Federated learning for millimeter-wave spectrum in 6G networks: applications, challenges, way forward and open research issues

Rationale and targeted audience

The study is essential for several reasons. Firstly, as 6G networks promise ultra-high data rates and connectivity, effectively utilizing the millimeter-wave spectrum is crucial for realizing their full potential. However, the unique propagation characteristics and security concerns of mmWave pose significant challenges. Traditional DL/ML approaches have limitations such as centralization and data privacy issues. Therefore, exploring FL as a decentralized paradigm becomes imperative, offering a solution to collaborate on model training while preserving data privacy. This study aims to address these challenges, identify specific applications, propose solutions, and highlight open research issues to pave the way for integrating FL and mmWave in 6G networks.

The targeted audience for this study includes researchers, engineers, and practitioners in the fields of telecommunications, wireless networking, machine learning, and artificial intelligence. Specifically, professionals involved in the development and deployment of 6G networks. Moreover, this study is particularly relevant for the research community interested in leveraging advanced techniques like FL to enhance network performance and security. Additionally, policymakers and stakeholders in the telecommunications industry seeking insights into the future directions of network technology and its implications on data privacy and security would benefit from the findings and recommendations presented in this review.

Related literature and their limitations

Recently, there has been a remarkable surge of research publications/reviews/surveys on the FL enabled 6G communication networks, spanning all the layers in the emerging network architecture. Table 2 summarizes the discussion on previous research works that are related.

The authors in Li et al. (2020) provide a detailed review of FL applications in mobile device and industrial domains but do not include mmWave applications. Similarly, the study Lim et al. (2020) is a comprehensive survey on FL integration with mobile edge computing but lacks a discussion on mmWave based wireless communication with FL. The study Wahab et al. (2021) offers a comprehensive understanding of federated learning’s intricacies, followed by exploring its evolving role and prospects in communication and networking. However, the study lacks a discussion on mmWave-specific FL applications and mmWave based concepts carried through in 6G networks. The study Boulogeorgos et al. (2021) explores the applications of THz wireless systems within the context of the Beyond fifth generation (B5G) era, delving into the emerging AI technologies that facilitate their implementation. However, the study does not cover the B5G network specifically, mmWave and FL enabled 6G networks. The authors in Pham et al. (2022) discuss the utilization of Mobile Edge Computing (MEC)-enhanced in Aerial Access Networks (AAN), commonly referred to as aerial computing, to explore the integration of FL in the context of 6G networks. However, the study does not discuss mmWave unitization in 6G networks. Similarly, the authors in Al-Quraan et al. (2023) seek to comprehensively explore cutting-edge FL applications within wireless technologies. Further, it aims to provide a comprehensive overview of the advancements and their significance. However, the areas related to mmWave spectrum and related concepts are only partially discussed. The study Duan et al. (2023) is one of the latest related work. It provides an in-depth examination of the latest technological advancements integrating FL and edge computing. Further, it covers a holistic perspective encompassing the intersection of FL and edge computing within the framework of 6G communication. However, the study does not discuss related issues related to mmWave communication in FL-enabled 6G networks. Likewise, the study Driss et al. (2023) explores the enhanced benefits of integrating FL across the entire protocol stack in the 6G technology.

Additionally, it highlights pivotal FL applications and delves into key concepts to offer valuable perspectives for future research and development endeavors. It contains only a brief discussion on mmWave utilization in FL enabled 6G network. The authors in Xiao et al. (2024) provide detailed analyses of various scenarios in FL over the air including, Multi-Input-Multi-Output (MIMO) and Reconfigurable Intelligent Surface (RIS). However, mmWave communication is only partially discussed with respect to RIS integration with FL. The study Lee et al. (2024) is a comprehensive survey on FL concepts in 5G and beyond network including network management, network core, network access etc. However, mmWave communication is discussed only partially in LTE communication domain. Notably, none of the previous reviews or surveys specifically cover the concepts involved in integrated scenarios of mmWave and FL enabled 6G communication.

Research methodology

This methodology ensures a systematic approach to reviewing and analyzing literature on federated learning for millimeter-wave spectrum management in 6G networks, thereby providing comprehensive insights into its applications, challenges, and future research directions. The following steps have been adapted for an exhaustive literature review:

• Online search: A systematic approach was adopted to conduct a review of the literature. Renowned academic databases, including SCOPUS, Web of Science (WoS), IEEE Xplore, ScienceDirect, MDPI, and Google Scholar, were searched using carefully chosen keywords such as “Federated Learning,” “Millimeter-Wave Spectrum,” “6G Networks,” “FL Applications,” “Challenges in FL,” and “Open Research Issues in FL.” The search time spanned from 2019 to 2024, ensuring the inclusion of recent advancements in the field.

• Article selection: In selecting relevant publications, thoroughly scrutinize titles, abstracts, and keywords to ascertain their alignment with FL in the context of millimeter-wave spectrum and 6G networks. Subsequently, exclusion criteria were applied meticulously to eliminate non-novel, duplicate, or irrelevant publications, thus refining the selection to encompass only the most pertinent contributions.

• Analysis: The study employed bibliometric analysis tools such as VOSviewer to comprehensively analyze articles’ distribution based on titles and keywords, as shown in Fig. 1. By examining the distribution of relevant articles from various publishers, the trends and patterns in the literature pertaining to FL for millimeter-wave spectrum in 6G networks were identified. Moreover, from the VOSviewer based landscape of technological terms, the keywords can be classified as;

  • Core technologies: Millimeter wave, 5G network, 5G/6G, enabling technology, wireless federated learning, federated dropout learning,

  • Network and spectrum: Dynamic blockage recognition, coverage probability, allocation, beam management, channel estimation, beam squirt, overlap subarray,

  • Privacy and security: authentication techniques, security applications,

  • Applications and use cases: smart education, art, design,

  • Emerging trends: 6G Massive MIMO, Blockchain, AI/ML.

