In this paper, we investigate covert communications in simultaneously transmitting and reflecting reconfigurable intelligent surface (STAR-RIS)-assisted ambient backscatter communication (AmBC) system with rate-splitting multiple access (RSMA). Specifically, we use STAR-RIS as a backscatter device to enhance the backscatter link against interference from the direct link. In addition STAR-RIS utilizes the RSMA protocol to transmit information to both covert and public users, enabling more efficient use of spectrum resources. In order to evaluate the system performance, we derive expressions for detection error probability (DEP), outage probability (OP), covert rate (CR) and spectral efficiency (SE). Then, we perform an optimal detection threshold analysis for DEP, an asymptotic analysis for OP, as well as a lower bound analysis for CR, and analyze the impact of system parameter changes on the system performance. The numerical results show that the STAR-RIS-RSMA assisted ambient backscatter communication (AmBC) system performs better in terms of the covert performance than the system using tags as well as the system using the non-orthogonal multiple access (NOMA) protocol.

Automatic modulation recognition of unknown signals is key technology in cognitive radio and military systems. However, low recognition accuracy under low signal-to-noise ratio (SNR) conditions remain a critical challenge. To address the issue of signal feature ambiguity induced by low SNR, we propose an adaptive multimodal modulation recognition method. First, a multi-branch feature extraction network is designed to comprehensively exploit discriminative feature from the temporal domain, frequency domain, and modulation domain. By leveraging the inherent advantages of the gated attention fusion mechanism in dynamically allocating multimodal feature weights, we obtain complementary-enhanced feature representations through cross-domain fusion of in-phase and quadrature signals, time-frequency representations, and constellation diagrams. Second, to enhance intra-class compactness and inter-class separability, we develop a loss function adaptation algorithm driven by the deep deterministic policy gradient, which automatically optimizes the weight ratios among cross-entropy loss, center loss, and class loss. Finally, considering the real-world scenarios with superimposed interference, we further design a robust multimodal adversarial training strategy to mitigate decision boundary degradation. By injecting simulated interference samples into multimodal training data, the model is compelled to learn interference-resistant decision boundaries while systematically enhancing robustness against interference. Experimental results on the RadioML2016.10a, RadioML2016.10b, and RML-Realistic datasets demonstrate that the proposed method achieves superior performance compared to state-of-the-art methods, with particularly significant improvements in recognition accuracy under low SNR conditions.

Recently, deep learning (DL)-based semantic communication has emerged as a promising technique for enhancing transmission efficiency in various tasks. By transmitting the extracted semantic features instead of the original data, the communication delay can be significantly reduced. In this work, we design a semantic communication technique for low-latency device-edge co-inference systems which supports multi-stage semantic feature transmission. To reduce the average inference delay, we develop a multi-stage edge inference scheme and the corresponding early stopping strategy with the proposed semantic communication network. The edge inference would be directly terminated if the inference result at any stage is regarded as reliable. We also develop an algorithm to jointly optimize the transmission duration and decision threshold for both the two-stage and multi-stage inference schemes, aiming at minimizing the average inference delay while satisfying the inference accuracy requirement. Simulation results show that the proposed multi-stage edge inference scheme with early stopping strategy can effectively reduce the average edge inference delay.

We propose a new cognitive technique for blind adaptive beamforming which uses a pre-trained deep learningbased signal classifier to protect a signal of interest (SOI) from interference. The method is an application of transfer learning since the features learned from noisy single-antenna captures are transferred to the task of beamforming using an antenna array. The proposed technique, called the classification-based transfer learning (CBTL) beamformer, forms a beam towards the SOI and nulls towards interference by maximizing a measure of classification confidence for the SOI. It does not require synchronization to the SOI or knowledge of the array manifold. We describe how the beamformer is optimized and the desired characteristics of the classifier. Extensive simulations show that the CBTL beamformer can improve the array signal-to-interference-plus-noise ratio by 8 dB or more compared to state-of-the-art techniques when using 64 samples from a 4-element array, which can be useful for dealing with dynamic environments. We validate this result using real-world data from an open-access radio testbed. The novelty of the proposed approach is in its ability to repurpose a classifier for a task that it was not trained for, eliminate the need for training data from an antenna array, and automate the development of beamformers without relying on expert-designed features.

Graph neural networks (GNNs) have been widely applied in software-defined network (SDN) to enhance network modeling and performance forecasting. However, the black-box nature of deep learning makes GNNs difficult to interpret, hindering their broad use and the application of GNN-based SDN systems in engineering. In this paper, we propose a novel interpretation framework named GEEK-Explainer, designed to efficiently provide instance-level interpretation of GNNs in SDN. Specifically, we introduce a KernelSHAP-based scoring module to generate intuitive and human-friendly explanations for each performance prediction. To address conflicts in computation cost, we propose a soft discrete mask matrix that identifies a critical set of important nodes. Extensive experiments demonstrate that the RouteNet model can effectively learn the relationships among features, which can provide a better understanding of the prediction process with less computation cost. These findings improve the transparency and robustness of the model and promote the application of GNN-based SDN systems in engineering practice.

