Artificial Intelligence in Networking Research in the Arab World

Intelligent 5G and 6G networks.  In the context of 6G, AI and ML technologies are leveraged to manage network resources more intelligently by automating processes such as dynamic spectrum allocation, traffic-load balancing, and predictive maintenance.⁷ These capabilities are essential for maintaining the quality of service (QoS) in environments where data demand fluctuates rapidly, such as urban centers and large-scale events. Researchers in the Arab world have made several contributions to this area, particularly in the development of ML algorithms that predict network traffic patterns and adapt network configurations in real time. For instance, researchers in Zhang et al.¹⁷ proposed dual attention-based federated learning (FL) models to perform accurate traffic prediction in heterogeneous terrestrial networks, ensuring optimal data routing and enhancing the overall efficiency of the network. Another notable contribution in Alwarafy et al.⁶ is the development of the deep reinforcement learning (DRL)-based radio access technology (RAT) called DeepRAT, which is a hierarchical, multi-agent, DRL-based solution for jointly optimizing RAT’s assignment and dynamic resource allocation in heterogeneous networks. This work was done in collaboration with Ooredoo, a local Qatari telecommunication company, where a dataset was collected to understand network behavior. The DeepRAT framework has demonstrated superior performance compared to existing heuristic approaches, maximizing downlink sum-rate and adapting effectively to dynamic network changes, such as edge-device mobility.

Intelligent integration of non-terrestrial networks.  To extend the connectivity to remote and underserved regions, the use of non-terrestrial networks (NTNs), such as satellites and unmanned aerial vehicles (UAV)s, is becoming increasingly important. The new generation of low earth orbit (LEO) satellites plays a critical role in providing high-capacity, low-latency communications that complement terrestrial networks. Researchers have explored methods for optimizing traffic offloading between integrated satellite and terrestrial networks, aiming to relieve congestion on terrestrial backhauls. The proposed approaches involve offloading ultra-reliable low-latency communication traffic to terrestrial networks while routing enhanced mobile broadband traffic through satellites to meet diverse latency and data-rate requirements.¹

Further, advancements in UAV-based communication have led to innovative solutions for network coverage in urban environments. The “Learn-As-You-Fly” algorithm in El Hammouti et al.¹⁰ dynamically adjusts UAV positions and associations with ground users to maximize network performance. This algorithm uses reinforcement learning (RL) techniques to adapt to changing traffic conditions, providing a flexible and efficient means of managing data traffic.¹⁰ In Hevesli,¹³ researchers proposed a dynamic digital twin edge air-ground network (D2TEAGN) for 6G-enabled industrial Internet of Things (IoT) that integrates UAVs, edge computing, and digital-twin technology to optimize resource allocation and task offloading. The problem was mapped into a Markov Decision Process and then solved using a customized DDPG algorithm. The D2TEAGN framework reduces energy consumption by up to 27% compared to baseline methods by jointly optimizing UAV trajectory, device association, and edge resources while respecting delay and energy constraints.

Intelligent IoT traffic management.  IoT networks face unique traffic patterns characterized by frequent, variable, and often unpredictable data transmissions. To address these challenges, researchers have developed intelligent traffic-management systems that use AI/ML to enhance the QoS. For instance, efforts in the region have led to the creation of a smart IoT QoS system, known as software defined networking-based application-aware dynamic IoT QoS (SADIQ), which dynamically adjusts network parameters based on the specific requirements of IoT applications.² SADIQ has significantly improved error rates in applications such as weather monitoring and smart parking, with reductions in temperature reporting errors and incorrect parking-status detection by up to 45x and 30x, respectively. By observing the spatial and temporal correlations in IoT traffic, researchers have developed a traffic prediction at the edge using the federated meta-learning approach. Such intelligent data handling ensures that critical information is prioritized while reducing network congestion and minimizing data loss and thus, increasing network reliability and performance.²

Intelligent and secure decentralized IoT architectures.  The traditional centralized approach to IoT network management presents challenges in terms of scalability, latency, and security, especially as the number of connected devices continues to grow. To overcome these limitations, a shift toward decentralized and automated IoT architectures is essential. Researchers in the Arab world have proposed edge-centric IoT models based on blockchain technology, which provide a more scalable and secure framework for managing IoT data. An example of this approach is the development of a blockchain-based system called Synopsis, which uses Byzantine fault-tolerance consensus protocols to facilitate secure and efficient data exchange in IoT networks.⁴ This system has achieved substantial improvements, including a reduction in memory footprint from megabytes to a few kilobytes and a decrease in message complexity and commitment delay by 85% and 99.4%, respectively.