• Synthesis: The literature review and analysis in this study reveal several findings regarding the applications, challenges, and open research issues associated with FL in leveraging millimeter-wave spectrum for 6G networks. Additionally, the study identifies challenges, including bandwidth consumption, power consumption, and synchronization requirements, while also providing a way forward for further investigation into areas like dynamic synchronization and bandwidth optimization for FL-enabled millimeter-wave communication in 6G networks. The overall research methodology opted in this research is depicted in Fig. 2.

Organization and contributions

It is pertinent to highlight that almost all existing reviews and surveys have partially covered the topic. Moreover, the existing research lacks insight into the latest challenges and open research issues.

• FL enabled 6G networks (“FL Enabled 6G Networks”): This section initially discusses the basic concepts of FL; thereby, to establish FL potential, this section critically the FL application for 6G domains; including, enhanced Mobile Broad-Band (eMBB), Un-Conventional Data Communications (UCDC), Secure Ultra-Reliable Low-Latency Communications (SURLLC), Three-Dimensional Communications (3DCom) and Big Communications (BigCom).

• Millimeter-Wave Applications in FL Enabled 6G (“mmWave Applications in Fl Enabled 6G”): This section describes mmWave utilization for 6G networks including innovative concepts such as mmWave spectrum for integrated 6G, mmWave Chip for hybrid Massive MIMO (M-MIMO), sub-6GHz and mmWave dual-band antennas, joint venture of BF, mmWave and M-MIMO, wireless backhaul with mmWave and BF design and process.

• Challenges (“Open Research Issues”): This section highlights the primary challenges and limitations in embracing mmWave with FL; including bandwidth consumption, energy consumption and synchronization requirements.

• Proposed way forward (“Conclusion”): This section provides a way forward based of FL integration in mmWave communication architecture. The proposed approach is named Federated Energy-Aware Dynamic Synchronization with Bandwidth-Optimization (FEADSBO).

• Future research directions (“Way Forward”): Thereby, this review identifies the specific future research areas in FL enabled mmWave; including, FL for high-frequency mmWave communication, FL for energy efficiency in massive antenna systems, secure integration of mmWave in FL enabled 6G and mobility in mmWave with FL enabled 6G.

• Conclusion (“Open Research Issues”): The article concludes in this section with a brief on the aim of this review and the corresponding contribution to fill the research gap.

FL IN eMBB

eMBB, featuring advanced smartphones, dynamic gaming, and high-res multimedia, exhibits asymmetry in data rate needs, ranging from Mbps to 1+ Gbps. In the context of eMBB applications, FL supports efficient model training while minimizing the need to transmit large amounts of sensitive data over the network. FL enhances the adaptability of models to the dynamic nature of eMBB by allowing continuous learning from data generated across the network (Guo et al., 2022). The authors in Hong, Park & Choi (2023) introduce a novel over-the-air aggregation framework for FL in broadband wireless networks. Unlike conventional FL setups, this framework accommodates local datasets on edge devices and base stations, thereby enhancing the efficiency and collaboration in model training across the network.

The integration of FL in eMBB aligns with the goals of achieving robust and secure communication in 6G networks while accommodating modern applications’ diverse data rate requirements. Similarly, the study Balakrishnan et al. (2020) introduces resource management schemes that consider the importance of computing communication and data and optimizing training metrics. The proposed algorithms demonstrate a significant 4x–10x reduction in convergence time without compromising test performance, achieving a balance between model effectiveness and overall training efficiency on benchmark datasets.

FL in SURLLC

In 6G networks, FL plays a pivotal role in SURLLC, which encompasses applications in smart tools, industries, and healthcare. FL facilitates collaborative model training across distributed devices, ensuring data privacy and security in SURLLC applications. The authors in Khowaja et al. (2021) proposed a Distributed Federated Learning (DBFL) framework that addresses challenges associated with long-range communication by mitigating the need for devices to increase transmission power, thereby alleviating energy efficiency concerns. This framework is designed to overcome connectivity issues for distant devices and seamlessly integrates with a mobile edge computing architecture. DBFL facilitates distributed communication among devices by utilizing clustering protocols, offering an efficient solution for collaborative learning while minimizing energy consumption. This decentralized approach allows SURLLC devices to collectively improve model accuracy without transmitting sensitive data to a centralized server, mitigating latency concerns. In smart tools and industries, FL enables real-time decision-making with low-latency communication, enhancing the reliability of critical processes. The study Lu et al. (2020) introduces the concept of Digital Twin Wireless Networks (DTWN), integrating digital twins into wireless networks to shift real-time data processing to the edge plane. This approach involves implementing a blockchain-empowered federated learning framework within DTWN, fostering collaborative computing to enhance system reliability, security, and data privacy. This innovative combination improves the overall functionality of the wireless network. In healthcare, FL supports secure collaborative learning on patient data distributed across devices, ensuring both reliability and privacy in the delivery of critical medical services within the 6G framework.

FL in 3DCom

In the 3DCom of 6G communication, FL can potentially play a pivotal role in addressing challenges associated with airborne, multi-dimensional and high-rise platforms such as drones and underwater communication. FL enables decentralized machine learning models to be trained on these platforms, allowing them to adapt and learn from local data without transmitting sensitive information to a centralized server. This is particularly crucial for 3DCom due to its spatiotemporal characteristics, ensuring that machine learning models can adapt dynamically to the changing environmental conditions of airborne and underwater scenarios. The study Qu et al. (2007) introduces an innovative concept known as AGIFL (Air-Ground Integrated FL), which seamlessly combines air-ground integrated networks with FL. Within the AGIFL framework, the adaptable and on-demand 3D deployment of aerial nodes, such as unmanned aerial vehicles (UAVs), enables all nodes to collectively train a proficient learning model through FL (Sandamini et al., 2023). Similarly, in order to overcome UWA tough conditions, the authors in Zhao et al. (2021) introduce C-DNN, a deep neural network-based receiver designed for underwater acoustic chirp communication. They also present an innovative approach by combining federated meta-learning with acoustic radio technology to enhance the performance and generalization of the DL model. The study provides tractable expressions for the convergence rate of FL in a wireless network by considering scheduling ratio, local epoch, and data volume on individual nodes. The potential outcome is improved, efficient, and reliable data transfer for underwater communication systems. However, challenges in this approach includes, potential complexities in coordinating and managing the federated meta-learning process, and limitations in addressing real-world variations in underwater environments. FL enhances privacy and security by keeping sensitive data local, mitigating potential risks associated with data transmission over these specialized 6G communication channels. Leveraging FL in 3DCom of 6G networks facilitates efficient and adaptive machine learning applications on diverse smart platforms.

