AIoT stands for Artificial Intelligence of Things and refers to the synergy between Internet of Things and artificial intelligence, where new frontiers are opening for developing intelligent autonomous systems. The integration of AI and IoT enables devices to operate beyond mere data collection and transmission by analyzing data in real time, making independent decisions, and adapting to environmental changes. However, AIoT has several deployment challenges, each potentially being limiting factors against the potential of AIoT and security or privacy of data. Some of the most important integration challenges for AIoT involve the discussion of device and protocol compatibility on one hand and vast amounts of data on the other, latency, and power consumption. The article further discusses the complexity brought in by a multitude of heterogeneous hardware and software platforms, making standardization and inter-operability between systems difficult. Much attention is given to the problems of security, as AIoT systems are becoming gradually vulnerable to possible cybersecurity attacks, which include unauthorized access to data, loss, or leakage. The article also covers the potential threats regarding the privacy and security of AI algorithms, including data poisoning attacks and manipulations with machine learning models. Any of these might be solved by developing secure mechanisms for authentication and authorization, advanced encryption methods, and attack-resistant AI models. Finally, the article points out the relevance of standardization and the development of international protocols with the aim of guaranteeing interoperability and security of the AIoT systems. Distributed computing, including edge computing, is also fundamental to the decrease in latency and increase in efficiency for the processing of data. The next section shall discuss the need for compliance with legislation on the protection of personal data and privacy and the application of security principles in the whole Cycle of creation and operation of an AIoT system. This paper presents the overall landscape of current challenges and security issues related to the field of AIoT and provides guidelines for researchers, developers, and practitioners on how to integrate AI and IoT efficiently.
One of the important challenges in integrating AI and IoT is interoperability between
disparate devices with different communication protocols. IoT connotes a number of devices
from various manufacturers; many of these may have their individual hardware platforms,
operating systems, and protocols [
Hardware heterogeneity means that these devices have different computational powers,
memory, and energy resources. Some of them are capable of complex AI algorithms, while
others are simple in structure. All this constitutes a big challenge in developing software with
which such diverse devices can provide the needed functionality realization [
The communication protocols also vary, starting from short-range wireless technologies,
including Bluetooth and Zigbee, up to their long-range protocols like LoRaWAN. Others
include MQTT, CoAP, and HTTP, among others, which in one way or another have certain
advantages and limits. Since there is no preference for the use of a common standard, it
becomes hard to integrate devices into applications and could demand additional gateways or
protocol converters [
Furthermore, the format and model of information are not the same in all devices, which
may be an issue during the processing and analysis of the collected information. AI algorithms
require relevant and consistent data in order to learn and operate effectively. Such differences
in data format only lead to errors or inaccuracies in the models, thereby leading to poor
performance [
The use of middleware serves for abstraction from peculiarities of devices and unification
of interaction interface. Certainly, interoperability issues will be improved only in the case of
collaboration between manufacturers and developers. An open platform, an ecosystem that
allows co-design and knowledge sharing, could really accelerate the integration processes:
based on this, a development of semantic technologies and ontologies will enable unification
of data models, enabling semantic interoperability among systems [
The AIoT model is subject to constant threats, as shown in Figure 1.
The integration of AI and IoT generates such huge volumes of data that processing and
analysis introduce serious challenges. IoT devices constantly collect data from the
environment, sensors, and users that needs to be processed as quickly as possible for real-time
decisions. Traditional methods are becoming inefficient and resource-intensive while
processing this amount of data. One of the main difficulties is the deficit of computational
resources at the IoT device level. Because of these power and memory limitations, most of
them cannot run complex AI algorithms. Transferring all collected data to be processed by
cloud services may lead to network overload, delays, and swelling security risks [
In such circumstances, edge computing is becoming crucial. Data processing closer to, or
even on the device itself, reduces latency and decreases the loads on networks [
Another crucial factor during the processing of extensive volumes of data is power
consumption. Sometimes, running complex AI algorithms is power-consuming, and IoT
devices may not have this power every time. So, energy-efficient algorithm development and
the use of some special hardware-like neuromorphic processors-will reduce the power
consumption [
In general, big data processing in AIoT needs a blend of technical solution, resource optimization, and security management. Meeting these challenges effectively will allow the full exploitation of AIoT's potential and ensure successful implementations either in industry or everyday life.
