While AI technology can initially solve some SDN problems, challenges remain before AI can be fully integrated with SDN. These include high-quality data, inter-domain communication, scalability, and security prediction. Additionally, this section will analyze how to use AI to fully integrate SDN. The advantages of centralized SDN control combine with development trends in fields such as 5G, network function virtualization, IoT, edge computing, information center networks, and wireless networks.
4.5.1. Future Challenges
High-Quality Data
Although mature big data processing technologies and AI-based data analysis technologies exist, AI algorithms are proposed in the field of image processing rather than in the field of networks. Most algorithms cannot be transferred or used directly in network scenarios, and network data is incomplete. To meet the data format requirements of AI algorithms and apply AI to the network, network data must be translated into matrix or vector form in advance. Therefore, the prerequisite for applying AI to SDN for accurate classification prediction is that training is precise and of high quality, compensating for deficiencies in data analysis formats, data compilation, etc. Current problems are mainly divided into two aspects: on the one hand, the size of the SDN network dataset, SDN network characteristics, and the AI-based SDN network model have not yet formed. On the other hand, high-quality data can only be cleaned manually, and no public datasets are available for researchers to experiment with. Therefore, all parties need to integrate high-quality public network datasets.
The emergence of multiple controllers aims to solve the disadvantages of increased complexity between switches and controllers caused by network-scale expansion. Research has shown that inter-domain information exchange is beneficial for improving network performance. However, data flow transmission between domains requires multiple controllers; such inter-domain design breaks the original SDN modularity. To account for the increase in multiple controllers and network complexity, traditional optimization methods are no longer sufficient to optimize the network. Although AI algorithms can be used in network optimization at all layers, analyzing the collected inter-domain information, including data link layer information, application plane information, and other information to optimize the network, such as routing mechanisms, congestion control, load balancing, etc., inter-layer optimization methods based on recursive multi-layer networks are required. Moreover, there is no mature model for inter-domain communication that facilitates future research.
On the other hand, inter-domain communication involves too many parameters, often affecting the entire system. Optimization algorithms that are truly suitable for inter-domain communication are lacking. Although many target optimization algorithms can solve the above problems with fewer parameters [
103], in real situations, multi-objective optimization algorithms will cause the Pareto non-dominated problem to fail due to too many parameters, leading to the failure of the entire optimization method. Therefore, the current research aim is to establish inter-domain communication problem models and optimization methods as soon as possible.
- b.
Scalability
The advantages of centralized SDN control and management have increasingly attracted researchers from academia and industry to dedicate themselves to this field. While SDN offers significant opportunities for network development, it encounters numerous practical obstacles. For instance, as the network scale expands, increasing the number of controllers is necessary to share tasks and enhance the performance of the centralized controller. However, as the number of controllers increases, issues related to controller placement and dynamic control arise [
104].
From this perspective, to improve the scalability of SDN networks, a multi-level RL (reinforcement learning) scheme can be considered, where the root controller acts as a high-level learning agent and the local controller as a low-level learning agent. Each lower-level learning agent acquires knowledge of directing traffic within its domain by leveraging the state information of its immediate network, thereby making optimal decisions. Conversely, the higher-level learning agent oversees inter-domain traffic management by maintaining a global perspective of the entire network. To expedite system responsiveness, the root control intermittently deploys the trained reinforcement learning (RL) model within the local controller. This trained RL model then directs the local controller to process inter-domain traffic directly. The multi-tiered RL approach not only diminishes the latency in handling commercial flows but also bolsters the scalability of the SDN network [
105]. Despite the theoretical consideration of controller categorization to enhance SDN scalability using the reinforcement learning framework, there is no real-world scenario to validate the feasibility of the algorithm. Therefore, real-world situations must be considered to ensure the robustness of the entire system.
- c.