FL in UCDU

FL in the context of UCDC within 6G networks, offers an innovative approach to decentralized machine learning. UCDC, being an open-ended edge technology, benefits from the ability to train models across a distributed network of unconventional devices, such as smart human bond applications. This enables the dynamic and diverse nature of UCDC to be harnessed for collaborative learning without centralizing sensitive data. The authors in Sozinov, Vlassov & Girdzijauskas (2018) assess the effectiveness of FL in training a classifier for Human Activity Recognition (HAR), contrasting its performance with centralized learning. The findings indicate that while FL yields slightly lower accuracy than centralized models for human activity recognition, the difference is deemed acceptable. Likewise, the study Ouyang et al. (2021) introduces ClusterFL, a FL system designed for HAR applications. It utilizes a unique clustered multi-task federated learning framework to enhance model accuracy by efficiently capturing inherent clustering relationships among data from various nodes, minimizing communication overhead. The open-ended nature of UCDC aligns well with FL adaptability, facilitating secure and efficient attribute selection for authentication and confidentiality in this evolving technological landscape.

FL in BigCom

FL in the context of BigCom in 6G contributes to a comprehensive global communication paradigm by decentralizing machine learning processes. This approach enables devices to collaboratively train models without transmitting raw data to a centralized server, ensuring privacy and security. The authors in Chen, Xiao & Pang (2022) advocate for implementing FL in Low Earth Orbit (LEO)-based satellite communication networks to enhance support for highly interconnected devices with intelligent adaptive learning and mitigate costly traffic. Moreover, the study comprehensively reviews the current state-of-the-art LEO-based satellite communication and explores relevant ML techniques (Siddiqui et al., 2022). In 6G BigCom, FL accommodates the diverse aspects of communication technologies, allowing for dynamic and efficient learning across various devices and applications.

Similarly, integrating the Space-air-ground Integrated Network (SAGIN) with satellite, aerial, and terrestrial networks is enhanced through the innovative use of FL, a distributed learning method. This approach intelligently addresses resource scheduling challenges in SAGIN while ensuring security and safeguarding user privacy (Tang et al., 2022). The distributed nature of FL aligns with the need for robust communication in the face of evolving challenges, such as those presented by eMBB and other components of 6G. By fostering collaboration and learning across a globally distributed network, FL becomes an integral part of the holistic approach to communication in the BigCom paradigm of 6G.

FL in mmWave spectrum for integrated 6G networks

The rapid-paced development of new technologies and smart modes of communication, as discussed in the 6G communication domains, is poised to overwhelm and strain the capabilities of the existing 5G network infrastructure, resulting in congestion and inadequate performance (Hong et al., 2021). The upcoming 6G will be a network of intelligent mobile communication, where almost everything will communicate with each other or the environment (Giordani & Zorzi, 2020). In this context, mmWave, and sub-mmWave technologies have the full potential to raise the bar and set new standards of intelligent communication. However, unconventional applications and heterogeneous architectures require highly dependable and distributed approaches to fully harness mmWave potential for 6G communication (Qamar et al., 2018).

AI, a revolutionizing concept, is a highly researched technology for integrating various technologies in 6G intelligent communication. FL is a privacy-protected and distributed approach that can provide an intelligent integration of mmWave with 6G communication. The authors in Catak et al. (2022) suggest an adversarial attack mitigation scheme ML models for mmWave beam prediction in 6G. The method involves gradient sign techniques; however, this method requires further enhancement to avoid the trivial vulnerability of centralized DL/ML models. FL is a potential candidate to rescue the situation in combination with security mechanisms such as differential privacy (DP). It is a challenge to model the space-time non-stationarity properties of UAV based mmWave communication in 6G, including average fading time, envelope crossing rate, doppler power density and space-time frequency correlation (Bai et al., 2021). The authors in Qi, Liu & Yang (2020) utilize FL for dynamic adaptability to mobility patterns for improved hand-in mmWave communication. Moreover, the scheme is compatible with the client, who has limited storage capacity.

Similarly, the study Salehi et al. (2022) proposes a FL approach to expedite sector selection in the mmWave band for vehicular mobility by leveraging machine learning techniques that integrate data from LiDAR, GPS, and camera sensors. This includes a multimodal FL framework for collaborative model training among vehicles. Validation on a real-world dataset shows a significant 52.75% reduction in sector selection time compared to the 802.11ad standard, while maintaining 89.32% throughput with globally optimal solutions. However, this approach depends on the quality and availability of non-RF sensor data. Additionally, the effectiveness of the proposed solution is constrained by security issues inherent in sharing model weights among vehicles.

In mmWave systems, hybrid precoding architecture is considered suitable for mm-wave systems to achieve high beamforming gain with reduced hardware complexity, but channel estimation becomes challenging due to the split between analog and digital domains. The authors in Zhao et al. (2022) proposed a FL-based channel state information (CSI) estimation and feedback (FCEF) scheme, where users train local models and exchange parameters with the base station, reducing transmission overhead compared to centralized learning. However, with rapidly changing channels or highly dynamic user distributions, the effectiveness of the local model updates could be compromised, leading to degraded performance in CSI estimation and feedback. Similarly, the study in Al-Abiad, Hassan & Hossain (2023) introduces a resource-efficient FL framework tailored for millimeter-wave aerial-terrestrial integrated networks. It leverages decentralized model dissemination through UAV-to-UAV and device-to-device communication, enabling increased participation of user devices without relying on a central server. However, reliance on UAVs and complex scheduling algorithms can pose implementation challenges, particularly in real-world scenarios with dynamic network conditions and hardware constraints. The above analysis is summarized in Table 3.

