The Internet of Things (IoT) is increasingly evolving into a more intelligent and environmentally sustainable ecosystem. However, as IoT systems proliferate, they face persistent challenges, including limited spectrum and energy resources. These challenges can be mitigated by integrating cognitive radio (CR) technology and energy harvesting (EH) techniques. CR enables wireless devices to increase their spectral resources by opportunistically utilizing licensed spectrum bands, while EH allows energy-constrained devices to become self-sustaining and extend their operational lifetime. The increasing diversity of wireless devices and data, combined with rapidly changing wireless environments and unpredictable energy availability, presents significant challenges. As a result, traditional optimization-based approaches often fall short. The advent of artificial intelligence (AI) offers a promising alternative, enabling low-power wireless (LPW) networks to become self-managing, self-sustaining, and adaptive to changing conditions through data-driven methods. However, existing surveys rarely address cognitive LPW networks in depth and often overlook the role of AI in enabling EH and smart operations. This survey fills that gap by reviewing state-of-the-art AI approaches that address key challenges in EH-enabled cognitive LPW systems, including power allocation, resource optimization, and task scheduling. We review key advancements in cognitive LPW systems, such as IoT, backscattering networks, multi-hop and relay-assisted networks, and multiple access networks. Finally, we highlight key insights from the literature and outline promising directions for future research on cognitive LPW systems that enable smart, green, scalable, and secure deployments.

Spectrum sensing in large-scale cognitive radio networks (CRNs) presents significant challenges, as it typically necessitates numerous static secondary users (SUs) to determine the spectrum state. Current cooperative spectrum sensing (CSS) methods require SUs to transmit their private sensing data to a central unit. This centralized approach not only raises security concerns but also leads to considerable communication overhead. To address these issues, this paper introduces FeRAP, a novel CSS framework based on unsupervised federated representation learning. We leverage the mobility of multiple SUs to collect spectrum sensing data, allowing them to collaboratively yet distributively train a learning model to determine the spectrum state. The FeRAP framework employs a novel deep federated$\beta$variational autoencoder ($\beta$-VAE) for distributed representation learning, which identifies independent latent variables and learns disentangled representations of the sensing data in a lowerdimensional space. Furthermore, Affinity Propagation (AP) is then trained locally on the learned representations at each cooperating SU to securely and autonomously infer the spectrum state. FeRAP is a fully data-driven solution, requiring no modelbased assumptions or prior knowledge of channel or signal characteristics for training. Numerical results demonstrate that FeRAP's CSS performance is on par with supervised deep learning-based CSS techniques. Extensive simulations conducted under various network settings and propagation environments confirm the effectiveness and scalability of FeRAP.

In interweave cognitive radio (CR), users can transmit data only when licensed bands are vacant. Deep learning (DL) allows CRs to intelligently sense and identify channel availability. However, most supervised DL detectors require a substantial amount of labeled training data, which can be challenging to obtain in interweave CR. In this paper, we introduce G-VAP, an unsupervised deep generative approach for cooperative spectrum sensing (CSS). G-VAP utilizes an advanced variant of variational autoencoders (VAEs) called β-VAE to identify independent latent variables and encourage accurate and disentangled sensing data representation in a lower dimensional latent space. Furthermore, G-VAP utilizes the affinity propagation (AP) algorithm for unsupervised clustering to detect primary user activity. Unlike other unsupervised clustering methods, AP’s performance is not reliant on the initialization of cluster centroids. Furthermore, G-VAP leverages the cooperation among CRs to sustain a high detection performance. Our approach is fully data-driven, operates without any model-driven assumptions, and does not require prior knowledge of channel or signal characteristics for training. Numerical results indicate that G-VAP for CSS performs comparably to benchmark supervised DL-based CSS. Extensive simulations have been conducted in diverse network settings, propagation environments, and fading conditions, which have proved the effectiveness of G-VAP.

The increasing demand for data transmissions in next-generation wireless networks necessitates effective spectrum utilization, a challenge addressed by cognitive radio (CR) through enhancing spectral efficiency. In hybrid underlay-interweave CR, secondary users (SUs) adapt their transmissions when primary users (PUs) are active to avoid causing interference and operate at full power during idle spectrum periods. The primary network's tolerance for interference is directly influenced by the currently active PUs. Consequently, the primary network's interference threshold exhibits a dynamic characteristic. By accurately determining the channel activity of the primary network, SUs can effectively optimize spectrum usage. This strategy enables the SUs to have higher transmit power when permitted, resulting in higher performance gains. Therefore, we propose an unsupervised learning framework for sensing in cooperative hybrid CR networks to precisely determine the channel state of the primary network. Our unsupervised approach utilizes principal component analysis (PCA) for feature preprocessing and dimensionality reduction and a Gaussian mixture model (GMM) for channel state identification. Furthermore, our approach requires no prior knowledge and learns on a small amount of unlabeled sensing data. Our findings suggest that our proposed CR network can accurately and efficiently determine primary network channel states based on a variety of performance metrics. Moreover, we demonstrate that the proposed unsupervised framework outper-forms popular supervised learning techniques. Furthermore, it is shown that the proposed learning approach offers reduced complexity and is robust to low signal-to-noise ratios.