Security Prediction
In the previous section on network security, it was explained that the SDN controller uses AI to analyze historical data, detect network anomalies, and address network attacks. However, anomaly detection is an adversarial process. Although a mature model can be constructed using historical data to predict the next attack, malicious attackers rarely succeed in the same manner; they continuously create new attacks to evade controller detection [
106]. In this context, using historical data to train an AI model is not an effective attack detection method, as it requires creating new attacks for detection. To address the aforementioned issue, two solutions can be adopted.
One solution is the Generative Adversarial Network (GAN) [
107]. GAN is a method to resolve problems by predicting new attacks. GAN consists of two neural networks: a generative neural network and a discriminative neural network. The generator generates fresh data, while the discriminator assesses the genuineness of the new data by comparing it with the authentic training dataset. Both the generator and discriminator undergo joint training to enhance the realism of the newly generated data. GAN has the capability to produce potential new attack data by leveraging historical data and integrating this newly generated data with historical data to train the machine learning (ML) model. This trained ML model can detect both known and potential new attacks. Upon detecting an attack, the controller proactively adjusts the flow table in the switch to thwart network attacks and restrict communication between the control plane and the data plane.
Another method employs a prior-posterior experience model [
108]. Prior experience and posterior experience can, respectively, represent historical network data information and real-time feedback information during the detection process in the model. This method retains historical information’s impact on controller detection while incorporating real-time feedback information into the model to correct it. This approach can accurately and in real-time analyze network attacks to ensure the network does not suffer new attacks and cause severe consequences.
4.5.2. Applications in Different Network Scenarios
5G networks are poised to accommodate a growing multitude of connected devices, furnish elevated user data rates, facilitate augmented mobile data traffic per unit geographical area, and diminish transmission latency and network power usage. In addition to possessing precise performance prerequisites, 5G also necessitates catering to heterogeneous services, devices, and access networks [
109,
110]. By separating the control and data planes, SDN significantly simplifies the network structure, reduces control signaling between network nodes, and improves load balancing, mobility management, and network control flexibility. The literature [
111] proposes a centralized wireless network controller based on SDN to control multiple nodes and gateways in the core network. The purpose of the 5G network architecture based on ISDN is to meet the functional and performance requirements of a new generation of services and devices. A key feature is the flexibility required to effectively support heterogeneous service sets, including machine-type communications and IoT communications.
These applications face significant challenges regarding latency, reliability, and end-to-end availability. Challenging objectives in terms of scalability have been added. To address the challenges posed by the heterogeneous wireless environment, the complexity of network management, the growing needs of mobile communications, and the diversified service requirements in 5G mobile networks, AI technology should be used to implement smarter networks in motion [
112].
- b.
Network Function Virtualization (NFV)
NFV enables the deployment of virtual network functions to improve performance, security, and management. Network functions are decoupled from the underlying dedicated hardware through NFV to provide flexibility in network architecture. NFVs are implemented in software running on real-world commercial devices and can be centrally controlled via an ISDN controller. Compared to traditional network functions implemented by dedicated hardware devices, NFVs have significantly lower operational and capital expenses and the potential to improve service agility. With the application of NFV, numerous related studies have become crucial services promoting network flexibility and cost-effectiveness. For instance, RL is used to dynamically create service function chains (SFC) based on resource usage to support efficient service delivery [
113,
114]. Combined with AI concepts, the online routing problem of SDN with NFV can be described as a linear programming model. According to this architecture, the required delay can be obtained based on resource conditions in the network, load, resource utilization, and other data to promote dynamic service delivery and optimize network resource utilization. For example, the network function allocation problem is described as a two-stage Stackelberg game, where the server acts as a network function seller and the user as a network function buyer. Wu [
115] applied the RL algorithm to obtain the optimal network function allocation strategy.
- c.