FL in mmWave chip for hybrid M-MIMO

The mmWave chip integrates both digital and analog components to efficiently process signals in M-MIMO systems, which are essential for next-generation wireless communication networks. By leveraging mmWave frequencies and hybrid beamforming techniques, this chip can support high data rates and enhance the capacity and coverage of wireless networks. A modification in the antenna design is also required to prosperously conduct the mmWave RF signals, especially sub-mmWave signals, in a closely packed hybrid M-MIMO system. ML algorithms can be employed for intelligent signal processing, adapting dynamically to varying channel conditions and optimizing beamforming parameters for improved system performance. Similarly, DL models can aid in complex signal processing tasks, such as channel estimation and interference mitigation, enabling the chip to achieve higher efficiency and reliability in M-MIMO systems. However, the centralized architecture of DL/ML can pose challenges related to processing and power consumption at the chip level. Therefore, FL is a potential distributed computing solution for resource constraints in the mmWave chips. The authors in Elbir & Coleri (2020) presented a FL based hybrid beamforming in mmWave. Here, the model training is performed at base stations (BS), where input data is the channel data used for analog beamforming. This technique aims to counter bandwidth consumption during data collection in traditional centralized learning ML/DL approaches.

The study Onizawa et al. (2019) suggests a training chip by utilizing a mainstream 65-nanometer CMOS technology for ML. The chip employs invertible logic and stochastic computing to achieve training without backpropagation. It enables direct extraction of weight values from input/output training data with low precision, making it well-suited for inference tasks. CMOS is developed for edge inference through a pre-trained model. However, the study established the effective utilization of CMOS-based processing in memory (PIM) Computing for model training and learning in FL applications (Qamar et al., 2023). The authors in Vu et al. (2020) introduce an innovative approach for cell-free massive multiple-input multiple-output (CFmMIMO) networks, specifically designed to facilitate any FL framework. The proposed scheme ensures the occurrence of individual iterations, rather than all, within an extensive coherence time to ensure the stable operation of the FL process.

Similarly, the authors in Farooq et al. (2023) proposed half-duplex (HD) and full-duplex (FD) communication schemes to handle simultaneous FL and non-FL user groups sharing the same resource, enabling privacy preservation and communication efficiency in MIMO system. Meanwhile, self-interference or small FL model updates can affect the performance of full-duplex communication. Moreover, the generality and scalability of this approach require the evaluation of real-world heterogeneous scenarios.

Similarly, the study Zhong, Yang & Yuan (2022) discusses Over-the-Air Federated Multi-Task Learning (OA-FMTL) over the MIMO multiple access channel, leveraging analog superposition of electromagnetic waves for computation. It introduces a novel model aggregation technique to align local gradients from different devices, mitigating the straggler problem arising from channel heterogeneity. However, dealing with a large number of devices can potentially impact the effectiveness of the proposed approach. The above discussion is summarized in Table 4.

Sub-6 GHZ and mmWave dual-band antennas

The NR 5G mobile communication is compatible with earlier technologies, and simultaneous operation of both sub-6 GHz and mmWave spectrums on the same antenna nodes is desirable. Indeed, new shared aperture antennas that support the operation of both frequency bands are a serious challenge and have recently emerged as a hot topic (Hasan, Bashir & Chu, 2019). The major design issue is the dimension requirements, as the gap between both frequency bands is very large. Research in the sub-6 GHz spectrum for dual-band antennas typically focuses on the frequency bands, including low bandwidth (Sub-1 GHz for LTE) and mid bandwidth (1–6 GHz for WiFi). While dual-band antennas are typically associated with sub-6 GHz frequencies, mmWave frequencies (24 GHz and above) are used for high-capacity and ultra-fast communication (Alieldin et al., 2018; Sang et al., 2023).

Utilizing sub-6 GHz channels for predicting millimeter wave (mmWave) beams and blockages holds promise for enhancing mobility and reliability in scalable mmWave systems. By employing a sufficiently large neural network, achieving high success probabilities in predicting mmWave beams and blockages is feasible, approaching near certainty (Alrabeiah & Alkhateeb, 2020). Here, the decentralized nature of FL can facilitate the aggregation of insights from multiple devices, leading to improved success probabilities in predicting mmWave beams and blockages, with the potential to achieve high accuracy and lower bandwidth consumption. The authors in Al-Quraan et al. (2023) introduce Radar-aided Dynamic Blockage Recognition (RaDaR), which combines radar data and FL to train a dual-output neural network model. This model can predict both blockage status and time concurrently for proactive handover or beam switching. This approach mitigates latency in 5G new radio procedures, ensuring a high-quality of experience for users.

Similarly, the authors Chafaa et al. (2021) introduced an FL framework for a wireless network with multiple communication links (access points and users). This approach involves individual access points training their local deep neural networks using local data, sharing only model parameters to achieve a collectively improved global model for predicting downlink mmWave beamforming vectors based on uplink sub-6GHz channels, thereby; enhancing data rate predictions.

However, due to the heterogeneous and varying nature of wireless network, the data is considered highly Non-Independently Distributed (Non-IID), which can severely affect FL’s performance. The research Bai et al. (2021) proposes two personalized FL approaches to address non-IID effects. The first involves fine-tuning FL models on individual private datasets of clients, while the second employs Adaptive Expert Models for FL to directly predict the optimal mmWave beamforming vector from non-IID sub-6 GHz channel vectors generated by Deep-MIMO.