Data latency is a major barrier to AIoT integration, as it can negatively impact system
performance, especially for systems operating in real-time. Instant information exchange and
fast decision-making are essential for many AIoT applications, including industrial control
systems, medical devices, and autonomous vehicles. Many factors, such as limited network
bandwidth, heavy traffic, or distance between devices and data centres, can cause delays [
In typical IoT systems, data is often sent to cloud servers for processing, which can cause
significant delays, especially if network resources are limited. In mission-critical applications
where milliseconds can matter, this is not always acceptable. For example, delays in traffic
management systems can lead to accidents [
The emergence of edge computing, which processes data closer to the data source at the
network edge, offers a solution to this problem. Since the data does not need to be transported
over long distances for processing, it reduces latency. Real-time pre-analysis, filtering and
decision-making can be performed by edge devices, which send only the data that is needed
for further analysis or long-term storage to the centre [
In addition, because 5G and other high-speed network technologies have higher bandwidth
and lower latency, they can significantly reduce data latency. New AIoT applications now
have prospects that were previously impossible due to network limitations [
Optimising the data transfer protocol is another important factor. Using lightweight
protocols such as MQTT or CoAP, which are designed specifically for the Internet of Things,
helps to reduce overhead and speed up transmission. These protocols can function well even
in low-bandwidth networks because they are designed to work in resource-constrained
environments [
Thus, minimising data delays is essential for successful AIoT integration. This requires a comprehensive strategy that includes the creation of adaptive AI algorithms, optimisation of transmission protocols, and the introduction of new network infrastructure technologies. This is the only way to guarantee the efficient and reliable operation of AIoT solutions in real time.
Another point in implementation that arises as a challenge pertains to power consumption:
most IoT devices are either powered on batteries or at very low levels of power input. Adding
AI functionality to devices from these classes increases the need for power significantly and
amply necessitates recharging the battery more often [
Devices with limited resources of energy mostly are unable to perform the complex
algorithm calculations required by AI. This therefore constrains the possibilities of integrating
AI at the device level, and hence, one has to make a trade-off between functionality and
energy efficiency. Another approach is the use of specialized hardware that has been
optimized for low-power AI jobs. For instance, neuromorphic processors or
ApplicationSpecific Integrated Circuits designed to execute particular AI operations could cut energy
costs by an order of magnitude [
In addition to that, there needs to be energy-efficient machine learning algorithms and
models. This includes developing lightweight models with fewer parameters, using techniques
like quantization and pruning to match a model's size without significant loss of accuracy, and
low-power modes of operation [
Employing edge computing will further help in reducing energy consumption by avoiding
the continuous transfer of large volumes of data to the cloud, which is itself an
energyintensive process. Besides, edge devices can carry out simple processing of data and transmit
only information needed to the centre [
Most AIoT systems run physical processes and are capable of processing highly sensitive
information. These threats, in the field, such as unauthorized access to systems, data leakage,
or physical harm, have a great influence. One of the major threats involves unauthorized
access to AIoT devices and networks. This means access to the system through exploited
software vulnerabilities, poor passwords, or no encryption, attackers would successfully get
hold of it, after which they can steal information, disrupt devices, or use them as a part of a
botnet to attack other systems [
In most instances, it is grossly observed that these AIoT systems collect and process personal
information of users, which gives way to a series of grave privacy issues, we have proposed a
Table of Security Risks in AIoT and Corresponding Protection Methods (Table 1). Improper
handling of such data can result in breaches of privacy, discrimination, or other negative
consequences. Therefore, privacy necessarily demands compliance with legislation, such as
the General Data Protection Regulation (GDPR) by the European Union, and ethical data
processing principles should be implemented [
Attackers may gain access to devices or networks using vulnerabilities or weak passwords. - Implement strong authentication and authorization - Use multi-factor authentication - Regularly update credentials - Implement mechanisms to detect and block DoS attacks - Use firewalls and intrusion prevention systems (IPS) - Validate and verify training data - Use anomaly detection methods - Develop robust AI algorithms resistant to attacks - Implement defenses against adversarial examples - Regularize and enhance model robustness - Monitor AI outputs for anomalies - Verify integrity and authenticity of updates - Use digital signatures - Control suppliers and partners
Lack of Transparency in Data
Processing
The AI algorithms used in the AIoT systems can be targeted by various forms of attacks. For instance, poisoning attacks are performed by manipulating input or training data with the intent of skewing the model's output. In turn, it would result in incorrect decisions that, within the context of AIoT, can have serious consequences [41].
