key for monitoring environmental
conditions, providing accurate temperature and humidity readings.
– 10 × MPU6050 Gyroscope and Accelerometer Sensors: Employed to track motion
and orientation, crucial for applications requiring movement analysis.
– 8 × LDR Light Sensors: These sensors are tasked with detecting changes in light
intensity, useful in both indoor and outdoor settings.
– 7 × HC-SR04 Ultrasonic Distance Sensors: Utilized primarily for distance measure-
ment and object detection, they play a pivotal role in spatial analysis.
– 5 × Soil Moisture Sensors: Specifically selected for agricultural applications, these
sensors provide valuable data for smart farming solutions.
Future Internet 2024, 16, 42 12 of 29

This hardware ensemble, consisting of Raspberry Pi devices and a diverse set of sen-
sors, constitutes the core of our IoT network. It is adeptly designed to handle a wide spec-
trum of data collection and processing operations. The Raspberry Pi 4 models, with their
advanced capabilities, are integral for more demanding computational tasks. In contrast,
the Raspberry Pi Zero W units offer a compact energy-efficient solution for simpler ac-
tivities. The assortment of sensors capture a broad range of environmental and physical
parameters, which are vital for the thorough deployment and effectiveness of the TinyML
algorithms central to our research.

3.5. Computational Framework for IoT Model Training and Evaluation

In this study, the computational framework consists of two key elements: a centralized
High-Performance Computing (HPC) cluster and a network of 10 Raspberry Pi 4 units.
Each Raspberry Pi is configured with 4 GB of RAM and a 1.5 GHz quad-core processor.
Importantly, these Raspberry Pi units are connected to a distinct network, separate from the
HPC cluster, to emulate a realistic communication scenario as we have proposed in [89,90].
A primary function of the HPC server is to aggregate and analyze the individual
models developed on each Raspberry Pi, culminating in the evaluation of a comprehensive
global model, denoted as MG .
On each Raspberry Pi, a Flask server manages crucial tasks such as the exchange
of models, execution of local training, and monitoring of memory usage. This aspect is
particularly vital in contexts with limited hardware resources. Additionally, the server
enables the handling of requests and facilitates essential updates over the air.
The setup with Raspberry Pi units is designed to support experiments involving
more than 10 clients, indicated as K > 10, with a constraint that no more than 10 clients
(S ≤ 10) are sampled in each communication round. This capability is achieved by storing
all K client datasets on each Raspberry Pi. During each communication round l, a subset
of these datasets, corresponding to the S sampled clients, is allocated to a Raspberry Pi
for processing.
The Raspberry Pis are connected to a switch, which links them to the router via
a cable. In scenarios where the switch is not in use, the Raspberry Pis switch to a Wi-Fi
connection, allowing for the evaluation of communication overhead in two distinct network
conditions: the faster Ethernet and the slower Wi-Fi. The network operates at a bandwidth
of 100/100 Mbit/s, with the Ethernet utilizing full capacity and the Wi-Fi about 10% of it
(10/10 Mbit/s). Each Raspberry Pi represents a client corresponding to a partition of the
dataset, labeled as k1 , . . . , k S , with Raspberry Pi 1 processing data as client k i and Raspberry
Pi 10 as k j .

3.6. Dataset Configuration for TinyML Evaluation

In this study, we have thoroughly assembled a dataset to effectively evaluate our
TinyML algorithms within a comprehensive IoT framework. This dataset is characterized
by its diversity and volume, mirroring the complexities encountered in large-scale IoT
systems. The following are the key aspects of our dataset:
• Sensor Array Composition:
– Environmental Data: Sourced from DHT22 sensors, providing continuous insights
into temperature and humidity.
– Motion and Orientation Data: Collected via MPU6050 sensors, capturing detailed
information on movement and angular positions.
– Light Intensity Measurements: Obtained from LDR sensors, these readings reflect
variations in ambient lighting conditions.
– Distance and Proximity Data: Acquired from HC-SR04 ultrasonic sensors, essential
for spatial analysis and object detection.
– Soil Moisture Levels: Recorded by specialized sensors, pivotal for applications in
smart agriculture.
Future Internet 2024, 16, 42 13 of 29

• Data Volume:
– The dataset encompasses over 1 terabyte of collected raw sensor data, providing
a substantial foundation for algorithmic testing and optimization.
• Data Collection Frequency:
– Sensor readings are captured at varying intervals, ranging from high-frequency
real-time data streams to periodic updates. This variability simulates different
real-world operational scenarios, ensuring robust algorithm testing.
• Data Preparation:
– Prior to analysis, the data were subjected to essential preprocessing steps, in-
cluding cleaning and normalization, to ensure consistency and reliability for
subsequent TinyML processing.
This dataset, with its rich variety and significant volume, plays a crucial role in the
assessment of our TinyML algorithms. It not only provides a realistic environment for
testing but also ensures that the algorithms are evaluated across a range of conditions
reflective of real-world IoT systems. The frequency of data collection, in particular, allows
us to examine the algorithms’ performance under various data flow scenarios, which is
critical for their application in diverse IoT settings.

3.7. Proposed Algorithms

TinyCleanEDF, as presented in Algorithm 1, is an advanced solution tailored for data
cleaning and anomaly detection in IoT systems using federated learning. This algorithm
partitions the data stream into subsets, each processed on separate edge devices. Each
device operates a federated learning model, trained locally with its data subset. These
models synchronize with a central server, which aggregates their parameters to refine the
global model. This distributed framework enables efficient anomaly detection and data
cleaning directly at the edge, with anomalies being pinpointed when data deviate from
established patterns.
Moreover, TinyCleanEDF employs an autoencoder at each node for feature extraction.
These autoencoders are trained to reconstruct inputs from their compressed representations,
effectively distilling significant data characteristics from complex datasets. This process
is crucial for simplifying data, making them more manageable and highlighting vital
information. The algorithm is inherently dynamic, consistently updating both federated
models and autoencoders to integrate new data observations. Such continuous adaptation
ensures the system’s ongoing relevance and efficacy. Through its integration of local
data processing, anomaly detection, and feature extraction, TinyCleanEDF stands out as a
robust and comprehensive solution for upholding data integrity and quality in intricate
IoT environments.
EdgeClusterML, outlined in Algorithm 2, is an innovative algorithm designed for
dynamic and self-optimizing clustering in IoT networks. This algorithm combines re-
inforcement learning (RL) with a self-organizing map (SOM) to adaptively cluster data
at the edge. In the initial step, EdgeClusterML initializes a dynamic clustering model
using RL, where the quality of actions in different states is evaluated using a Q-function,
Q(s, a) and optimized through a defined reward function, R(s, a). The learning rate α and
discount factor γ are set to guide the learning process. Subsequently, the algorithm deploys
an SOM for efficient data clustering. The SOM is initialized with random weights and
fine-tuned using a neighborhood function and learning rate. As data flow through the
system, EdgeClusterML dynamically clusters them using the SOM, constantly updating
the model based on the data’s characteristics. RL is then employed to optimize clustering
decisions, adjusting parameters through an ϵ-greedy policy and updating the Q-function
based on observed rewards. This process leads to intelligent and responsive clustering,
tailoring the data organization to the changing patterns and needs of the IoT environment.
Finally, the clustered data are stored, ensuring organized and efficient data management at
the edge.
Future Internet 2024, 16, 42 14 of 29

Algorithm 1 TinyCleanEDF: Federated Learning for Data Cleaning and Anomaly Detection
with Autoencoder-based Feature Extraction
1: procedure T INY C LEAN EDF (dataStream)
2: Step 1: Initialize Federated Learning Models
3: Partition dataStream into subsets { D1 , D2 , . . . , Dn } for distributed processing
4: Deploy federated learning models { M1 , M2 , . . . , Mn } on edge devices
5: Train each model Mi with its subset Di
6: Models periodically execute Mi → Sync( Mi ) with central server
7: Central server performs Aggregate({ M1 , M2 , . . . , Mn })
8: Step 2: Federated Model for Data Cleaning and Anomaly Detection
9: Apply f anomaly ( x; Mi ) to detect and clean anomalies locally
10: Anomalies identified as x ∈ / ExpectedPattern( Mi )
11: Cleaned data {C1 , C2 , . . . , Cn } sent to central server
12: Step 3: Deploy Autoencoder for Feature Extraction
13: Implement autoencoder AEi at each node i
14: Train AEi to reconstruct input x from compressed representation z
15: Feature extraction: f features ( x; AEi ) = HiddenLayer( AEi ( x ))
16: Step 4: Continuous Adaptation and Feature Extraction
17: for each dataPoint in dataStream do
18: cleanDataPoint ← f clean (dataPoint; Mi )
19: anomaly ← f anomaly (cleanDataPoint; Mi )
20: f eatures ← f features (cleanDataPoint; AEi )
21: Update Mi and AEi with dataPoint for continuous learning
22: Store (cleanDataPoint, anomaly, f eatures)
23: end for
24: end procedure