The study Hu et al. (2023) proposes combining multi-band- reconfigurable holographic surfaces (RHSs) and FL to offer precise and environment-adaptive user positioning services. It employs an FL framework for collaborative training of a position estimator, leveraging transfer learning to address the absence of position labels among users, and introduces a scheduling algorithm for base stations to select users for training, considering both FL convergence and efficiency. However, implementing such a system across diverse network environments can be challenging. Similarly, the study Chafaa et al. (2022) presents a self-supervised DL method for predicting beamforming vectors in mmWave communication, utilizing sub-6 GHz channels and DeepMIMO dataset. The extension to multiple links using FL efficiently predicts mmWave beams with limited local data. However, potential limitations may arise from network synchronization challenges and the need for extensive computational resources for training and coordination. Here, the potential challenge is the synchronization of multiple access points in FL setups. The above analysis is summarized in Table 5.

Joint venture of BF, mmWAVE, and M-MIMO

Since the conventional sub-6 GHz spectrum is fully occupied, the researcher’s community is performing extensive investigations on successfully utilizing higher frequency spectrum in the mmWave and mmWave operable frequency spots. Herein, both high-frequency radio wave signals are regarded as commendatory technologies for forthcoming mobile communication (Song, Yang & Sun, 2017). However, it is undesirable for the cellular data transmission environment due to its limited coverage support in a given geographical area (Goudarzi et al., 2020). Therefore, some advanced multi-beam or beam-steerable antenna schemes have been lately adopted to conduct upper spectrum transmissions with M-MIMO antenna in B5G cellular networks. The remote channel state information (CSI) inference involves DL structures in a communication system, to accurately assess channel conditions. DL/ML approach enables a more sophisticated and adaptable method for inferring CSI in remote communication scenarios where conventional models may fall short. The study in Jiang et al. (2019) aims to assess the effectiveness of a DL-based approaches in the context of mmWave multicellular networks. The study focuses on optimizing beamforming configurations through two neural networks trained to minimize mean square error (MSE). The first network considers requested spectral efficiency (SE) per active sector as input, while the second network addresses the corresponding energy efficiency (EE), enabling adaptive beamforming to account for channel and power variations. Nonetheless, the proposed approach achieve satisfactory performance in enhanced data rates at the cost of increased radiating nodes and blocking probability.

Similarly, utilizing FL has emerged as a promising approach to enhance the performance of learning-based mmWave BF systems for efficient link configuration. In the study Elbir, Coleri & Mishra (2021), two distinct frameworks, namely model-based and model-free, are introduced for the design of beamformers in multi-user Spatial Path Index Modulation (SPIM) for joint mmWave-MIMO systems. The study employs FL coupled with dropout learning techniques. This approach involves training a learning model on the local datasets of individual users. These users estimate beamformers by inputting their channel data into the model, thereby; facilitating collaborative learning across the network.

Despite the potential benefits of FL, FL systems are susceptible to backdoor attacks during training in FL-based beam selection for mmWave systems. A backdoor attacker aims to inaccurate the out of ML model a predefined set of inputs. This manipulation is designed to undermine the model’s performance on targeted sub-tasks while appearing normal for other inputs. The authors in Zhang et al. (2022) presented a backdoor attack strategy where the model activates upon encountering specific obstacles in designated locations. When the model processes input featuring these obstacles, the backdoor is triggered, causing the model to produce the predefined output beam specified by the attacker.

Similarly, the study Huang et al. (2022) proposed an approach that integrates joint device scheduling and receive beamforming to minimize the FL convergence gap over shared wireless MIMO networks, maximizing the number of weighted devices under latency and power constraints. However, the complexity of jointly optimizing device scheduling and receive beamforming requires significant computational resources.

Likewise, the study Asaad et al. (2024) investigates Over-the-Air FL (OTA-FL) in massive MIMO systems with limited RF-chains, addressing the challenge of joint antenna selection and beamforming for model aggregation. This scheme is based on a two-tier approach utilizing penalty dual decomposition and treating antenna selection as a sparse recovery problem using the Lasso algorithm. Here, scalability can become complex with the increase in the number of antennas and RF-chains. The above discussion is summarized in Table 6.

Wireless backhaul with mmWAVE

Currently, conventional backhaul wireless links are largely based on microwave frequency bands (often owned by operators) and fiber/copper cables (often leased) with varying proportions per operator and country area (Allen, Chevalier & Bora, 2014). Many other types of research and simulation tests have also been developed to evaluate small-cell wireless backhaul communication, including software-defined networking (SDN) based techniques (Niephaus et al., 2015; Zhang et al., 2015). In contrast, some recent research works using AI and ML on the agility of dynamic configuration of backhaul streams have also been discussed. Many mmWave-based SDN wireless backhaul access systems focus on optimizing different parameters, which are exhaustively discussed in the literature (Santos et al., 2019; Santos, 2020; Camps-Mur et al., 2019).

With the emergence of heterogeneously connected devices and advanced applications like IoT, V2V communications, and wearables, industry experts foresee complex management requirements in backhaul communication. FL resolves these privacy and bottleneck issues by training a global model and also plays a crucial role in optimizing wireless backhaul communication networks (Chehri et al., 2023). The authors in Yang, Hong & Park (2021) introduce an adaptive power allocation method, where each client dynamically allocates its transmit power based on the magnitude of the FL based gradient information, optimizing communication efficiency. The article Wang et al. (2021) explores the optimization of energy and time consumption in mobile-edge computing-enabled balloon networks, where high-altitude balloons serve as flying wireless base stations. The proposed solution employs a support vector machine (SVM)-based FL algorithm to proactively determine user associations, allowing high-altitude balloons to dynamically adjust resource allocation for computational tasks without transmitting historical associations or tasks.