In the evolving landscape of the Internet of Things (IoT), integrating cognitive radio (CR) has become a practical solution to address the challenge of spectrum scarcity, leading to the development of Cognitive IoT (CIoT). However, the vulnerability of radio communications makes radio jamming attacks a key concern in CIoT networks. In this article, we introduce a novel deep reinforcement learning (DRL) approach designed to optimize throughput and extend network lifetime of an energy-constrained CIoT system under jamming attacks. This DRL framework equips a CIoT device with the autonomy to manage energy harvesting (EH) and data transmission, while also regulating its transmit power to respect spectrum-sharing constraints. We formulate the optimization problem under various constraints, and we model the CIoT device’s interactions within the channel as a model-free Markov decision process (MDP). The MDP serves as a foundation to develop a double deep Q-network (DDQN), designed to help the CIoT agent learn the optimal communication policy to navigate challenges, such as dynamic channel occupancy, jamming attacks, and channel fading while achieving its goal. Additionally, we introduce a variant of the upper confidence bound (UCB) algorithm, named UCB interference-aware (UCB-IA), which enhances the CIoT network’s ability to efficiently navigate jamming attacks within the channel. The proposed DRL algorithm does not rely on prior knowledge and uses locally observable information, such as channel occupancy, jamming activity, channel gain, and energy arrival to make decisions. Extensive simulations prove that our proposed DRL algorithm that utilizes the UCB-IA strategy surpasses existing benchmarks, allowing for a more adaptive, energy-efficient, and secure spectrum sharing in CIoT networks.

The Internet of Things (IoT) has attracted a lot of interest owing to its various applications. Cognitive IoT (CIoT) networks utilize the cognitive radio (CR) technology to relieve spectrum congestion and boost network performance. In this context, this article proposes a novel deep reinforcement learning (DRL) approach for joint power control and channel access coordination, tailored to energy-constrained CIoT networks. Unlike the existing works, our approach considers coordination dynamics between the competing devices and adopts a realistic energy harvesting (EH) model. The goal of the CIoT transmitter is to meet the interference constraint imposed by the primary network and coordinate channel access with the other CIoT devices while optimizing its lifetime and performance. We model the joint power control and access coordination problem as a model-free Markov decision process (MDP) and introduce a novel deep Q-network (DQN) architecture. This architecture enables a CIoT transmitter to autonomously make decisions regarding EH and data transmission, while also regulating transmit power to maximize the network’s performance and lifetime. These decisions incorporate critical factors, such as channel occupancy by other devices, EH opportunities, and interference constraints without prior knowledge. Through extensive simulations we demonstrate that the proposed DQN strategy achieves faster convergence than the benchmarks, facilitating adaptive, energy-efficient, and realistic spectrum sharing in CIoT networks. Additionally, our algorithm consistently achieves higher performance in terms of average sum rate, interference ratio, and rewards compared to the benchmarks.

In this paper, we present DEAP learning, an unsupervised approach for cooperative spectrum sensing. DEAP learning uses a sparse autoencoder (SAE) to learn a useful representation of the sensing data, followed by unsupervised clustering using affinity propagation (AP) algorithm for identifying primary user activity. Our suggested approach does not require prior information about the channel characteristics or the sensing data for training. Moreover, in contrast to other unsupervised clustering methods, the performance of AP is not dependent on cluster centroids initialization. DEAP learning leverages a few cooperating secondary users to minimize cooperation costs while achieving a high detection performance. Our numerical results suggest that our proposed sensing approach achieves comparable performance to supervised and state-of-the-art unsupervised deep learning-based sensing, without requiring a substantial amount of training data. To demonstrate the merits of DEAP learning, extensive simulations have been conducted under a wide range of primary and secondary network settings. Moreover, we evaluate DEAP learning while considering communication channel impairments. Overall, our DEAP learning approach illustrates the potential for robust performance in cooperative spectrum sensing.

Cognitive radio (CR) users can transmit data as vacant licensed bands become available. By using machine learning, CR users can intelligently sense channel activity and determine the availability of empty channels. Learning-based CR systems that use supervised learning for spectrum sensing require labeled training data. Furthermore, the majority of existing deep learning-based detectors are supervised, requiring a lot of labeled training data to achieve adequate performance. On the other hand, obtaining a large amount of labeled data in practical CR may be difficult. To address this gap, we propose DeepSense, which is an unsupervised cooperative sensing approach that uses representation learning by a sparse autoencoder (SAE) and unsupervised clustering by a Gaussian mixture model (GMM). DeepSense does not rely on cooperation among many SUs. Instead, it uses the learned representation to improve the detection performance, which significantly decreases the network's cooperation overhead. DeepSense does not require any prior knowledge, such as noise characteristics or channel state information, to operate. Furthermore, only a small amount of unlabeled data is needed for training. Extensive simulations have been conducted, which suggest that the proposed detector is able to learn hidden features in the sensing data that allows it to achieve an excellent detection performance. Moreover, our results show that DeepSense outperforms pure GMM, and attains comparable detection performance to benchmark deep supervised learning-based cooperative sensing.