Algorithm 2 EdgeClusterML: Dynamic and Self-Optimizing Clustering at the Edge

1: procedure E DGE C LUSTER ML (dataStream)
2: Step 1: Initialize Dynamic Clustering Model with RL
3: Let Q(s, a) represent the quality of action a in state s
4: Initialize Q(s, a) for all state-action pairs
5: Define reward function R(s, a) for evaluating clustering actions
6: Set learning rate α and discount factor γ
7: Step 2: Deploy Self-Organizing Map (SOM) for Clustering
8: Initialize SOM with random weights W
9: Define neighborhood function hci (t) for neuron i at time t
10: Set SOM learning rate η
11: for each dataPoint in dataStream do
12: Step 3: Dynamic Clustering with SOM
13: Find Best Matching Unit (BMU) for dataPoint in SOM
14: Update weights W using hci (t) and η
15: Step 4: RL-based Optimization of Clustering
16: Observe state s (current clustering configuration)
17: Choose action a (adjusting clustering parameters) using ϵ-greedy policy
18: Apply action a, observe new state s′ and reward r
19: Update Q(s, a) using the Bellman equation:
20: Q(s, a) ← Q(s, a) + α[r + γ maxa′ Q(s′ , a′ ) − Q(s, a)]
21: Step 5: Store Clustered Data
22: clusterID ← BMU index for dataPoint
23: Store dataPoint in clusterID
24: end for
25: end procedure
Future Internet 2024, 16, 42 15 of 29

CompressEdgeML, as described in Algorithm 3, is designed to introduce smart and

adaptive data compression capabilities to the edge of IoT networks. This algorithm uti-
lizes a neural network, denoted as NN, specifically customized for data compression
tasks. It is trained on sample data to efficiently identify and execute compression pat-
terns, setting an initial compression ratio CR based on the characteristics of the training
data. CompressEdgeML dynamically adapts its compression techniques to align with
the current network conditions (represented as Ncond ), ensuring an optimal balance be-
tween the quality of compression and operational efficiency. As data stream through the
network, CompressEdgeML uses its neural network model to compress each datapoint.
The algorithm continually updates the compression ratio in response to changing net-
work bandwidth and storage capacities, ensuring that the size of the compressed data
(|compressedData|) is always suitable for the network and storage constraints. The final step
involves securely storing the compressed data in a specific storage system. This approach
significantly enhances data management efficiency in IoT environments by reducing the
size of data for transmission and storage while maintaining the integrity and usability of
the information.

Algorithm 3 CompressEdgeML: Adaptive Data Compression

1: procedure C OMPRESS E DGE ML (dataStream)
2: Step 1: Initialize Neural Network-based Selective Compression Model
3: Define neural network NNcomp for data compression
4: Train NNcomp on dataset Dtrain for compression patterns
5: Initialize compression ratio CR ← CRinit ( Dtrain )
6: Step 2: Deploy Adaptive Compression Techniques
7: Monitor network conditions Ncond
8: Define adaptive function FCR ( Ncond , CR) for compression ratio
9: Balance between compression quality and network efficiency
10: for each dataPoint ∈ dataStream do
11: Step 3: Data Compression Using Neural Network
12: compressedData ← NNcomp (dataPoint, CR)
13: Measure size |compressedData|
14: Step 4: Adapt Compression Level
15: Update Ncond based on network bandwidth and storage
16: Adjust CR using FCR ( Ncond , CR)
17: Ensure |compressedData| fits network and storage constraints
18: Step 5: Store Compressed Data
19: Store compressedData in targeted storage system
20: end for
21: end procedure

CacheEdgeML, outlined in Algorithm 4, represents an advanced approach to data

caching in IoT networks. This algorithm efficiently handles data requests by utilizing a
predictive analytics model that bases its forecasts on historical access patterns. By training
the model, denoted as P, to predict the likelihood of future data requests, CacheEdgeML
can effectively identify and prioritize high-importance data. It employs a multi-tier caching
strategy, where data are organized into various tiers (T1 , T2 , . . . , Tn ) according to priority
and frequency of access. This setup not only simplifies the management of data but also
ensures the efficient utilization of storage resources. Moreover, CacheEdgeML works
together with cloud services to manage data storage effectively. It establishes a synchro-
nization mechanism with cloud storage and devises strategies for offloading data when the
cache reaches capacity, maintaining consistency between edge and cloud storage. The al-
gorithm dynamically manages cache tiers, continuously re-evaluating data priority and
relocating frequently accessed data to higher tiers. This adaptability ensures that the
most relevant data are readily available, thus enhancing the system’s overall efficiency
and responsiveness.
Future Internet 2024, 16, 42 16 of 29

Algorithm 4 CacheEdgeML: Predictive and Tiered Data Caching Strategy

1: procedure C ACHE E DGE ML (dataRequests)
2: Step 1: Initialize Predictive Analytics Model for Anticipatory Caching
3: Define predictive model Pcache based on historical patterns
4: Train Pcache on dataset Dhist to estimate request probabilities
5: Set priority threshold θ for high-importance data classification
6: Step 2: Implement Multi-Tier Caching Based on Priority and Frequency
7: Define cache tiers T = { T1 , T2 , ..., Tn }
8: Assign frequency thresholds F = { f 1 , f 2 , ..., f n } for each Ti
9: Store data in tier Ti based on access frequency f i
10: Step 3: Collaborate with Cloud Services for Cache Management
11: Establish synchronization Ssync with cloud storage
12: Define offloading strategy Ωoffload when cache full
13: Maintain consistency Cconsist between edge and cloud
14: for each request ∈ dataRequests do
15: if request ∈ Cache then
16: Step 4: Serve from Cache
17: Serve data from cache
18: else
19: Step 5: Predict Future High-Priority Requests
20: Calculate Pcache (request), compare with θ
21: Update cache based on Pcache (request)
22: Step 6: Dynamic Management of Cache Tiers
23: Re-evaluate priority for request, adjust T
24: Relocate data to appropriate tier based on F
25: Step 7: Serve Requested Data
26: Fetch data from external source or cloud if not in cache
27: Store new data in appropriate tier based on priority
28: end if
29: end for
30: end procedure

TinyHybridSenseQ, as detailed in Algorithm 5, is a sophisticated TinyML-based algo-

rithm specifically designed for IoT environments. It employs advanced machine learning
models deployed on edge devices, focusing on the critical tasks of analyzing, categorizing,
and efficiently storing data collected from a wide array of sensors. The core competency of
this algorithm is in its robust data quality assessment module, which thoroughly evaluates
the integrity and accuracy of sensor measurements. This evaluation is crucial for filtering
out incorrect data, ensuring that only reliable and high-quality information is processed
further. TinyHybridSenseQ also performs well in data management, implementing a
dynamic and data-aware hybrid storage strategy. It efficiently determines the optimal
storage location for each data packet—be it local storage for less critical data or immediate
transfer to a central database for high-priority information. This strategic approach not
only simplifies data handling but also significantly enhances the overall efficiency of data
transfer processes in IoT networks. Furthermore, TinyHybridSenseQ continuously evolves
through its model adaptation feature, which refines its analytical capabilities based on
incoming sensor data, thereby maintaining high accuracy and relevance in ever-changing
IoT environments.

3.8. Comparison among Proposed TinyML Algorithms

The following Table 9 provides a comparative overview of the five proposed algo-
rithms: TinyCleanEDF, EdgeClusterML, CompressEdgeML, CacheEdgeML, and TinyHy-
bridSenseQ. The table outlines key features such as federated learning, anomaly detection,
data compression, caching strategy, and data quality assessment. This comparison helps in
Future Internet 2024, 16, 42 17 of 29

understanding the unique capabilities and functionalities that each algorithm brings to IoT
Big Data management.