The study Feng & Mao (2019) proposes a solution to the challenge of backhaul resource allocation in mmWave systems using deep reinforcement learning (DRL). By leveraging DRL, the system learns the blockage patterns and dynamics, enabling efficient allocation of backhaul resources to users with highly varying data rates. DRL models in real-world mmWave face potential difficulty in generalizing learned policies across different network scenarios; here, personalized FL approaches can be used to overcome the challenges related to generalization.

The study Mahmood et al. (2024) proposes (0.1 to 10 THz) Terahertz-based networks to improve FL convergence time and reduce training loss, particularly in 1 km links, by exploiting efficient data transmission conditions in comparison. This approach involves implementing FL with THz-based wireless backhaul and a virtual private network (VPN) infrastructure, catering to end-users for enhanced privacy and network efficiency. However, signal attenuation and hardware complexities can hinder the adoption of FL deployment in real-world scenarios. Similarly, the authors in Khan et al. (2023) propose a latency-aware vision-aided FL approach for predicting beam blockage in 6G wireless networks. It utilizes multi-sensor data and advanced deep learning techniques by employing distributed learning on edge nodes for data processing and model training. This framework aims to mitigate communication costs, bandwidth inefficiency, and latency issues associated with centralized training; however, this approach can increase the complexity of coordination and synchronization across numerous edge nodes. The above analysis is summarized in Table 7.

BF design and process

In multi-stream multi-user mmWave transmission, the BS with large-scale antenna elements simultaneously serves many users with the directed beam patterns (Sun, Qi & Li, 2019; Kazmi et al., 2022). A hybrid precoding mechanism is commonly adopted for mmWave M-MIMO communication, wherein analog precoding/BF generates directional beams under constant amplitude, limited phase shifter resolution constraints, and digital precoding multiplexed independent user data streams. Many M-MIMO hybrid precoding solutions have been proposed for mmWave wireless communication (Alkhateeb et al., 2014; Xiao et al., 2016; Xiao, Xia & Xia, 2017). One of the challenging aspects of BF is Beam squinting. This is a phenomenon associated with non-uniformities in the array elements, causing the beam to deviate from its intended direction. In 6G communication systems, beam training is crucial for optimizing the directional transmission and reception of signals between base stations and user devices. This involves dynamically adjusting the beamforming parameters, such as beam direction and width, to adapt to changing environmental conditions and mitigate beam squinting effects through adaptive algorithms and sophisticated signal processing (Chen, Chen & Jiang, 2021). Designing BF for large-scale antenna arrays, particularly with constraints such as limited radio frequency chains and the use of phase-shifter-based analog BF architecture, poses a significant challenge in millimeter-wave communication systems. Therefore, the complexity further intensifies in the presence of imperfect CSI. The study Niephaus et al. (2015) introduces a novel approach using DL for BF design to address this. The proposed method involves the development of an NN that can be trained to optimize the beamformer, considering hardware limitations and dealing with the challenges posed by imperfect CSI, ultimately maximizing the spectral efficiency. The effectiveness of FL aggregation is directly linked to the increased participation of devices. The article Zhang et al. (2015) presents a design based on BF vectors and the selection of devices for FL using over-the-air computation (AirComp). The results show that the approach maximizes the count of selected devices while adhering to a predefined target aggregation MSE.

Similarly, the authors in Xiao et al. (2007) introduce a fully decentralized FL framework featuring an inexact stochastic parallel random walk alternating direction method of multipliers (ISPW-ADMM). The findings indicate that this framework exhibits significant resilience against the effects of time-varying dynamic networks and stochastic data collection while maintaining rapid convergence. Leveraging the advantages of stochastic gradients and biased first-order moment estimation, the proposed framework is a dynamic approach for decentralized FL tasks across time-varying graphs.

Similarly, the authors in Elbir & Coleri (2021) introduce a FL framework for channel estimation in both conventional and RIS-assisted massive MIMO systems. They utilize a CNN trained on local datasets of users, avoiding the need to transmit them to the BS. Their approach demonstrates significantly reduced overhead compared to conventional learning, approximately 16 times lower, while maintaining performance close to CL. However, coordination in the FL process across numerous distributed users can impact the overall efficiency and scalability of the system.

Likewise, the study Jeon et al. (2020) introduces a compressive sensing approach for FL in MIMO communication systems. It addresses the challenge of accurately reconstructing local gradient vectors sent from wireless devices by establishing a transmission strategy for constructing sparse transmitted signals and proposing a compressive sensing algorithm for the central server to iteratively find the linear minimum-mean-square-error (LMMSE) estimate of the transmitted signal. This approach relies on sparsity assumptions and linear minimum-mean-square-error estimation, which can be challenging in heterogeneous 6G networks. The above discussion is summarized in Table 8.

Bandwidth consumption by FL

The integration of mmWave technology into FL-enabled 6G communication is linked with substantial bandwidth consumption due to the continuous exchange of training details among edge devices. The inherent high data rates of mmWave communication cause bandwidth limitations, specifically during the phases of FL processes and updating (Al-Quraan et al., 2023). Therefore, the utilization of mmWave frequencies is dependent on the efficient data transfer to accommodate model updates exchanged among federated nodes. This increased demand poses a severe challenge to the seamless integration of mmWave into FL enabled 6G communication. Similarly, to address this, innovative solutions must be devised to optimize data transfer protocols (Hafi et al., 2023). Techniques such as advanced compression algorithms, adaptive modulation schemes, and prioritized data transmission strategies can potentially effectively manage the critical data volumes exchanged during FL operations over mmWave frequencies (Li et al., 2023). Additionally, integrating intelligent bandwidth management mechanisms and incorporating dynamic allocation and spectrum-sharing strategies are essential to mitigate potential congestion issues arising from the augmented bandwidth demand. Therefore, bandwidth limitation is a multifaceted challenge that requires advanced compression techniques with dynamic bandwidth management to achieve efficient and congestion-free FL processes in the context of mmWave-enabled 6G communication.