Security remains a significant concern for IoT networks, given the ever-increasing number of cyber threats and vulnerabilities. To address these risks, researchers have turned to AI/ML techniques capable of real-time detection and mitigation of potential threats, thereby enhancing the overall resilience of IoT systems. In this context, FL emerges as a powerful approach that enables secure, scalable, and resource-efficient AI-driven IoT solutions. These benefits are realized through techniques, which collectively optimize the learning performance as well as resource consumption considering both data and resource heterogeneity. For example, to handle statistical heterogeneity and varying resource constraints, client selection and hierarchical FL frameworks have been proposed, reducing energy consumption while sustaining robust model performance.¹² Specifically, clustered FL (CFL)¹² further enhances convergence and scalability by grouping clients with similar data distributions, outperforming traditional FL solutions in complex IoT settings. Real-world IoT deployments often entail large volumes of unlabeled data, limited resources, and latency constraints, prompting additional refinements to FL-based methods. In this context, combining FL and CFL with semi-supervised learning (SSL) has effectively leveraged unlabeled data at the network edge, boosting global model accuracy while minimizing communication and computation overheads. Extending this synergy, integrating CFL and SSL¹¹ has demonstrated improvements in model accuracy and reduced energy usage by employing specialized models and judicious client-scheduling strategies. These multifaceted FL-based strategies, including CFL and SSL integrations, have collectively improved data handling, resource allocation, and scalability in decentralized IoT environments. Consequently, IoT deployments can become more intelligent and resilient, as these methods enable robust collaboration among diverse devices while preserving privacy and optimizing resource usage.

Despite FL’s intrinsic privacy advantages, the technology remains susceptible to adversarial threats, such as poisoning and evasion attacks that compromise global model reliability.³ To strengthen security, the authors implement DRL-based solutions that facilitate reputation mechanisms capable of detecting and mitigating malicious participants. The DRL agent continuously assesses the behavior of IoT devices during the FL process, assigning reputation scores based on their contributions and reliability.³ Devices exhibiting anomalous behavior receive lower scores, reducing their influence on the global model and thereby enhancing the overall robustness of the FL system. Beyond DRL-based solutions, the framework incorporates divergence-based detection methods to identify adversarial activities. These methods effectively safeguard IoT applications against potential threats by analyzing the differences between benign and malicious model updates. This dual-layered approach ensures that both known and emerging adversarial behaviors are promptly detected and addressed, maintaining the integrity of the FL process.

Intelligent and secure smart grids.  Smart grids represent a critical component of next-generation IoT ecosystems, bridging energy distribution, metering, and consumer demand through sophisticated communication infrastructures. However, these infrastructures remain highly susceptible to cybersecurity vulnerabilities and adversarial ML attacks. Researchers have conducted extensive work to protect smart-grid communication protocols against these evolving threats, focusing on collaborative frameworks, anomaly detection, and resilient system design. A key solution proposed is FedPot, an FL–based framework for improving the security of advanced metering infrastructure. FedPot enables small-scale power suppliers (SPSs) to deploy honeypots and securely share data with traditional power retailers, ensuring both data privacy and enhanced detection of malicious activities. This framework incorporates a quality-aware FL mechanism that integrates a data-quality measure into an enhanced FedAvg algorithm. FedPot also introduces a fair and transparent incentive mechanism to reward SPSs based on the quality and usefulness of their shared data, thereby promoting active collaboration. Simulation results validate FedPot’s capacity to strengthen security frameworks for smart grids while maintaining fairness in reward distribution.⁵ Researchers have also developed a dual-layered defense strategy against cyber and adversarial ML attacks within the IEC 60870-5-104 protocol. This approach combines conventional cybersecurity measures with adversarial training, reinforcing known defenses to address protocol vulnerabilities and using adversarial examples to train intrusion detection systems for enhanced robustness. Such a layered strategy mitigates risks from both traditional cyber threats and sophisticated adversarial behaviors, ensuring reliable grid operations.¹⁵ Lastly, electricity theft remains a major concern in smart-grid systems, where adversaries manipulate meter readings to inflate energy injection. To counter this, a fine-tuned recurrent neural network model conducts time-series analyses of meter data and incorporates additional features such as daily and monthly consumption patterns. Experimental results demonstrate robust detection of theft scenarios, even under minimal data perturbations. These cutting-edge security mechanisms, ranging from collaborative federated approaches to sophisticated anomaly detection, collectively form a comprehensive defensive framework for safeguarding smart grid networks.⁹