This paper investigates physical-layer security (PLS) in UAV-assisted communication networks featuring an untrusted amplify-and-forward aerial relay. To counter potential eavesdropping threats, a destination-aided jamming strategy is employed. The proposed solution adopts a hybrid framework that integrates particle swarm optimization (PSO) and a dual-layer deep neural network (DNN). In this framework, PSO is responsible for optimizing the three-dimensional (3-D) placement of the UAV relay by minimizing the total aggregate path loss, while the DNN tackles the transmit antenna selection (TAS) problem at the source to maximize the average secrecy rate. The effectiveness of the suggested approach is assessed by comparing diverse relay placement techniques and antenna selection methods. The simulation results demonstrate that the proposed framework consistently outperforms conventional machine learning (ML)-based approaches and random baselines. Interestingly, single-antenna selection achieves slightly better secrecy rates than dual-antenna configurations. Overall, the proposed method provides a scalable and computationally efficient solution for enhancing the security of UAV-enabled communications in scenarios involving untrusted relays.

In this paper, multiple relay selection (MRS) schemes for untrusted unmanned aerial vehicle (UAV)-enabled networks are proposed. In this context, various machine learning (ML) models are employed to improve secrecy performance by optimally selecting multiple aerial relays from the available ones, all of which are regarded as untrusted. Notably, these ML models can cope with the quickly changing and random positioning of the aerial relays, effectively decoupling the intricate coupling relationship between the secrecy rate, channel coefficients, and inter-node distances. The results indicate that the ML-based MRS schemes obtain sufficient accuracy and better decoupling than a respective exhaustive searching (ES) approach, while also maintaining a lower computational complexity.

An appealing alternative towards dealing with mobility and coverage challenges in future wireless networks, are hybrid satellite-terrestrial networks (HSTNs). The combination of network coding (NC) and non-orthogonal multiple access (NOMA) techniques, have recently gained popularity due to their performance gains in improving the quality of a wireless transmission and exploiting the accessible spectrum. This paper introduces a joint NOMA-NC approach, integrated to an HSTN comprising a satellite, a terrestrial base station (BS), and multiple ground-based mobile terminals (MTs). The suggested method facilitates a cluster of end-users to be concurrently served through NOMA from a satellite using optimal power allocation, whereas the BS utilizes random linear NC to augment the MTs’ reception in the event of errors. The simulation results confirm the effectiveness of the proposed methodology in improving system sum throughput, demonstrating substantial enhancements compared with both OMA with NC, and standalone NOMA and OMA approaches. The resilience of the suggested method is proved, improving the sum throughput even at low power transmissions and in the worst case channel conditions.

This paper proposes a scheme for rapidly deploying a UAV-enabled emergency cellular network (UECN) in disaster scenarios, such as earthquakes or floods, to support rescue operations. The unsupervised placement of UAV aerial base stations (ABSs) is achieved through machine learning (ML) techniques. Specifically, the k-medoids algorithm is utilized to cluster ground users in the disaster area and determine the minimum number of ABSs and their position. ABS's altitude is defined based on its capacity capabilities and propagation environment. The UECN cooperates with undamaged cellular infrastructure via joint coordinated multipoint transmission and reception (CoMP) with neighbouring functional terrestrial macrocell base stations (TBSs) to improve end-user signal quality. Finally, the proposed scheme is comparatively evaluated through simulation, and the results demonstrate its efficiency even in the presence of users and ABSs positioning errors.

This paper presents prediction path loss models in an urban environment for cellular networks with the help of machine learning methods. For this goal, Support Vector Regression (SVR), Random Forest (RF) and K-Nearest Neighbor (KNN) algorithms are exploited and assessed. The training and testing procedure is carried out with the help of a path loss dataset generated by simulated results considering a Long Term Evolution (LTE) network utilizing a digital terrain model. The simulation takes into account an urban environment for both line-of-sight (LOS) and non-LOS (NLOS) propagation condition. The results reveal that all the evaluated algorithms forecast path loss with a remarkable accuracy, providing root-mean-square errors on the order of 2. 1-2.2dB for LOS and 3. 4-4.1dB for NLOS locations, respectively. Among the examined algorithms, KNN shows the best performance, thus being an appealing option to predict path loss in urban areas. For comparison purposes, the COST231 Walfisch-Ikegami empirical model was applied, which presents the worst performance, providing the highest errors under-predicting path loss, especially in NLOS locations.