Flying ad hoc networks (FANETs), consisting of unmanned aerial vehicles (UAV), and aircraft have become popular in the 5G era. However, a fast-response adaptive waveform control mechanism tailored for FANET applications has not yet been deeply investigated. In various scenarios, such as slow but crowded drone swarming networks for search and rescue, fast drone-show networks, and aircraft formations, it is a significant challenge to ensure that the RF modulation method can timely cope with the Doppler effect caused by node mobility changes. This paper introduces a Doppler adaptive waveform engine (DAWE) to adjust the node’s modulation method and parameters, according to the highly dynamic FANET conditions. An enhanced orthogonal time frequency space (OTFS) modulation method is implemented in DAWE to mitigate high Doppler shifts and multipath propagation across diverse FANET scenarios. We further propose a time-fused Transformer mechanism-based reinforcement learning (TTRL) algorithm to adjust the modulation parameters. This algorithm can focus on long-range dependencies better than traditional time-series forecast models and does not require the actor-critic dual model structure that is commonly used in deep reinforcement learning. Our simulations show that, across link SNRs of 20 – 40 dB and user speeds ranging from 100 km h¹ (pedestrian UAV loiter) up to 1000 km h¹ (high-speed fixed-wing transit), the DAWE protocol sustains a bit-error rate roughly one order of magnitude ( 10 ×) lower than a conventional orthogonal frequency-division multiplexing(OFDM)-only baseline.

Mobile edge computing (MEC)-enabled air-ground networks advance 6G wireless networks by utilizing aerial base stations (ABSs) such as autonomous aerial vehicles (AAVs) and high altitude platform stations (HAPS) to provide dynamic services to ground IoT devices (IoTDs). These IoTDs support real-time applications like multimedia and Metaverse services, which demand high computational resources and strict quality of service (QoS) guarantees, specifically in terms of latency and efficient task queue management. However, IoTDs often face constraints in energy and computational power, requiring efficient queue management and task scheduling to maintain QoS. To address these challenges, AAVs and HAPS are deployed to supplement the computational limitations of IoTDs by offloading tasks for distributed processing. Due to AAVs’ resource limitations, particularly in terms of power and coverage area, HAPS are used to enhance their capabilities and extend coverage. Overloaded AAVs may relay tasks to HAPS, creating a multi-tier computing system. This paper addresses the overall energy minimization problem in the MEC-enabled air-ground integrated network (MAGIN) by optimizing AAV trajectories, computing resource allocation, and queue-aware task offloading decisions. The optimization problem is highly complex due to the nonconvex and nonlinear nature of this hierarchical system, which traditional methods cannot effectively solve. To tackle this, we reformulate the problem as a multi-agent Markov decision process (MDP) with continuous action spaces and heterogeneous agents. We propose a novel variant of multi-agent proximal policy optimization (MAPPO) with Beta distribution (MAPPO-BD) to solve this problem. Extensive simulations show that MAPPO-BD significantly outperforms other baselines, achieving superior energy savings and efficient resource management in MAGIN, while adhering to constraints related to queue delays and edge computing capabilities.

With the explosion of computation-intensive and latency-sensitive vehicular applications, multi-tier edge computing systems consisting of vehicular edge computing (VEC) server, server vehicles (SeVs) and task vehicles (TaVs) have been proposed to provide task offloading services for resource-limited TaVs. Since normally both edge servers (VEC server and SeVs) and TaVs are selfish and care about only their own benefits, reasonable pricing strategies and task offloading decisions should be carefully designed to motivate them to help TaVs. To this end, we propose a multi-leader multi-follower Stackelberg game-based joint task offloading and pricing strategy for multi-tier VEC systems by assuming edge servers (ESs) to be leaders and TaVs to be the followers. With the game, ESs maximize their own utility by optimizing the pricing strategies of their computing resources. According to the computing resources’ prices set by ESs, the TaVs as the followers make task offloading decisions to maximize their individual utility by jointly designing user association, bandwidth purchasing volume and computing resources purchasing volume. In solving the optimization problem, we prove there exists a unique Stackelberg equilibrium (SE) solution. The task offloading problem of the followers is decomposed into a resource purchasing sub-problem and a user association sub-problem, which are first solved by a joint optimization of bandwidth and computing (JOBC) algorithm and a Gale-Shapley-based matching game bilateral selection (GSMB) algorithm, respectively. Then, based on the followers’ task offloading decisions, the pricing problem is solved by a bisection search method. Finally, a joint optimization task offloading and pricing strategy (JOTPS) algorithm is established to obtain the SE. Simulation results show that the proposed scheme can maximize the individual utility of ESs and TaVs compared to those existing benchmark schemes.

To jointly tackle the challenges of data and node heterogeneity in decentralized learning, we propose a distributed strong lottery ticket hypothesis (DSLTH), based on which a communication-efficient personalized learning algorithm is developed. In the proposed method, each local model is represented as the ‌Hadamard product of global real-valued parameters and a personalized binary mask for pruning. The local model is learned by updating and fusing the personalized binary masks while the real-valued parameters are fixed among different agents. To further reduce the complexity of hardware implementation, we incorporate a group sparse regularization term in the loss function, enabling the learned local model to achieve structured sparsity. Then, a binary mask aggregation algorithm is designed by introducing an intermediate aggregation tensor and adding a personalized fine-tuning step in each iteration, which constrains model updates towards the local data distribution. The proposed method effectively leverages the relativity among agents while meeting personalized requirements in heterogeneous node conditions. We also provide a theoretical proof for the DSLTH, establishing it as the foundation of the proposed method. Numerical simulations confirm the validity of the DSLTH and demonstrate the effectiveness of the proposed algorithm.