Algorithm 5 TinyHybridSenseQ: Data-Aware Hybrid Storage and Quality Assessment for

IoT Sensors
1: procedure T INY H YBRID S ENSE Q (sensorData)
2: Step 1: Initialize Data Quality Assessment Model
3: Deploy TinyML models { M1 , M2 , . . . , Mn } for data quality assessment
4: Train each model Mi to identify anomalies and inconsistencies
5: Step 2: Hybrid Storage Strategy Initialization
6: Define data storage strategies S = {S1 , S2 , . . . , Sn } based on type and priority
7: Initialize storage resources Rlocal and Rcloud
8: for each dataPacket ∈ sensorData do
9: Step 3: Assess Data Quality
10: Quality score Q ← f quality (dataPacket, Mi )
11: Categorize dataPacket as high-quality or low-quality based on Q
12: Step 4: Data Categorization and Prioritization
13: Categorize dataPacket into type T
14: Prioritize dataPacket for storage based on Q and T
15: Step 5: Efficient Data Transfer and Storage
16: if Q is high and dataPacket is high-priority then
17: Transfer to central database DB
18: else
19: Store in Rlocal or Rcloud based on S
20: end if
21: Step 6: Continuous Model Adaptation and Reporting
22: Update Mi with new dataPacket for continual learning
23: Generate and store reports on data quality and storage
24: end for
25: end procedure

Table 9. Comparison of proposed algorithms.

Federated Anomaly Caching Data

Algorithm Data Compression
Learning Detection Strategy Quality
TinyCleanEDF (Algorithm 1) ✓ ✓ × × ✓
EdgeClusterML (Algorithm 2) × ✓ × × ×
CompressEdgeML (Algorithm 3) × × ✓ × ×
CacheEdgeML (Algorithm 4) × × × ✓ ×
TinyHybridSenseQ (Algorithm 5) × ✓ × ✓ ✓

4. Experimental Results
4.1. Overview
In this section, we assess the performance of our five proposed algorithms—TinyCleanEDF,
EdgeClusterML, CompressEdgeML, CacheEdgeML, and TinyHybridSenseQ—across mul-
tiple key performance metrics. The evaluation is conducted by varying the number of
Raspberry Pi devices used in the deployment, ranging from one to ten. Each algorithm’s
performance is measured across elements such as accuracy, compression efficiency, data
processing time (ms), training time (ms), overall efficiency, and scalability. The metrics
utilized in this work are provided in detail in Section 4.2 below.

4.2. Metrics and Methods

For the evaluation of the proposed techniques, the following metrics are utilized.
Future Internet 2024, 16, 42 18 of 29

1. Data Processing Time (ms): To measure the data processing time in a distributed
system such as the one proposed where we have multiple Raspberry Pi devices, we
can consider the maximum time taken by any single device as well as the average
time across all devices. The equation is provided in Equation (1).

∑in=1 Ti
Data Processing Timetotal = max( T1 , T2 , . . . , Tn ) and Data Processing Timeavg = (1)
n

2. Model Training Time (ms): For the model training time, we want to measure both
the total cumulative time and the longest individual training time across all devices.
The calculation is provided in Equation (2).
n
Model Training Timetotal = ∑ Ti and Model Training Timemax = max( T1 , T2 , . . . , Tn ) (2)
i =1

3. Anomaly Detection Accuracy: For a distributed system, we want to consider not only
the overall accuracy but also the consistency of anomaly detection across different
nodes. A weighted approach is used where the accuracy of each node is weighted by
the number of instances it processes. This is provided in Equation (3).

∑in=1 (wi × Accuracyi )

Anomaly Detection Accuracy = (3)
∑in=1 wi

4. Communication Efficiency: In a large-scale distributed setup, communication effi-

ciency should account for the data transmission efficiency, the overhead of synchro-
nization among nodes, the error rate in data transmission, and the effective utilization
of available bandwidth. This comprehensive approach ensures a realistic assessment
of communication performance in a distributed system.
 
Du
CE = × (1 − ER) × BU (4)
Dt + Osync

• CE represents Communication Efficiency.

• Du is the symbol for Useful Data Transmitted.
• Dt stands for Total Data Transmitted.
• Osync is the Synchronization Overhead.
• ER denotes the Error Rate.
• BU symbolizes Bandwidth Utilization.
5. Scalability: Scalability in a distributed system can be quantified by measuring how the
system’s performance changes with the addition of more nodes, considering factors
like throughput, response time, load balancing, system capacity, and cost-effectiveness.
A higher throughput ratio, a lower response time ratio, and efficient load balancing
with increased nodes indicate better scalability.

Throughput at n nodes
Scalability =
Throughput at a single node
Response Time at n nodes
Response Time Ratio =
Response Time at a single node
∑in=1 Load on Node i
Load Balancing Efficiency = (5)
Ideal Load per Node × n
Total Processed Load
System Capacity Utilization =
Total System Capacity
Total System Cost at n nodes
Cost-Effectiveness Ratio =
Performance Improvement Factor
Future Internet 2024, 16, 42 19 of 29

Starting with the evaluation of the first Algorithm 1, the results are shown in Figure 2.

TinyCleanEDF Evaluation for Various Number of Nodes

Number of Nodes
1
102 2
5
10

Log-scaled Values

101

Anomaly Detection Accuracy Data Processing Time (ms) Model Training Time (ms) Communication Efficiency Scalability
Evaluation Metrics

Figure 2. Performance evaluation of TinyCleanEDF.

Incorporating the TinyCleanEDF algorithm into an IoT data management system has
demonstrated quantifiable improvements in several key performance metrics. As shown
in the preceding Figure, the deployment of this algorithm across an increasing number of
nodes—from 1 to 10—has yielded substantial benefits. Specifically, the anomaly detection
accuracy improved with more nodes. For instance, there was a 10% increase in the accuracy
on a single node compared to ten nodes. This improvement highlights the algorithm’s
enhanced capability to identify and respond to data anomalies as the collaborative network
of nodes expands.
Moreover, the data processing and model training times, both critical for the efficient
operation of IoT systems, show a decreasing trend as more nodes are engaged. Log-
scaled values indicate that processing time decreased fourfold when the number of nodes
increased from 1 to 10, which suggests a notable enhancement in the speed of data handling.
Communication efficiency also saw a rise, which is particularly relevant in scenarios where
network bandwidth is a limiting factor. This increase indicates a more optimal use of
available resources, allowing for smoother data transfer between nodes and the central
server. Lastly, scalability, which is also a significant metric, reflects the algorithm’s ability
to maintain performance despite the growing scale of the network. The consistent upward
trend across nodes validates that TinyCleanEDF is well-suited for environments where
expansion is anticipated, ensuring that the system not only sustains its performance but
actually improves as it scales.
These results underscore the effectiveness of TinyCleanEDF in enhancing data quality
and system robustness, making it a compelling choice for federated learning applications
in distributed networks. Moving on to Algorithm 2, the results are presented in Figure 3.
The integration of the EdgeClusterML algorithm within an edge computing framework
such as FL has yielded remarkable improvements in critical performance metrics. Notably,
the algorithm achieved an impressive accuracy rate of approximately 90% when applied to
real-world data streams. This represents a significant enhancement in the precision of data
clustering, making it well-suited for applications like anomaly detection and data-driven
decisionmaking. The observed increase in accuracy is particularly noteworthy as it directly
impacts the algorithm’s ability to effectively group datapoints.
Future Internet 2024, 16, 42 20 of 29

EdgeClusterML Evaluation for Various Device Configurations

102 Number of Devices
1
2
5
10

Values (Log-scaled for Speed)

101

Clustering Accuracy (%) Clustering Speed (ms) Resource Utilization (%) Adaptability Score (%)
Evaluation Metrics

Figure 3. Performance evaluation of EdgeClusterML.

Furthermore, our analysis reveals significant reductions in the clusterind speed. Specif-
ically, the algorithm exhibited a time reduction of approximately 10% in the clustering
speed when transitioning from one to ten nodes. These reductions are crucial in edge
computing scenarios, ensuring real-time responsiveness and rapid adaptation to changing
data patterns. The improvement in resource utilization and the adaptability score, with a
roughly 15% increase as nodes scaled, signifies more efficient resource utilization and data
transfer, particularly valuable in resource-constrained edge environments.
In conclusion, EdgeClusterML emerges as a robust solution for edge computing en-
vironments, offering concrete benefits in terms of accuracy, clustering speed, resource
utilization, and adaptability. Its reinforcement-learning-driven dynamic clustering ap-
proach positions it as a valuable asset for real-time data analysis and decisionmaking in
dynamic edge scenarios. In the next steps, we evaluate Algorithm 3 in Figure 4.