Energy consumption by FL

Integrating mmWave technology into 6G communication, particularly in the context of FL, presents a significant challenge due to heightened energy consumption during training and data processing at the edge nodes. The resource-intensive nature of FL processing is compounded by the energy-demanding characteristics inherent in mmWave transmission, involving; DL based training at local devices and further, iterative model updates across distributed nodes (Nguyen et al., 2021). In mmWave communication, the shorter wavelengths and higher frequencies lead to increased path loss and susceptibility to atmospheric absorption, necessitating higher transmit power to maintain communication reliability. This inherent limitation of mmWave contributes to substantial energy requirements in the overall system (Lim et al., 2020). To address this challenge, the development of energy-efficient algorithms and protocols are required alongside standard FL processing over mmWave frequencies. One approach involves optimizing the model update process to minimize redundant transmissions, employing techniques such as sparsity-aware algorithms to reduce the amount of data exchanged (Dinh et al., 2020). Additionally, incorporating adaptive power control mechanisms can help optimize the transmit power levels, dynamically adjusting them based on channel conditions to balance reliability and energy efficiency. Furthermore, the design of compression algorithms tailored for mmWave FL can significantly alleviate energy consumption. Thus, addressing the challenge of increased energy consumption in 6G communication with mmWave-enabled FL demands a holistic approach, from optimizing model updates to implementing adaptive power control to ensure sustainable and optimized energy usage in this complex and dynamic communication paradigm (Liu & Simeone, 2020).

Sychnozation requirements in FL

The joint integration of mmWave technology and FL in 6G communication introduces the challenge of maintaining synchronization in communication and FL processing (Rodríguez-Fernández, 2021). Due to the significant susceptibility of mmWave signals to propagation conditions, special attention must be given to maintaining synchronization among federated nodes. This synchronization is paramount for the seamless and coherent aggregation of models distributed across various nodes in the FL processing. Similarly, ensuring a reliable synchronization in Heterogeneous FL enabled 6G networks is critical for achieving coherent collaboration among distributed nodes (Qi et al., 2023). Similarly, including satellite links in the BigCom domain of 6G networks introduces additional latency considerations, necessitating sophisticated synchronization protocols to account for varying transmission delays (Giordani & Zorzi, 2020). Therefore, the solution in this domain must exhibit high resilience and adaptability to ensure precise and timely coordination among distributed nodes throughout the FL process. Achieving synchronization is challenging in scenarios where propagation delays and signal distortions have high recurrent probabilities (Chukhno et al., 2023). The intricacies of mmWave communication, characterized by short wavelengths and susceptibility to blockages, further complicate the synchronization challenge. Consequently, the successful resolution of synchronization challenges assumes pivotal importance in unlocking the full potential of mmWave in FL enabled 6G communication.

Security and privacy in FL

The integration of federated learning (FL) with millimeter-Wave (mmWave) technology in cellular networks, spanning 4G, 5G, and emerging 6G systems, presents significant security and privacy challenges (Liu et al., 2020). FL, which enables decentralized model training across devices without sharing raw data, is attractive for maintaining user privacy. The study in Yang et al. (2022) presents an estimation algorithm to prevent privacy leakage in Cybertwin-Driven 6G system. This infers local data distribution from clients without accessing their raw data. This approach addresses two scenarios in FL: one where the server receives individual trained models from each device, and another where it receives an aggregated model, with device selection strategies formulated to optimize training performance in both cases. However, the decentralized nature of FL introduces vulnerabilities, such as adversarial attacks where malicious participants can corrupt the global model or infer sensitive information through model updates. The study in Bárcena et al. (2023) introduces a novel FL-as-a-Service framework designed for B5G/6G networks, particularly in a vehicular networking scenario by leveraging FL with eXplainable AI (XAI) models. This framework enhances both the accuracy of QoE predictions in local learning and the trustworthiness. Similarly, FL with mmWave technology, with its high-frequency bands, enhances network performance but also poses unique security risks, including increased susceptibility to eavesdropping due to the narrow beam and short-range characteristics of mmWave signals (Catak, Catak & Moldsvor, 2021). Additionally, the reliance on dense infrastructure for mmWave deployment raises concerns about physical security and the potential for infrastructure attacks. As cellular networks evolve from 4G to 6G, ensuring robust security and privacy in the integration of FL with mmWave requires advanced cryptographic techniques, secure aggregation methods, and enhanced physical layer security to mitigate these risks effectively.

Dynamic synchronization

In FEADSBO, participating devices dynamically regulate their synchronization frequency based on energy availability and computational resources, ensuring that energy-constrained devices synchronize less frequently to conserve power. Devices dynamically adjust the frequency of synchronization based on their energy availability and computational resources. Devices with sufficient energy and computational resources participate in more frequent synchronization rounds. The system can optimize bandwidth usage and conserve energy more effectively by allowing participating devices to adjust their synchronization frequency according to their energy constraints and computational capabilities. This dynamic adjustment mechanism helps strike a balance between synchronization requirements and resource constraints, making mmWave-enabled FL more sustainable and efficient.

Energy-aware bandwidth optimization

FEADSBO utilizes quantization and pruning-based model compression techniques to optimize bandwidth consumption during model updates. Prioritizes transmission of compressed model updates and aggregates them at the server side, reducing the overall data transmitted over the network. Compressed model updates reduce bandwidth consumption, allowing for efficient utilization of network resources. Quantization involves reducing the precision of numerical values in the model, typically from floating-point to fixed-point representation. This reduces the memory and computational requirements during model inference, making it more suitable for resource-constrained devices in federated settings. Likewise, pruning involves removing unnecessary parameters or connections from the model, effectively reducing its complexity and size. Pruning techniques can be applied globally or locally across layers, identifying and eliminating redundant information while preserving accuracy. Together, quantization and pruning-based compression techniques enable FL to achieve better scalability, faster inference times, and reduced communication overhead.