Internet of Things (IoT)
IoT has become a global network infrastructure by connecting many different heterogeneous devices using heterogeneous communication technologies (wired/wireless). Because of the extensive coverage and high mobility associated with these devices, a range of radio access technologies have found widespread application in the IoT. Nonetheless, the intricate nature of heterogeneous communication technologies and device infrastructure poses numerous significant challenges. For instance, as the number of devices increases, the traffic load on switches can become exceedingly burdensome, necessitating the appropriate allocation of multiple channels to links. Furthermore, heterogeneous devices have different strategies for data detection and collection, leading to the occurrence of uneven traffic bursts arriving at the switch [
116,
117]. To better adapt to large-scale heterogeneous IoT, ISDNs have become a novel solution for connecting distributed heterogeneous devices to centralized shared systems, termed Intelligent SDN-IoT. Within Intelligent SDN-IoT, a multitude of devices are extensively distributed across the sensing plane. The sensing data amassed by this plane is relayed and conveyed to the gateway via switches situated in the data plane. AI technology is used in combination with partially overlapping channel allocation to solve adaptive channel allocation issues and ensure the QoS of communications in wireless networks.
- d.
Edge Computing
Edge computing (EC) entails the deployment of computing and storage resources at the periphery of the network. Within edge computing, the term “edge” encompasses any computing and networking asset positioned between the data source and the cloud data center [
118]. The integration of SDN intelligence with edge computing can yield advantages across various domains, including enhanced resolution and control, increased flexibility with fewer innovation hurdles, implementation centered around services, mobility of virtual machines, adaptability, interoperability, cost-effectiveness, and extensive coverage. Nonetheless, both Smart SDN and the current standardization of OpenFlow are not yet sufficiently developed to handle all the potential use cases and management operations outlined. As requirements change rapidly, even the relatively new standards-based realm of infrastructure as a service is also constantly evolving. Faced with increasingly complex requirements, merely deploying SDN-enabled network devices will not easily integrate SDN into existing networks. Moreover, the investment cost in hardware is very high, and therefore, smart SDN-supporting hardware cannot reach the required level in terms of functionality and deployment. Using AI technology to provide finer control granularity and abstraction granularity offers more possibilities in use cases and management.
- e.
Information-Centric Network (ICN)
ICNs represent an architecture designed to furnish users with content accessibility through names as opposed to establishing communication channels between hosts [
119]. Integrating SDN functionalities into ICN amalgamates the network control plane with the data plane, thereby facilitating centralized programming and network control from a holistic standpoint. The amalgamated SDN-ICN framework, which combines ICN and SDN, proves advantageous in capitalizing on the data autonomy of the data plane within ICN and the centralized control afforded by the control plane in SDN for enhanced global network management. Initially, content devoid of address constraints in ICN is routed and forwarded as an autonomous entity at the network layer. Centralized scheduling facilitated by SDN can optimize the allocation of network resources [
120,
121].
There are numerous computing nodes in SDN-ICN. The use of AI can simultaneously learn with ICN, reducing the delay caused by ICN data processing. Therefore, due to the use of rich computing resources scattered across various SDN nodes, it can address the computation time issue during the AI model training process. Secondly, based on the unification of data collection and analysis from a global perspective, multi-scale traffic prediction can be achieved with flow-oriented characteristics for greater accuracy. Multiple switches in distributed SDN collect features of the requested target content. These features contain inherent spatiotemporal and social correlations, which DL networks can comprehend from a global view. Consequently, due to the unsupervised learning capabilities of DL networks, they can find the distribution pattern of content popularity without missing small pattern changes, thus improving prediction accuracy. Foremost, SDN-ICN holds the capability to modify the structure of the DL network. This stems from the programmability inherent in SDN and the overarching control exerted by the SDN controller over the network. Consequently, hidden layers and neurons within SDN-ICN can be readily reconfigured at each layer. Leveraging this programmable infrastructure, diverse network models can be seamlessly integrated to address computational challenges encountered within the SDN-ICN framework [
122].
- f.