Another type of threat is that of attacks through the injection of crafted input data, causing the model to make mistakes or generate certain results. They are called adversarial example attacks. For such types of threats, attack-resistant AI models need to be developed. It includes regularization, integrity checks of training data, monitoring abnormal model behavior, and providing explainable AI mechanisms to clear up how the model makes its decisions [42].Besides, it's relevant to ensure security in infrastructures where AI algorithms are executed on the servers, databases, and networks against cyber threats. This is highly in contrast to routine cybersecurity training of staff, as well as the development and application of security policies [43].
Thus, security issues in AIoT are multifaceted and require a comprehensive approach. Ensuring cybersecurity, data privacy protection, and the security of AI algorithms are critical for user trust and the successful implementation of AIoT technologies in various spheres of life [44].
The workable integration of AIoT into life is impossible without solving such problems as interoperability, processing of data, security, and meeting the regulatory requirements comprehensively. Below we consider the basic ways of solving listed problems.
Interoperability between the huge number of devices and systems is a key role that standardisation plays in AIoT. Interoperability issues arise due to the lack of standards common for the market unification; this seriously complicates integration of newer technologies. The use of common applied communication protocols, data formats, and interfaces will make the devices of different manufacturers intercommunicate easily. ISO, IEEE, IETF, and many international organizations are working on the development of standards related to IoT and AI. For example, the IEEE 2413 standard defines an architecture for IoT that promotes interoperability and security [45].
Standardization will enhance the security of AIoT. Security standards allow for the definition of the necessary level of protection against diverse cyber threats, such as IEC 62443 for industrial systems. Their adoption minimizes any weaknesses and inspires more user confidence in AIoT technologies [46]. Besides contributing to various standardization processes, which give organizations an opportunity to have their say in developing the industry and ensuring that standards meet the organization's needs, manufacturers and developers must work together with the regulators. They constitute a vital part of successful standardization processes that are the major interoperability barriers.
Distributed data processing is an efficient approach to managing large amounts of information in AIoT systems. The use of:
Edge computing for data processing near devices with reduced latency and network load. This is invaluable for real-time applications like autonomous vehicles or industrial control systems [47].
Fog computing provides complementary services by introducing data processing closer to the cloud, reducing the distance between devices and data centers. Fog nodes can conduct some initial processing, aggregate data, and enhance another layer of security.
Cloud computing remains essential for large data storage over the long term and power-intensive AI workloads, including machine learning model training. Combining edge, fog, and cloud creates a comprehensive system that maximizes resource utilization and reliability [48].