CompressEdgeML Evaluation for Various Device Configurations

102 Number of Devices
1
2
5
10
Values (Log-scaled for Speed)

101

Compression Efficiency (%) Compression Speed (ms) Data Integrity Post-Compression (%) Resource Utilization (%)
Evaluation Metrics

Figure 4. Performance evaluation of CompressEdgeML.

In our evaluation of the CompressEdgeML algorithm, significant improvements were

observed across various performance metrics in edge computing environments. The al-
gorithm, however, shows a good compression efficiency of up to 95% on single-device
configurations, highlighting its effectiveness in data size reduction. This efficiency slightly
decreases as the number of devices increases, stabilizing at 88% for configurations with
10 devices, indicating a high level of data compression consistency. Compression speed,
measured in milliseconds, displayed a marked improvement with increasing device num-
bers. For a single device, the compression time was logged at approximately 1200 ms,
Future Internet 2024, 16, 42 21 of 29

which reduced logarithmically to around 300 ms for 10 devices. This reduction showcases
the algorithm’s capability to handle larger data streams more efficiently, a critical attribute
in real-time edge computing scenarios.
Data integrity post-compression was maintained above 90% across all device configu-
rations, peaking at 98% in a single-device setup. This metric underscores the algorithm’s
reliability in preserving essential data characteristics during the compression process. Re-
source utilization also showed a positive trend, with efficiency increasing from 70% in a
single-device scenario to 85% in a 10-device configuration. This improvement indicates
the algorithm’s scalability and its efficient use of computational resources, which is vital in
resource-constrained edge environments.
In summary, CompressEdgeML demonstrates robust performance in adaptive data
compression, marked by high compression efficiency, accelerated processing speeds, reli-
able data integrity, and efficient resource utilization. Its adaptability and scalability make it
well-suited for diverse edge computing applications. The next algorithm is Algorithm 4,
which is evaluated in Figure 5.

CacheEdgeML Evaluation for Various Device Configurations

Number of Devices
1
2
102 5
10
Values (Log-scaled for Time Metrics)

101

Cache Hit Rate (%) Cache Update Speed (ms) Cloud Synchronization Latency (ms) Data Retrieval Efficiency (%)
Evaluation Metrics

Figure 5. Performance evaluation of CacheEdgeML.

In our assessment of the CacheEdgeML algorithm, tailored for predictive and tiered
data caching in edge computing settings, we observed substantial enhancements in pivotal
performance metrics. The algorithm exhibited a cache hit rate of 85% in a single-device
environment, which progressively increased to 92% with the addition of more devices.
This upward trend signifies the algorithm’s enhanced accuracy in predicting data requests,
a crucial factor in reducing redundant data retrieval operations.
The cache update speed, a critical measure in dynamic environments, improved
logarithmically from 400 ms for one device to 250 ms for ten devices. This acceleration high-
lights the algorithm’s efficiency in adapting to changing data patterns, thereby optimizing
caching strategies in real time. Cloud synchronization latency, through the HPC server, is
essential for maintaining data consistency between edge and cloud storage, and it was also
optimized. It decreased from 95 ms to 99 ms as the number of devices increased, demon-
strating the algorithm’s effectiveness in synchronizing large volumes of data swiftly across
distributed networks. Data retrieval efficiency, indicative of the algorithm’s performance in
providing timely access to cached data, showed, however, a negative trajectory, decreasing
from 200% in single-device setups to 140% in scenarios involving ten devices. This decrease
shows that the algorithm requires more time to streamline data access, particularly in
multi-device edge computing networks where data are distributed.
In summary, CacheEdgeML emerges as a robust and adaptive solution for data caching
in edge computing environments. Its strengths lie in its high cache hit rate, improved cache
update speed, and reduced cloud synchronization latency. These attributes collectively
Future Internet 2024, 16, 42 22 of 29

ensure that CacheEdgeML is well-equipped for optimizing data caching processes in

complex real-time edge computing scenarios. Lastly, we evaluate Algorithm 5 in Figure 6.

TinyHybridSenseQ Evaluation for Various Sensor Configurations

102 Number of Sensors
1
2
5
Values (Log-scaled for Time Metrics) 10

101

Data Quality Score (%) Anomaly Detection Speed (ms) Storage Efficiency (%) Data Transfer Latency (ms)
Evaluation Metrics

Figure 6. Performance evaluation of TinyHybridSenseQ.

In our evaluation of the TinyHybridSenseQ algorithm, specifically designed for IoT

sensors with a focus on data quality assessment and hybrid storage, we observed notable
improvements in several key performance areas. The algorithm achieved a high Data
Quality Score of 95% in a single-sensor setup, which marginally decreased to 90% as the
number of sensors increased to ten. This trend highlights the algorithm’s robustness in
maintaining high data quality standards even as the sensor network scales.
Anomaly Detection Speed, a crucial metric in real-time environments, showed signifi-
cant optimization. It improved logarithmically from 500 ms in a single-sensor scenario to
350 ms with ten sensors, illustrating the algorithm’s accelerated response in identifying data
anomalies. This enhancement is pivotal for timely anomaly detection in dynamic sensor
networks. Storage efficiency, essential in optimizing data storage across local and cloud
resources, also saw progressive improvements. It increased from 85% to 91% as the number
of sensors expanded. This increase indicates the algorithm’s capability to efficiently manage
storage resources, a vital aspect in IoT environments where data volume is substantial.
Data transfer latency, critical in ensuring the swift movement of data between sensors and
storage facilities, was optimized with the expansion of the sensor network. It reduced from
250 ms for a single sensor to 160 ms for ten sensors, signifying the algorithm’s effectiveness
in reducing data transfer times across a distributed network.
In conclusion, TinyHybridSenseQ stands out as an effective solution for IoT sensors,
excelling in data quality assessment and hybrid storage management. Its strengths lie
in maintaining high data quality, fast anomaly detection, efficient storage utilization,
and reduced data transfer latency. These attributes collectively position TinyHybridSenseQ
as a highly capable tool in managing complex data workflows in IoT sensor networks,
ensuring both data integrity and operational efficiency.

4.3. Comparison with Similar Works

In this section, we evaluate our proposed methods with similar works and approaches.
In particular, we evaluate the CacheEdgeML with similar approaches, namely the Proximal
Policy Optimization (PPO) caching method, the Markov Chain Monte Carlo (MCMC)
method, and the deep reinforcement learning (DRL) method, based on our previous
work [91], as well as methods like Least Frequently Used (LFU) and Least Recently Used
(LRU) based on works [92–94]. The results are provided in Figure 7.
Future Internet 2024, 16, 42 23 of 29

Cache Hit Rate Comparison: CacheEdgeML vs Similar Methods

80

60
Cache Hit Rate (%)

40

Methods
PPO
20 MCMC
LFU
LRU
DRL
CacheEdgeML
0
1

10
Number of Devices

Figure 7. Cache hit rate comparison of CacheEdgeML with similar methods.

As can be seen from the preceding figure, the cache hit rate of the proposed CacheEdgeML
method outperforms the other five methods in terms of cache hit rate, reaching above
80% across all assessments. Additionally, we assess our proposed method named Com-
pressEdgeML with a similar method presented in [72] named TAC. The results for the
compression efficiency are provided in Figure 8, while the compression speed is provided
in Figure 9.
Compression Efficiency: CompressEdgeML vs Similar Method
100 CompressEdgeML
TAC

80
Compression Efficiency (%)

60

40

20

0
1

10

Number of Devices

Figure 8. Compression efficiency of CompressEdgeML compared to similar method.

Compression Speed: CompressEdgeML vs Similar Method

3.5 CompressEdgeML
TAC
3.0

2.5
Compression Speed (Log-scaled)

2.0

1.5

1.0

0.5

0.0
1

10

Number of Devices

Figure 9. Compression speed of CompressEdgeML compared to similar method.

Future Internet 2024, 16, 42 24 of 29

As can be seen from Figure 8, the compression efficiency of our method is lower
than TAC in the beginning; however, it reaches the performance of TAC when utilized
on five devices and finally overtakes it using ten devices. Note that the TAC method in
the original method is implemented only once and its performance remains stable across
replication on our devices. As per the compression speed, for the assessment, we replicated
the TAC method in our devices and assessed the compression while increasing the size
of data in both methods. As per the compression speed, as can be seen from Figure 9,
the compression speed of CompressEdgeML is lower, meaning faster compression while
maintaining a good speed across all devices.