Wireless Network
A wireless network (WN) consists of many nodes and transmits through wireless channels. Unlike wired networks, wireless network channels always change with user mobility, channel fading, and interference. In dense wireless networks with mobile users and small cell sizes, channel capacity changes are more challenging to handle [
123,
124]. Integrated SDN-based solutions can support user mobility in dense wireless networks, making it possible to implement software in traditional and emerging wireless environments. Software-Defined Wireless Networks (SDWNs) have raised concerns about the inherent security of the SDWN architecture [
125,
126]. The SDWN paradigm faces security challenges akin to those present in the traditional wired SDN framework, exacerbated by the introduction of the wireless medium, which introduces additional avenues for attacks. Attackers can target the control plane, data forwarding elements, and individual wireless applications. Despite the controller’s susceptibility as a single point of failure, its dynamic migration across network servers enhances resilience against potential compromises to the control plane. Furthermore, the comprehensive network visibility afforded by global monitoring empowers security administrators to monitor real-time traffic statistics and adapt security policies promptly as required.
Moreover, SDWN environments can leverage the capabilities of individual nodes to implement security at different parts of the network chain, thereby establishing a layered security model without overburdening a single network entity. Thus, if the existing centralized SDWN design is used appropriately, the framework itself can be transformed into a programmable security barrier, including functions that can be changed via programmable data plane equipment. Combining AI-related technologies with SDWNs to address SDWN security issues, integrating historical data analysis to predict malicious attacks, and accurately and effectively improving SDWN security and reliability are crucial. In SDWNs, ML algorithms play a vital role in managing numerous heterogeneous sensor nodes, optimizing each node’s resource utilization, and flexibly and efficiently scheduling communication links. Currently, routing optimization, node clustering, and data aggregation in wireless sensor networks, event detection and query processing, positioning, intrusion detection, fault detection, and other problems use ML technology [
127].
- g.
WiFi
The authors in [
128] propose integrating SDN and ML techniques to improve operational efficiency and QoS in wireless local area networks (WLANs). This offers an innovative approach to wireless network management, enabling dynamic adaptation and smarter, data-driven decision-making. The second study [
129] provides a comprehensive overview of advances in optimizing wireless networks through SDN and AI techniques, highlighting the potential benefits of this integration, including improved performance, efficiency, and security of WiFi networks. Finally, Ref. [
130] presents a specific approach using an ML-based SDN controller for managing wireless LANs. This proposed model uses ML algorithms to optimize resource allocation and enhance WiFi network performance, demonstrating AI’s potential to optimize network operations in wireless environments. Collectively, these studies emphasize the crucial role of integrating SDN and AI in the evolution of wireless networks, opening new opportunities for continuous performance and efficiency improvement [
131]. WiFi challenges with ISDN include refining AI algorithms for more dynamic optimization, ensuring data security and privacy, exploring the scalability and interoperability of proposed solutions, and experimentally validating their effectiveness in wireless network environments.
- h.
Other Application Trends for ISDNs
The convergence of diverse advanced technologies is transforming industrial IoT and cloud computing, optimizing the efficiency and reliability of productive services, and managing critical infrastructures. Four key approaches stand out in this field.
The authors of [
132] introduce a precision mechanism to allocate resources in edge computing, improving operational efficiency in smart cities. This approach allows an adaptive and dynamic response to changing demands, ensuring optimal management of urban resources. In [
133], the authors address semantic segmentation in networks (Unmanned Aerial Vehicle—UAV). Techniques such as convolutional neural networks and generative adversarial networks correct biases in training data, improving the accuracy and reliability of segmentation models. Additionally, in [
134], researchers apply deep neural evolution networks to detect faults in the interconnection of cloud data centers. This method combines deep learning and evolutionary algorithms, allowing accurate and timely identification of failures, as well as improving operational resilience. Finally, in [
135], the authors explore the use of federated learning for productive service procurement in industrial IoT. This approach, inspired by the human brain, facilitates decentralized learning, preserves data privacy, and improves the efficiency and security of industrial processes. Together, these developments represent a significant leap towards optimization and security in resources and services management for industrial and urban environments, highlighting the transformative potential of emerging technologies in IoT and cloud computing that can be supported by ISDN.