In Cognitive radio (CR), users take advantage of vacant licensed bands to transmit their data as they become available. Cognitive users employ an autonomous perception-action decision cycle that starts with sensing the activity of licensed users. By using machine learning techniques, CR users can attain their full cognitive potential and smartly detect empty frequency bands. Learning-based CR systems that utilize supervised learning for spectrum sensing require labeled data for model training. Having readily accessible labeled data is a challenging task for CR networks, since it necessitates cooperation between licensed and unlicensed users. In interweave CR networks, such cooperation is not feasible and imposes a significant communication overhead. Motivated by the above, we address the practical limitation of labeled data scarcity in learning-based CR networks by designing a novel unsupervised learning framework for cooperative spectrum sensing based on a Gaussian mixture model (GMM) and principal component analysis (PCA) that uses a small amount of unlabeled data for training and requires no prior knowledge of the radio environment. The system is mathematically simulated, and its performance is evaluated based on various detection performance metrics. According to our findings, our proposed approach outperforms the GMM algorithm and is on par with supervised learning algorithms such as SVM, RF, and DT. Furthermore, the proposed approach is shown to be robust to low SNRs.

A cognitive radio (CR) network consists of wireless devices that opportunistically borrow vacant licensed bands. Cognitive users adaptively employ a perception-action decision cycle. Learning-based CR networks use past acquired knowledge of the radio environment to make smarter decisions. CR systems that use supervised learning for spectrum sensing require labeled data for training purposes. Having readily available labeled data is a complex task for CR networks, as it requires cooperation between the primary and secondary users. Such cooperation is not possible in interweave CR networks and imposes a cooperation overhead. Motivated by the above, we tackle the problem of labeled data scarcity in practical learning-based CR networks. We propose an unsupervised two-stage learning framework for cooperative spectrum sensing. The system combines the superior performance of the Support Vector Machine (SVM) and low cost training data of the Gaussian Mixture Model (GMM). A system model is proposed, and the system’s performance is evaluated based on the Receiver Operating Characteristics (ROC) and Area Under the ROC Curve (AUC). We obtain an upper and a lower performance bounds in terms of the AUC. The detection performance is compared for the SVM, GMM, and the proposed two-stage system based on the ROC. Additionally, we evaluate the detection performance of the CR network under different primary network sizes. Our results show that the two-stage learning approach attains a higher detection performance than the GMM algorithm, and achieves the same or comparable performance to the SVM algorithm.

In this paper, a cooperative Secondary Network (SN) is proposed that operates under a hybrid underlay-interweave model. The narrowband sensing problem under the interweave model was formulated as a binary hypothesis problem. The Fusion Center (FC) uses learning techniques, namely the Gaussian Mixture Model (GMM), Support Vector Machine (SVM), and Naive Bayes' (NB) to classify the state of the channel. An SVM kernel was chosen to best fit our hybrid network, since a nonlinear relationship exists between the energy vectors collected at the FC. Furthermore, the degree of the polynomial SVM kernel was manipulated to minimize classification errors. The multi-class SVM (MSVM) algorithm was reformulated to fit our multiple hypothesis problem in the underlay model. The performance of the hybrid network was evaluated based on the Receiver Operating Characteristics (ROC) and classification accuracy. In addition, the accuracy of the MSVM is improved through the cooperation of the SUs. Our results show that the proposed learning-based hybrid model is robust to low SNR environments, and yields an improved performance compared with traditional cooperative sensing techniques. Moreover, we show that the Gaussian SVM kernel surpasses other proposed learning algorithms achieving as high as an 80% detection rate with as low as 10% false alarm.

Vehicle-to-Everything (V2X) technology plays a critical role in maintaining road safety, avoiding accidents and controlling traffic flow. As self driving cars are expected to take over the roads, this paper discusses the intermediate phase in which manned and unmanned cars coexist. A heterogeneous network architecture that simultaneously serves manned and unmanned cars' different requirements in a suburban area is proposed and simulated using Riverbed Modeler. The feasibility of this architecture is examined in three different scenarios: Normal operation, congestion in both directions and Road Side Units (RSU) failure. In normal operation mode, traffic data is sent through Vehicle-to-Vehicle (V2V) and Vehicle-to-Infrastructure/Infrastructure-to-Vehicle (V2I/I2V or to RSU) using IEEE 802.11p and infotainment information is communicated as V2I/I2V using Long Term Evolution (LTE). A special case is highlighted and tested, in which congestion is in both directions. In such situation, data needs to be relayed to the nearest RSU using multi-hop communication. A fault-tolerant model is also proposed and analyzed in case of failure of RSU. The performance metrics are end-to-end delay, LTE response time, handover delay and packet loss ratio. The architecture proves its suitability by satisfying traffic control real time application requirements.