5. Conclusions and Future Work

In the context of IoT, where vast quantities of data are generated, the integration of
Edge AI and Big Data management plays a pivotal role in harnessing the full potential
of these technologies. Edge AI, by processing data at the source rather than relying on
distant servers, significantly reduces latency and enhances real-time data analysis. This
approach is particularly beneficial in IoT systems, where immediate decisionmaking based
on large-scale data is often required. In this context, effective data management becomes
crucial, entailing not just the storage and retrieval of data but also its processing, analysis,
and security. The intersection of Edge AI and Big Data management in IoT represents
a forward step in technology, offering novel solutions to manage and leverage the ever-
growing expanse of data in smart environments.
This study has methodically evaluated five different TinyML algorithms named Tiny-
CleanEDF, EdgeClusterML, CompressEdgeML, CacheEdgeML, and TinyHybridSenseQ—each
tailored for specific functions within IoT systems utilizing edge computing. Our findings
reveal that these algorithms substantially enhance the operational efficiency, data integrity,
and real-time processing capabilities of IoT networks, particularly when implemented
across a network of Raspberry Pi devices.
TinyCleanEDF excels in federated learning and real-time anomaly detection, thus
proving invaluable in scenarios requiring collaborative data processing and instantaneous
anomaly identification. EdgeClusterML, with its reinforcement-learning-based dynamic
clustering, demonstrates remarkable accuracy and optimal resource management, essential
for real-time data analysis and decisionmaking processes. CompressEdgeML showcases its
strength in adaptive data compression, achieving significant compression efficiency without
compromising data integrity. CacheEdgeML, through its innovative caching strategy,
ensures effective data retrieval and synchronization between edge and cloud storage, vital
for seamless data management. Lastly, TinyHybridSenseQ effectively manages data quality
and storage in IoT sensor networks, ensuring data reliability and operational efficiency.

Future Work
For future research, several key areas have been identified to further enhance the
capabilities of these algorithms:
1. Anomaly Detection: There is space for incorporating more advanced machine learning
models to enhance the accuracy and speed of anomaly detection, especially in envi-
ronments with complex or noisy data. This will allow for more precise identification
of irregularities, enhancing the overall data integrity.
2. Energy Efficiency: Optimizing the energy consumption of these algorithms is crucial,
particularly in environments where energy resources are limited. Research should
focus on developing energy-efficient methods that reduce the overall energy demand
of the system without sacrificing performance.
3. Cloud–Edge Integration: Enhancing the interaction between edge and cloud platforms
is essential for improved data synchronization and storage efficiency. This involves
developing methods for more seamless data processing and management in hybrid
cloud–edge environments.
Future Internet 2024, 16, 42 25 of 29

4. Real-Time Data Processing: Optimizing these algorithms for real-time processing of

streaming data is imperative. This would enable timely decisionmaking based on the
most current data, a critical aspect in dynamic IoT environments.
5. Security and Privacy: Strengthening the security and privacy features of these algo-
rithms is important, especially for applications handling sensitive information. This
involves implementing robust security measures to protect data from unauthorized
access and ensure user privacy.
6. Customization and Adaptability: Improving the adaptability of these algorithms
to various IoT environments is necessary. Future work should aim at developing
customizable solutions that can be tailored to meet specific requirements of differ-
ent applications.
7. Interoperability and Standardization: Promoting interoperability between diverse IoT
devices and platforms and contributing to standardization efforts is crucial. This will
facilitate smoother integration and communication across different systems and devices.
Ultimately, this study demonstrates a robust framework for future breakthroughs in
IoT data management within edge computing frameworks. The identified areas for future
exploration present promising opportunities for extending the current capabilities of these
algorithms and exploring novel possibilities in the era of IoT and edge computing.

Author Contributions: A.K., A.G., C.K., L.T., C.S.M., G.A.K. and S.S., conceived the idea, designed
and performed the experiments, analyzed the results, drafted the initial manuscript, and revised the
final manuscript. All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Data Availability Statement: Data are contained within the article.
Conflicts of Interest: The authors declare no conflicts of interest.

References
1. Mashayekhy, Y.; Babaei, A.; Yuan, X.M.; Xue, A. Impact of Internet of Things (IoT) on Inventory Management: A Literature
Survey. Logistics 2022, 6, 33. [CrossRef]
2. Vonitsanos, G.; Panagiotakopoulos, T.; Kanavos, A. Issues and challenges of using blockchain for iot data management in smart
healthcare. Biomed. J. Sci. Tech. Res. 2021, 40, 32052–32057.
3. Zaidi, S.A.R.; Hayajneh, A.M.; Hafeez, M.; Ahmed, Q.Z. Unlocking Edge Intelligence Through Tiny Machine Learning (TinyML).
IEEE Access 2022, 10, 100867–100877. [CrossRef]
4. Ersoy, M.; Şansal, U. Analyze Performance of Embedded Systems with Machine Learning Algorithms. In Proceedings of the
Trends in Data Engineering Methods for Intelligent Systems: Proceedings of the International Conference on Artificial Intelligence
and Applied Mathematics in Engineering (ICAIAME 2020), Antalya, Turkey, 18–20 April 2020; Springer: Berlin/Heidelberg,
Germany, 2021; pp. 231–236.
5. Khobragade, P.; Ghutke, P.; Kalbande, V.P.; Purohit, N. Advancement in Internet of Things (IoT) Based Solar Collector for Thermal
Energy Storage System Devices: A Review. In Proceedings of the 2022 2nd International Conference on Power Electronics & IoT
Applications in Renewable Energy and its Control (PARC), Mathura, India, 21–22 January 2022; pp. 1–5. [CrossRef]
6. Ayub Khan, A.; Laghari, A.A.; Shaikh, Z.A.; Dacko-Pikiewicz, Z.; Kot, S. Internet of Things (IoT) Security with Blockchain
Technology: A State-of-the-Art Review. IEEE Access 2022, 10, 122679–122695. [CrossRef]
7. Chauhan, C.; Ramaiya, M.K. Advanced Model for Improving IoT Security Using Blockchain Technology. In Proceedings of the
2022 4th International Conference on Smart Systems and Inventive Technology (ICSSIT), Tirunelveli, India, 20–22 January 2022;
pp. 83–89. [CrossRef]
8. Mohanta, B.K.; Jena, D.; Satapathy, U.; Patnaik, S. Survey on IoT security: Challenges and solution using machine learning,
artificial intelligence and blockchain technology. Internet Things 2020, 11, 100227. [CrossRef]
9. Jiang, Y.; Wang, C.; Wang, Y.; Gao, L. A Cross-Chain Solution to Integrating Multiple Blockchains for IoT Data Management.
Sensors 2019, 19, 2042. [CrossRef]
10. Ren, H.; Anicic, D.; Runkler, T. How to Manage Tiny Machine Learning at Scale: An Industrial Perspective. arXiv 2022,
arXiv:2202.09113.
11. Keserwani, P.K.; Govil, M.C.; Pilli, E.S.; Govil, P. A smart anomaly-based intrusion detection system for the Internet of Things
(IoT) network using GWO–PSO–RF model. J. Reliab. Intell. Environ. 2021, 7, 3–21. [CrossRef]
12. Gibbs, M.; Woodward, K.; Kanjo, E. Combining Multiple tinyML Models for Multimodal Context-Aware Stress Recognition on
Constrained Microcontrollers. IEEE Micro 2023, 1–9. [CrossRef]
Future Internet 2024, 16, 42 26 of 29