Considering this,
Table 8 shows findings that underscore key research voids within the intersection of SDN and AI, necessitating attention in forthcoming investigations. These voids delineate domains and facets that remain incompletely examined or comprehended within the scientific discourse. Addressing these lacunae promises enhancements in the methodologies and strategies deployed, as well as a deeper comprehension of AI’s relevance and efficacy within SDNs. Furthermore, bridging these gaps will facilitate the creation of more refined and sophisticated models, thereby making substantial contributions to the realm of AI as applied to SDNs.
Figure 2 shows a detailed analysis of annual publication trends, indicating a significant increase in research activity since 2018, with notable peaks in 2019 and 2021. This upward trend reflects the growing interest and rapid advances in integrating AI with SDN technologies. Key contributions in 2019, such as a highly cited review article that represents the synergy between machine learning algorithms and SDN concepts, exemplify the foundational work driving this field forward. Similarly, the 2021 perspective on smart city applications and vehicular intelligence systems underscores the innovative and practical applications of AI-enhanced SDN.
By comparison, the research in [
12] focuses on load balancing in SDN using AI, analyzing SDN architecture, and categorizing AI-based methods for load balancing. On the other hand, this research provides a broad focus on practical applications such as smart cities and vehicular systems, while [
12] offers a detailed assessment of load balancing mechanisms, highlighting challenges and future trends, with a specific focus on network efficiency and resource usage. Both studies underline the importance of AI improving SDN performance, although ours provides a broader view of practical and emerging applications.
Otherwise, this research emphasizes the integration of AI in network management and its innovative applications in various sectors. In contrast, the research on the era of generative AI in 6G wireless intelligence addresses the challenges and developments in next-generation wireless networks, highlighting the role of generative AI in improving 6G network intelligence [
59]. Also, it is important to emphasize that this research focuses on the synergy between AI and SDN, including applications in smart cities and vehicular systems. The research in 6G explores new fronts of wireless intelligence, covering the evolution of infrastructure and the development of advanced algorithms for 6G networks. Finally, the research in [
33] focuses on the problem of controller placement during SDN deployment in telecommunication internet service provider networks, analyzing strategies to optimize controller placement to improve network performance and efficiency. Contrasting with the obtained results of this research, the integration of AI to improve network management and practical applications is highlighted.
Thus, it becomes evident that the integration of AI with SDN not only enhances network management capabilities but also paves the way for innovative applications across various sectors. However, the exploration of AI-driven SDN is still in its early stages, particularly in the context of emerging technologies such as 6G wireless intelligence. As this field continues to evolve, challenges of scalability, security, and real-time optimization remain paramount. Future research should aim to bridge these gaps by developing robust AI models tailored to the dynamic and complex nature of SDN environments, unlocking the full potential of intelligent network management systems.
Finally, and from the research results, several promising directions for future investigations in the field of AI-integrated SDN can be outlined. First, the scalability and efficiency of AI algorithms in SDN environments need to be further explored, especially in large-scale contexts such as telecommunication networks and data centers. Secondly, future research could focus on developing AI models that optimize routing and load balancing in real-time, dynamically adapting to changing network traffic conditions. Furthermore, given the growing interest in 6G wireless intelligence, the integration of generative AI to improve wireless network security and management represents another crucial area of research. It is also essential to investigate the interoperability and integration of SDN with other emerging technologies, such as edge computing and virtualized function networks, in order to create more robust and efficient systems. Finally, developing high-quality data sets and creating advanced predictive models for threat detection and cyberattack mitigation in SDN networks are critical areas that require attention to ensure the security and resilience of future network infrastructures.