Adaptive synchronization thresholds

FEADSBO dynamically sets synchronization thresholds based on network conditions and model convergence speed. Network conditions encompass several properties that can influence the synchronization thresholds in a dynamic system. These properties include network latency, bandwidth availability, packet loss rate, and network topology. Latency refers to the time delay between data transmission and reception and can vary based on the distance between devices, network congestion, and routing efficiency. Bandwidth availability determines the amount of data that can be transferred within a given timeframe, affecting the speed and efficiency of synchronization processes. Packet loss rate measures the percentage of data packets lost during transmission, impacting data integrity and reliability. Network topology defines the structure and connectivity of devices in the network, influencing communication paths and potential bottlenecks. By considering these properties, dynamic systems can adaptively adjust synchronization thresholds to optimize data exchange, minimize latency, mitigate packet loss, and enhance overall performance based on prevailing network conditions and model convergence speed in FL environments. Devices with faster convergence or stable network conditions can trigger synchronization less frequently, while devices experiencing slower convergence or fluctuating network conditions synchronize more frequently to ensure timely updates. Dynamic adjustment of synchronization frequency ensures timely updates while adapting to varying network conditions and device capabilities.

By integrating these above-discussed three aspects into the FEADSBO, FL systems can achieve improved efficiency in terms of bandwidth consumption, energy consumption, and synchronization requirements for mmWave spectrum utilization in 6G environment. For the FEADSBO, several deep learning-based features can be extracted from different system components. The potential features that can be extracted from each component, include: Received Signal Strength (RSS) Channel State Information (CSI) Carrier Frequency Offset (CFO) Amplitude & Phase Offset Signal-to-Noise Ratio (SNR) Antenna Array Configuration ADC and DAC Performance Metrics Energy Consumption Metrics. The proposed FEADSBO utilizes decentralized FL (DFL) and centralized FL (CFL) to achieve fast convergence by exploiting node-to-node communication. The feature extraction is performed at each node and the local model is trained on these features. This established a local level of intelligent synchronization. The local model are shared from node to node for DFL and simultaneously local model are shared with a centralized server for the global model in generation. The architecture of the proposed way-forward is depicted in Fig. 3.

FL for high-frequency mmWave communication

The extremely high-frequency radio waves are utterly useful in IAB propagation links and thus can contribute to the backhaul packet data transmission in either standalone (SA) or no-standalone (NSA) modes. Nonetheless, with the simultaneous operation of the mm Wave and mmWave spectrums and the M-MIMO operation, the beam width has an inverse relationship with the number of directions to be scanned at the network setup time (Qamar et al., 2019). Narrower directional beams with analog BF and sequential scans can verily cost initial access delay. Also, a very small size wavelength can easily create hardware composition problems and thus demand advancement in the hardware equipment to make it ultra-reliable with massive antenna nodes for both access and backhaul links (You et al., 2020). The mmWave communication is a promising technology for future B5G/6G mobile communication services. The massive leap in wireless devices, 3D applications, immersive multimedia activities, and various Internet-based services will insist on the different antenna configurations and access to spectrum above 300 GHz. The collaborative work of mmWave with ultra M-MIMO antenna elements would enable the management of network resources, different controlling functional parameters, and network operations affluently (Borges et al., 2021). Yet, the close packet ultra M-MIMO deployment would complicate signal processing and estimation; therefore, an FL-based assessment with minimum complexity and computing time is needed to support 6G applications.

FL for energy efficiency in massive antenna system

Severe path loss is an inevitable impediment to mmWave communication. With M-MIMO deployment, we can satisfactorily manage the problem at the high cost of EE and hardware complications. To surmount the poor energy management, a new array design has been developed wherein the mutual coupling between closely placed antennas is exploited to form super directive pairs (Borges et al., 2021). Thereby, the effectiveness of the proposed scheme is shown in three ways, minimize the number of antennas at the BS, increase EE, and ensure an achievable throughput rate. Principally, signaling cycles are adopted to construct the carrier leaner under low user mobility or traffic load in a cell(s). The NR 5G enables a longer sleep duration of up to 160 ms. Therefore, optimizing different parameters through FL integration mmWave concepts such as system information block 1, identification of cells, signaling pertinent to CSI, selection or reselection of signaling and process, and paging, are open research issues (López-Pérez et al., 2021).

Secure integration of mmWave in FL enabled 6G networks

While mmWave communication offers the potential for high data rates and low latency, its susceptibility to atmospheric absorption and limited range raises concerns about the reliability and security of communication links (Zhu et al., 2017). In the context of FL-enabled 6G networks, where decentralized learning models are collaboratively trained across distributed devices, ensuring the confidentiality and integrity of model updates becomes critical (Catak et al., 2022). Optimum utilization of mmWave communication within FL-enabled 6G networks addresses the challenges related to secure key exchange, robust authentication mechanisms, and privacy-preserving protocols while safeguarding against potential security threats and vulnerabilities (Kazmi et al., 2023a). Research efforts are needed to develop efficient cryptographic solutions and communication protocols that accommodate the unique characteristics of mmWave channels, providing a foundation for the reliable and secure integration of mmWave technology in future 6G networks employing FL.

Mobility in mmWave with FL enabled 6G networks

6G communication leverages mmWave bands for enhanced data rates and capacity. However, the challenges associated with mobility management have become more pronounced due to the unique propagation characteristics of mmWave signals (Qamar et al., 2017). Addressing seamless handovers, efficient beamforming, and context-aware mobility policies in mmWave environments is essential. Furthermore, incorporating FL techniques introduces complexities in distributed learning across heterogeneous devices while ensuring privacy and security (Fernandes et al., 2021). Balancing these aspects to design a robust and adaptive mobility management framework that optimally exploits mmWave frequencies while leveraging FL for intelligent decision-making, remains an intriguing challenge for researchers in the pursuit of realizing the full potential of 6G networks (Shome, Waqar & Khan, 2022). Thus, integrating mobility management in mmWave frequencies, coupled with FL mechanisms in 6G, represents a critical open research issue.