13. Chen, Z.; Gao, Y.; Liang, J. LOPdM: A Low-power On-device Predictive Maintenance System Based on Self-powered Sensing and
TinyML. IEEE Trans. Instrum. Meas. 2023, 72, 2525213. [CrossRef]
14. Savanna, R.L.; Hanyurwimfura, D.; Nsenga, J.; Rwigema, J. A Wearable Device for Respiratory Diseases Monitoring in Crowded
Spaces. Case Study of COVID-19. In Proceedings of the International Congress on Information and Communication Technology,
London, UK, 20–23 February 2023; Springer: Berlin/Heidelberg, Germany, 2023; pp. 515–528.
15. Nguyen, H.T.; Mai, N.D.; Lee, B.G.; Chung, W.Y. Behind-the-Ear EEG-Based Wearable Driver Drowsiness Detection System Using
Embedded Tiny Neural Networks. IEEE Sens. J. 2023, 23, 23875–23892. [CrossRef]
16. Hussein, D.; Bhat, G. SensorGAN: A Novel Data Recovery Approach for Wearable Human Activity Recognition. ACM Trans.
Embed. Comput. Syst. 2023. [CrossRef]
17. Zacharia, A.; Zacharia, D.; Karras, A.; Karras, C.; Giannoukou, I.; Giotopoulos, K.C.; Sioutas, S. An Intelligent Microprocessor
Integrating TinyML in Smart Hotels for Rapid Accident Prevention. In Proceedings of the 2022 7th South-East Europe Design
Automation, Computer Engineering, Computer Networks and Social Media Conference (SEEDA-CECNSM), Ioannina, Greece,
23–25 September 2022; pp. 1–7. [CrossRef]
18. Atanane, O.; Mourhir, A.; Benamar, N.; Zennaro, M. Smart Buildings: Water Leakage Detection Using TinyML. Sensors 2023, 23,
9210. [CrossRef]
19. Malche, T.; Maheshwary, P.; Tiwari, P.K.; Alkhayyat, A.H.; Bansal, A.; Kumar, R. Efficient solid waste inspection through
drone-based aerial imagery and TinyML vision model. Trans. Emerg. Telecommun. Technol. 2023, e4878. [CrossRef]
20. Hammad, S.S.; Iskandaryan, D.; Trilles, S. An unsupervised TinyML approach applied to the detection of urban noise anomalies
under the smart cities environment. Internet Things 2023, 23, 100848. [CrossRef]
21. Priya, S.K.; Balaganesh, N.; Karthika, K.P. Integration of AI, Blockchain, and IoT Technologies for Sustainable and Secured Indian
Public Distribution System. In AI Models for Blockchain-Based Intelligent Networks in IoT Systems: Concepts, Methodologies, Tools, and
Applications; Springer: Berlin/Heidelberg, Germany, 2023; pp. 347–371.
22. Flores, T.; Silva, M.; Azevedo, M.; Medeiros, T.; Medeiros, M.; Silva, I.; Dias Santos, M.M.; Costa, D.G. TinyML for Safe Driving:
The Use of Embedded Machine Learning for Detecting Driver Distraction. In Proceedings of the 2023 IEEE International
Workshop on Metrology for Automotive (MetroAutomotive), Modena, Italy, 28–30 June 2023; pp. 62–66. [CrossRef]
23. Nkuba, C.K.; Woo, S.; Lee, H.; Dietrich, S. ZMAD: Lightweight Model-Based Anomaly Detection for the Structured Z-Wave
Protocol. IEEE Access 2023, 11, 60562–60577. [CrossRef]
24. Shabir, M.Y.; Torta, G.; Basso, A.; Damiani, F. Toward Secure TinyML on a Standardized AI Architecture. In Device-Edge-Cloud
Continuum: Paradigms, Architectures and Applications; Springer: Berlin/Heidelberg, Germany, 2023; pp. 121–139.
25. Tsoukas, V.; Gkogkidis, A.; Boumpa, E.; Papafotikas, S.; Kakarountas, A. A Gas Leakage Detection Device Based on the Technology
of TinyML. Technologies 2023, 11, 45. [CrossRef]
26. Hayajneh, A.M.; Aldalahmeh, S.A.; Alasali, F.; Al-Obiedollah, H.; Zaidi, S.A.; McLernon, D. Tiny machine learning on the edge: A
framework for transfer learning empowered unmanned aerial vehicle assisted smart farming. IET Smart Cities 2023. [CrossRef]
27. Adeola, J.O.; Degila, J.; Zennaro, M. Recent Advances in Plant Diseases Detection With Machine Learning: Solution for Developing
Countries. In Proceedings of the 2022 IEEE International Conference on Smart Computing (SMARTCOMP), Helsinki, Finland,
20–24 June 2022; pp. 374–380. [CrossRef]
28. Tsoukas, V.; Gkogkidis, A.; Kakarountas, A. A TinyML-Based System for Smart Agriculture. In Proceedings of the 26th
Pan-Hellenic Conference on Informatics, New York, NY, USA, 25–27 November 2023; pp. 207–212. [CrossRef]
29. Nicolas, C.; Naila, B.; Amar, R.C. TinyML Smart Sensor for Energy Saving in Internet of Things Precision Agriculture platform.
In Proceedings of the 2022 Thirteenth International Conference on Ubiquitous and Future Networks (ICUFN), Barcelona, Spain,
5–8 July 2022; pp. 256–259. [CrossRef]
30. Nicolas, C.; Naila, B.; Amar, R.C. Energy efficient Firmware over the Air Update for TinyML models in LoRaWAN agricultural
networks. In Proceedings of the 2022 32nd International Telecommunication Networks and Applications Conference (ITNAC),
Wellington, New Zealand, 30 November–2 December 2022; pp. 21–27. [CrossRef]
31. Viswanatha, V.; Ramachandra, A.C.; Hegde, P.T.; Raghunatha Reddy, M.V.; Hegde, V.; Sabhahit, V. Implementation of Smart
Security System in Agriculture fields Using Embedded Machine Learning. In Proceedings of the 2023 International Conference
on Applied Intelligence and Sustainable Computing (ICAISC), Zakopane, Poland, 18–22 June 2023; pp. 1–6. [CrossRef]
32. Botero-Valencia, J.; Barrantes-Toro, C.; Marquez-Viloria, D.; Pearce, J.M. Low-cost air, noise, and light pollution measuring station
with wireless communication and tinyML. HardwareX 2023, 16, e00477. [CrossRef]
33. Li, T.; Luo, J.; Liang, K.; Yi, C.; Ma, L. Synergy of Patent and Open-Source-Driven Sustainable Climate Governance under Green
AI: A Case Study of TinyML. Sustainability 2023, 15, 13779. [CrossRef]
34. Ihoume, I.; Tadili, R.; Arbaoui, N.; Benchrifa, M.; Idrissi, A.; Daoudi, M. Developing a TinyML-Oriented Deep Learning Model
for an Intelligent Greenhouse Microclimate Control from Multivariate Sensed Data. In Intelligent Sustainable Systems: Selected
Papers of WorldS4 2022; Springer: Berlin/Heidelberg, Germany, 2023; Volume 2, pp. 283–291.
35. Prakash, S.; Stewart, M.; Banbury, C.; Mazumder, M.; Warden, P.; Plancher, B.; Reddi, V.J. Is TinyML Sustainable? Assessing the
Environmental Impacts of Machine Learning on Microcontrollers. arXiv 2023, arXiv:2301.11899.
36. Soni, S.; Khurshid, A.; Minase, A.M.; Bonkinpelliwar, A. A TinyML Approach for Quantification of BOD and COD in Water.
In Proceedings of the 2023 2nd International Conference on Paradigm Shifts in Communications Embedded Systems, Machine
Learning and Signal Processing (PCEMS), Nagpur, India, 5–6 April 2023; pp. 1–6. [CrossRef]
Future Internet 2024, 16, 42 27 of 29