The basic protection against cyber threats of AIoT systems is to develop secure software. Security integration at all development phases, from design to testing, makes it possible to address vulnerabilities before attackers can exploit them. Security by Design principles involve embedding security mechanisms within the system architecture. Employ strong authentication and authorization, use the latest cryptographic algorithms to encrypt data, and ensure information integrity [49]. Regularly downloading and installing software updates and patches is crucial. Detecting anomalous behavior through security monitoring tools helps to quickly respond to and contain threats. Security testing, including penetration testing and code analysis, is essential for identifying vulnerabilities [50].
Compliance with laws and regulations is extremely important in the successful operation of AIoT systems. Personal data protection laws (e.g., GDPR in the European Union, CCPA in California) significantly raise the requirements for assessments when collecting and processing personal information. Organizations must be transparent about the data they collect and its intended use, while also seeking explicit user consent [51]. It is critically important to enact and enforce data processing rules by specifying policies and procedures, as well as appointing oversight personnel, such as a privacy officer. Another key aspect is adherence to ethical principles in AI development and usage, including preventing unlawful surveillance, safeguarding against algorithmic bias, and ensuring that AI decisions are open to inspection [52].
Artificial intelligence of things (AIoT) is a powerful convergence of two of the leading technologies of our time: Internet of Things (IoT) and artificial intelligence (AI). The combination of IoT data collection and transmission capabilities with intelligent AI algorithms opens up new horizons for innovation in various fields, including industry, healthcare, transport, and everyday life. However, the integration of these technologies is accompanied by a number of challenges that require careful analysis and resolution.
One of the main challenges is device and protocol compatibility. The variety of hardware, operating systems and communication protocols makes it difficult to interoperate between devices from different manufacturers. The lack of common standards leads to market fragmentation and increases the complexity of system integration. This problem can be solved through active standardisation, participation in international organisations and cooperation between manufacturers and developers.
Processing large amounts of data is another critical aspect. IoT devices generate huge amounts of information that need to be processed and analysed efficiently. Distributed data processing, including the use of edge and fog computing, can optimise the use of computing resources and reduce latency. The integration of different computing layers creates a flexible and scalable architecture that meets the requirements of different applications.
Security issues in AIoT are extremely important. The growing number of connected devices increases the potential risks of cyber threats. Developing secure software that takes into account security principles at all stages of the life cycle is essential to protect systems from attacks. The use of modern authentication methods, encryption and regular software updates help to improve security.
Regulation and compliance are essential to ensure confidentiality and protection of personal data. Compliance with international and local laws, such as GDPR, as well as the implementation of ethical principles in the development and use of AI, contribute to user trust and the successful implementation of AIoT technologies.
This article analyses in detail the main challenges of AIoT integration and suggests ways to overcome them. Standardisation, distributed data processing, secure software development, and legal compliance are key components of a comprehensive approach to addressing the challenges.
Recommendations for future research and practice include:
Active engagement in standardisation processes to promote interoperability and interoperability of systems.
Developing new technologies and data processing methods that increase the efficiency and scalability of AIoT.
Improving cybersecurity through the implementation of advanced security methods and ongoing staff training.
Adherence to ethical standards and legislation governing data processing and use to ensure user privacy and trust.
Artificial intelligence of things has the potential to radically change various aspects of our lives, increasing efficiency, comfort and safety. However, in order to realise this potential, it is necessary to overcome the existing challenges and ensure that AIoT technologies are developed responsibly and ethically. Collaboration between academia, industry, government organisations, and society at large is key to a successful AIoT future. [39] Davari, M., & Bertino, E. (2019). Access Control Model Extensions to Support Data
Privacy Protection based on GDPR. «IEEE Big Data». [40] Marques, J. F., & Bernardino, J. (2020). Analysis of Data Anonymization Techniques.