37. Arratia, B.; Prades, J.; Peña-Haro, S.; Cecilia, J.M.; Manzoni, P. BODOQUE: An Energy-Efficient Flow Monitoring System for
Ephemeral Streams. In Proceedings of the Twenty-fourth International Symposium on Theory, Algorithmic Foundations, and
Protocol Design for Mobile Networks and Mobile Computing, Washington, DC, USA, 23–26 October 2023; pp. 358–363.
38. Wardana, I.N.K.; Fahmy, S.A.; Gardner, J.W. TinyML Models for a Low-Cost Air Quality Monitoring Device. IEEE Sens. Lett.
2023, 7, 1–4. [CrossRef]
39. Sanchez-Iborra, R. LPWAN and Embedded Machine Learning as Enablers for the Next Generation of Wearable Devices. Sensors
2021, 21, 5218. [CrossRef]
40. Hussein, M.; Mohammed, Y.S.; Galal, A.I.; Abd-Elrahman, E.; Zorkany, M. Smart Cognitive IoT Devices Using Multi-Layer
Perception Neural Network on Limited Microcontroller. Sensors 2022, 22, 5106. [CrossRef]
41. Prakash, S.; Callahan, T.; Bushagour, J.; Banbury, C.; Green, A.V.; Warden, P.; Ansell, T.; Reddi, V.J. CFU Playground: Full-Stack
Open-Source Framework for Tiny Machine Learning (TinyML) Acceleration on FPGAs. In Proceedings of the 2023 IEEE
International Symposium on Performance Analysis of Systems and Software (ISPASS), Raleigh, NC, USA, 23–25 April 2023;
pp. 157–167. [CrossRef]
42. Gibbs, M.; Kanjo, E. Realising the Power of Edge Intelligence: Addressing the Challenges in AI and tinyML Applications for
Edge Computing. In Proceedings of the 2023 IEEE International Conference on Edge Computing and Communications (EDGE),
Chicago, IL, USA, 2–8 July 2023; pp. 337–343. [CrossRef]
43. Shafique, M.; Theocharides, T.; Reddy, V.J.; Murmann, B. TinyML: Current Progress, Research Challenges, and Future Roadmap.
In Proceedings of the 2021 58th ACM/IEEE Design Automation Conference (DAC), San Francisco, CA, USA, 5–9 December 2021;
pp. 1303–1306. [CrossRef]
44. Banbury, C.R.; Reddi, V.J.; Lam, M.; Fu, W.; Fazel, A.; Holleman, J.; Huang, X.; Hurtado, R.; Kanter, D.; Lokhmotov, A.; et al.
Benchmarking tinyml systems: Challenges and direction. arXiv 2020, arXiv:2003.04821.
45. Ooko, S.O.; Muyonga Ogore, M.; Nsenga, J.; Zennaro, M. TinyML in Africa: Opportunities and Challenges. In Proceedings of the
2021 IEEE Globecom Workshops (GC Wkshps), Madrid, Spain, 7–11 December 2021; pp. 1–6. [CrossRef]
46. Sanchez-Iborra, R.; Skarmeta, A.F. TinyML-Enabled Frugal Smart Objects: Challenges and Opportunities. IEEE Circuits Syst.
Mag. 2020, 20, 4–18. [CrossRef]
47. Mishra, N.; Lin, C.C.; Chang, H.T. A Cognitive Oriented Framework for IoT Big-data Management Prospective. In Proceedings
of the 2014 IEEE International Conference on Communiction Problem-Solving, Beijing, China, 5–7 December 2014; pp. 124–127.
[CrossRef]
48. Mishra, N.; Lin, C.C.; Chang, H.T. A cognitive adopted framework for IoT big-data management and knowledge discovery
prospective. Int. J. Distrib. Sens. Netw. 2015, 11, 718390. [CrossRef]
49. Huang, X.; Fan, J.; Deng, Z.; Yan, J.; Li, J.; Wang, L. Efficient IoT data management for geological disasters based on big
data-turbocharged data lake architecture. ISPRS Int. J. Geo-Inf. 2021, 10, 743. [CrossRef]
50. Oktian, Y.E.; Lee, S.G.; Lee, B.G. Blockchain-based continued integrity service for IoT big data management: A comprehensive
design. Electronics 2020, 9, 1434. [CrossRef]
51. Lê, M.T.; Arbel, J. TinyMLOps for real-time ultra-low power MCUs applied to frame-based event classification. In Proceedings of
the 3rd Workshop on Machine Learning and Systems, Rome, Italy, 8 May 2023; pp. 148–153.
52. Doyu, H.; Morabito, R.; Brachmann, M. A TinyMLaaS Ecosystem for Machine Learning in IoT: Overview and Research
Challenges. In Proceedings of the 2021 International Symposium on VLSI Design, Automation and Test (VLSI-DAT), Hsinchu,
Taiwan, 19–22 April 2021; pp. 1–5. [CrossRef]
53. Lin, J.; Zhu, L.; Chen, W.M.; Wang, W.C.; Han, S. Tiny Machine Learning: Progress and Futures [Feature]. IEEE Circuits Syst. Mag.
2023, 23, 8–34. [CrossRef]
54. Schizas, N.; Karras, A.; Karras, C.; Sioutas, S. TinyML for Ultra-Low Power AI and Large Scale IoT Deployments: A Systematic
Review. Future Internet 2022, 14, 363. [CrossRef]
55. Alajlan, N.N.; Ibrahim, D.M. TinyML: Enabling of Inference Deep Learning Models on Ultra-Low-Power IoT Edge Devices for AI
Applications. Micromachines 2022, 13, 851. [CrossRef]
56. Han, H.; Siebert, J. TinyML: A Systematic Review and Synthesis of Existing Research. In Proceedings of the 2022 International
Conference on Artificial Intelligence in Information and Communication (ICAIIC), Jeju Island, Republic of Korea, 21–24 February
2022; pp. 269–274. [CrossRef]
57. Andrade, P.; Silva, I.; Silva, M.; Flores, T.; Cassiano, J.; Costa, D.G. A TinyML Soft-Sensor Approach for Low-Cost Detection and
Monitoring of Vehicular Emissions. Sensors 2022, 22, 3838. [CrossRef]
58. Wongthongtham, P.; Kaur, J.; Potdar, V.; Das, A. Big data challenges for the Internet of Things (IoT) paradigm. In Connected
Environments for the Internet of Things: Challenges and Solutions; Springer: Berlin/Heidelberg, Germany, 2017; pp. 41–62.
59. Shu, L.; Mukherjee, M.; Pecht, M.; Crespi, N.; Han, S.N. Challenges and Research Issues of Data Management in IoT for
Large-Scale Petrochemical Plants. IEEE Syst. J. 2018, 12, 2509–2523. [CrossRef]
60. Gore, R.; Valsan, S.P. Big Data challenges in smart Grid IoT (WAMS) deployment. In Proceedings of the 2016 8th International
Conference on Communication Systems and Networks (COMSNETS), Bangalore, India, 5–10 January 2016; pp. 1–6. [CrossRef]
61. Touqeer, H.; Zaman, S.; Amin, R.; Hussain, M.; Al-Turjman, F.; Bilal, M. Smart home security: Challenges, issues and solutions at
different IoT layers. J. Supercomput. 2021, 77, 14053–14089. [CrossRef]
Future Internet 2024, 16, 42 28 of 29

62. Kumari, K.; Mrunalini, M. A Framework for Analysis of Incompleteness and Security Challenges in IoT Big Data. Int. J. Inf. Secur.
Priv. (IJISP) 2022, 16, 1–13. [CrossRef]
63. Zhang, Y.; Adin, V.; Bader, S.; Oelmann, B. Leveraging Acoustic Emission and Machine Learning for Concrete Materials Damage
Classification on Embedded Devices. IEEE Trans. Instrum. Meas. 2023, 72, 2525108. [CrossRef]
64. Moin, A.; Challenger, M.; Badii, A.; Günnemann, S. Supporting AI Engineering on the IoT Edge through Model-Driven TinyML.
In Proceedings of the 2022 IEEE 46th Annual Computers, Software, and Applications Conference (COMPSAC), Los Alamitos, CA,
USA, 27 June–1 July 2022; pp. 884–893. [CrossRef]
65. David, R.; Duke, J.; Jain, A.; Janapa Reddi, V.; Jeffries, N.; Li, J.; Kreeger, N.; Nappier, I.; Natraj, M.; Wang, T.; et al. Tensorflow lite
micro: Embedded machine learning for tinyml systems. Proc. Mach. Learn. Syst. 2021, 3, 800–811.
66. Qian, C.; Einhaus, L.; Schiele, G. ElasticAI-Creator: Optimizing Neural Networks for Time-Series-Analysis for on-Device Machine
Learning in IoT Systems. In Proceedings of the 20th ACM Conference on Embedded Networked Sensor Systems; Association for
Computing Machinery: New York, NY, USA, 2023; pp. 941–946. [CrossRef]
67. Giordano, M.; Baumann, N.; Crabolu, M.; Fischer, R.; Bellusci, G.; Magno, M. Design and Performance Evaluation of an
Ultralow-Power Smart IoT Device with Embedded TinyML for Asset Activity Monitoring. IEEE Trans. Instrum. Meas. 2022,
71, 2510711. [CrossRef]
68. Bamoumen, H.; Temouden, A.; Benamar, N.; Chtouki, Y. How TinyML Can be Leveraged to Solve Environmental Problems:
A Survey. In Proceedings of the 2022 International Conference on Innovation and Intelligence for Informatics, Computing, and
Technologies (3ICT), Sakheer, Bahrain, 20–21 November 2022; pp. 338–343. [CrossRef]
69. Athanasakis, G.; Filios, G.; Katsidimas, I.; Nikoletseas, S.; Panagiotou, S.H. TinyML-based approach for Remaining Useful Life
Prediction of Turbofan Engines. In Proceedings of the 2022 IEEE 27th International Conference on Emerging Technologies and
Factory Automation (ETFA), Stuttgart, Germany, 6–9 September 2022; pp. 1–8. [CrossRef]
70. Silva, M.; Signoretti, G.; Flores, T.; Andrade, P.; Silva, J.; Silva, I.; Sisinni, E.; Ferrari, P. A data-stream TinyML compression
algorithm for vehicular applications: A case study. In Proceedings of the 2022 IEEE International Workshop on Metrology for
Industry 4.0 & IoT (MetroInd4.0&IoT), Trento, Italy, 7–9 June 2022; pp. 408–413. [CrossRef]
71. Ostrovan, E. TinyML On-Device Neural Network Training. Master’s Thesis, Politecnico di Milano, Milan, Italy, 2022.
72. Signoretti, G.; Silva, M.; Andrade, P.; Silva, I.; Sisinni, E.; Ferrari, P. An Evolving TinyML Compression Algorithm for IoT
Environments Based on Data Eccentricity. Sensors 2021, 21, 4153. [CrossRef]
73. Sharif, U.; Mueller-Gritschneder, D.; Stahl, R.; Schlichtmann, U. Efficient Software-Implemented HW Fault Tolerance for TinyML
Inference in Safety-critical Applications. In Proceedings of the 2023 Design, Automation & Test in Europe Conference & Exhibition
(DATE), Antwerp, Belgium, 17–19 April 2023; IEEE: Piscataway, NJ, USA, 2023; pp. 1–6.
74. Fedorov, I.; Matas, R.; Tann, H.; Zhou, C.; Mattina, M.; Whatmough, P. UDC: Unified DNAS for compressible TinyML models.
arXiv 2022, arXiv:2201.05842.
75. Nadalini, D.; Rusci, M.; Benini, L.; Conti, F. Reduced Precision Floating-Point Optimization for Deep Neural Network On-Device
Learning on MicroControllers. arXiv 2023, arXiv:2305.19167.
76. Silva, M.; Medeiros, T.; Azevedo, M.; Medeiros, M.; Themoteo, M.; Gois, T.; Silva, I.; Costa, D.G. An Adaptive TinyML
Unsupervised Online Learning Algorithm for Driver Behavior Analysis. In Proceedings of the 2023 IEEE International Workshop
on Metrology for Automotive (MetroAutomotive), Modena, Italy, 28–30 June 2023; IEEE: Piscataway, NJ, USA, 2023; pp. 199–204.
77. Pereira, E.S.; Marcondes, L.S.; Silva, J.M. On-Device Tiny Machine Learning for Anomaly Detection Based on the Extreme Values
Theory. IEEE Micro 2023, 43, 58–65. [CrossRef]
78. Zhuo, S.; Chen, H.; Ramakrishnan, R.K.; Chen, T.; Feng, C.; Lin, Y.; Zhang, P.; Shen, L. An empirical study of low precision
quantization for tinyml. arXiv 2022, arXiv:2203.05492.
79. Krishna, A.; Nudurupati, S.R.; Dwivedi, P.; van Schaik, A.; Mehendale, M.; Thakur, C.S. RAMAN: A Re-configurable and Sparse
tinyML Accelerator for Inference on Edge. arXiv 2023, arXiv:2306.06493.
80. Ren, H.; Anicic, D.; Runkler, T.A. TinyReptile: TinyML with Federated Meta-Learning. arXiv 2023, arXiv:2304.05201.
81. Ren, H.; Anicic, D.; Runkler, T.A. Towards Semantic Management of On-Device Applications in Industrial IoT. ACM Trans.
Internet Technol. 2022, 22, 1–30. [CrossRef]
82. Chen, J.I.Z.; Huang, P.F.; Pi, C.S. The implementation and performance evaluation for a smart robot with edge computing
algorithms. Ind. Robot. Int. J. Robot. Res. Appl. 2023, 50, 581–594. [CrossRef]
83. Mohammed, M.; Srinivasagan, R.; Alzahrani, A.; Alqahtani, N.K. Machine-Learning-Based Spectroscopic Technique for Non-
Destructive Estimation of Shelf Life and Quality of Fresh Fruits Packaged under Modified Atmospheres. Sustainability 2023, 15,
12871. [CrossRef]
84. Koufos, K.; EI Haloui, K.; Dianati, M.; Higgins, M.; Elmirghani, J.; Imran, M.A.; Tafazolli, R. Trends in Intelligent Communication
Systems: Review of Standards, Major Research Projects, and Identification of Research Gaps. J. Sens. Actuator Netw. 2021, 10, 60.
[CrossRef]
85. Ahad, M.A.; Tripathi, G.; Zafar, S.; Doja, F. IoT Data Management—Security Aspects of Information Linkage in IoT Systems.
In Principles of Internet of Things (IoT) Ecosystem: Insight Paradigm; Peng, S.L., Pal, S., Huang, L., Eds.; Springer International
Publishing: Cham, Switzerland, 2020; pp. 439–464. [CrossRef]
86. Liu, R.W.; Nie, J.; Garg, S.; Xiong, Z.; Zhang, Y.; Hossain, M.S. Data-Driven Trajectory Quality Improvement for Promoting
Intelligent Vessel Traffic Services in 6G-Enabled Maritime IoT Systems. IEEE Internet Things J. 2021, 8, 5374–5385. [CrossRef]
Future Internet 2024, 16, 42 29 of 29

87. Hnatiuc, B.; Paun, M.; Sintea, S.; Hnatiuc, M. Power management for supply of IoT Systems. In Proceedings of the 2022
26th International Conference on Circuits, Systems, Communications and Computers (CSCC), Crete, Greece, 19–22 July 2022;
pp. 216–221. [CrossRef]
88. Rajeswari, S.; Ponnusamy, V. AI-Based IoT analytics on the cloud for diabetic data management system. In Integrating AI in IoT
Analytics on the Cloud for Healthcare Applications; IGI Global: Hershey, PA, USA, 2022; pp. 143–161.
89. Karras, A.; Karras, C.; Giotopoulos, K.C.; Tsolis, D.; Oikonomou, K.; Sioutas, S. Peer to peer federated learning: Towards
decentralized machine learning on edge devices. In Proceedings of the 2022 7th South-East Europe Design Automation, Computer
Engineering, Computer Networks and Social Media Conference (SEEDA-CECNSM), Ioannina, Greece, 23–25 September 2022;
IEEE: Piscataway, NJ, USA, 2022; pp. 1–9.
90. Karras, A.; Karras, C.; Giotopoulos, K.C.; Tsolis, D.; Oikonomou, K.; Sioutas, S. Federated Edge Intelligence and Edge Caching
Mechanisms. Information 2023, 14, 414. [CrossRef]
91. Karras, A.; Karras, C.; Karydis, I.; Avlonitis, M.; Sioutas, S. An Adaptive, Energy-Efficient DRL-Based and MCMC-Based Caching
Strategy for IoT Systems. In Algorithmic Aspects of Cloud Computing; Chatzigiannakis, I., Karydis, I., Eds.; Springer: Cham,
Switzerland, 2024; pp. 66–85.
92. Meddeb, M.; Dhraief, A.; Belghith, A.; Monteil, T.; Drira, K. How to cache in ICN-based IoT environments? In Proceedings
of the 2017 IEEE/ACS 14th International Conference on Computer Systems and Applications (AICCSA), Hammamet, Tunisia,
30 October–3 November 2017; IEEE: Piscataway, NJ, USA, 2017; pp. 1117–1124.
93. Wang, S.; Zhang, X.; Zhang, Y.; Wang, L.; Yang, J.; Wang, W. A survey on mobile edge networks: Convergence of computing,
caching and communications. IEEE Access 2017, 5, 6757–6779. [CrossRef]
94. Zhong, C.; Gursoy, M.C.; Velipasalar, S. A deep reinforcement learning-based framework for content caching. In Proceedings
of the 2018 52nd Annual Conference on Information Sciences and Systems (CISS), Princeton, NJ, USA, 21–23 March 2018;
IEEE: Piscataway, NJ, USA, 2018; pp. 1–6.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.
Reproduced with permission of copyright owner. Further reproduction
prohibited without permission.