«International Conference on Data Science, E-learning and Information Systems». [41] Zhang, J., Song, B., Han, B., Liu, L., Niu, G., & Sugiyama, M. (2023). Assessing
Vulnerabilities of Adversarial Learning Algorithm through Poisoning Attacks. «arXiv». [42] Zheng, H., Zhao, J., & Tan, W. (2022). An Adversarial Defense Algorithm Based on Triplet
Network and Voting Decision. «IEEE ICUS». [43] Bouaziz, A., Nguyen, M.-D., Valdés, V., Cavalli, A., & Mallouli, W. (2023). Study on Adversarial Attacks Techniques, Learning Methods and Countermeasures: Application to Anomaly Detection. «International Conference on Information Systems Security and Privacy». [44] Ali, M., Hu, Y.-F., Luong, D., Oguntala, G., Li, J.-P., & Abdo, K. (2020). Adversarial Attacks on AI based Intrusion Detection System for Heterogeneous Wireless Communications Networks. «IEEE DASC». [45] Katrakazas, P., Kallinolitou, T., Markopoulou, S., & Chronopoulou, A. (2022). A Toolchain and Interoperability Framework to enhance privacy and individual control at the Edge. «IEEE ISC». [46] Prokhorenkov, D. (2022). Alternative Methodology and Framework for Assessing
Differential Privacy Constraints and Consequences From a GDPR Perspective. «IEEE». [47] Liu, Y., Hassan, K. A., Karlsson, M., Pang, Z., & Gong, S. (2019). A Data-Centric Internet of
Things Framework Based on Azure Cloud. «IEEE Access». [48] Tulkinbekov, K., & Kim, D. (2022). Blockchain-Enabled Approach for Big Data Processing in Edge Computing. «IEEE Internet of Things Journal». [49] Song, J., Gu, T., & Mohapatra, P. (2021). How BlockChain Can Help Enhance The Security And Privacy in Edge Computing? «ACM International Conference on Mobile Computing and Networking». [50] Ezeugwa, F. A. (2024). Evaluating the Integration of Edge Computing and Serverless Architectures for Enhancing Scalability and Sustainability in Cloud-based Big Data Management. «Journal of Engineering Research and Reports». [51] Issaoui, A., Örtensjö, J., & Islam, M. S. (2023). Exploring the General Data Protection Regulation (GDPR) compliance in cloud services: insights from Swedish public organizations on privacy compliance. «Financial Innovation». [52] Khalid, M. I., Ahmed, M., & Kim, J. (2023). Enhancing Data Protection in Dynamic Consent Management Systems: Formalizing Privacy and Security Definitions with Differential Privacy, Decentralization, and Zero-Knowledge Proofs. «Sensors». [53] Stukhliak, P., Martsenyuk, V., Totosko O., Stukhlyak, D., Didych, I. (2024). The use of neural networks for modeling the thermophysical characteristics of epoxy composites treated with electric spark water hammer. CEUR Workshop ProceedingsVolume 3742, CITI 2024. Ternopil 2nd International Workshop on Computer Information Technologies in Industry 4.0, Pages 13 – 242024. [54] Yasniy, O., Mykytyshyn, A., Didych, I., Kubashok, V., & Boiko, A. (2023). Application of artificial intelligence to improve the work of educational platforms. In ITTAP (pp. 605609). [55] Zhukovskyy, V., Shatnyi, S., Zhukovska, N., & Sverstiuk, A. (2021). Neural Network Clustering Technology for Cartographic Images Recognition. In IEEE EUROCON 2021 19th International Conference on Smart Technologies. IEEE EUROCON 2021 - 19th International Conference on Smart Technologies. IEEE. [56] Martsenyuk, V., Klos-Witkowska, A., Sverstiuk, A., Bahrii-Zaiats O., Bernas, M., Witos, K. (2021). Intelligent big data system based on scientific machine learning of cyberphysical systems of medical and biological processes. CEUR Workshop Proceedings, 2864, pp. 34–48. [57] Duda, O., Mykytyshyn, A., Mytnyk M., Stanko, A. (2023). Information technology sets formation and "TNTU Smart Campus" services network support. In 2nd International Workshop on Information Technologies: Theoretical and Applied Problems 2023, Ternopil, pp. 93-105.