With the rapid advancement of AI, deep learning has been increasingly integrated into communication systems, driving the evolution of semantic communication. By extracting and transmitting task-relevant semantics from raw signals, this paradigm enhances efficiency while alleviating bandwidth and redundancy constraints. Multimodal data, as a dominant representation form, has shown promising results in tasks like VQA and cross-modal retrieval. However, its application in semantic communication remains nascent, with most studies limited to unimodal scenarios. Challenges persist in achieving modality-coordinated compression, cross-modal semantic alignment (the process of ensuring consistency between transmitted and received semantics, or aligning representations across modalities), and robust task-oriented transmission under channel variations. Addressing these issues necessitates the development of a unified multimodal semantic communication (MMSC) framework. Such efforts are crucial for enabling intelligent communication in future systems, including 6G, the metaverse, and IoV. This article reviews recent progress in MMSC, outlines key challenges, and explores future research directions in this emerging field.
22 July 2025 in IEEE Transactions on Vehicular Technology
A deep learning (DL)-based wavenumber domain channel estimation scheme is proposed for holographic multipleinput multiple-output (HMIMO) systems. Firstly, considering the spatially-continuous characteristics and the super-dense antenna array architecture in the HMIMO systems, an electromagnetically consistent wavenumber domain channel modeling is adopted to depict physical electromagnetic (EM) propagation model and the actual EM field distribution. Furthermore, an effective Mamba-based channel estimation model, named CMamba, is proposed to learn the spatially-continuous and super-resolution wavenumber domain channel features. The proposed CMamba integrates the convolutional neural network and Mamba to capture the local features and long-range dependency of the wavenumber domain channel. In addition, considering the difficulties in obtaining the wavenumber domain channel and the high costs incurred in practical applications, we develop a maskingbased pre-training framework based on the proposed CMamba to reduce the data dependency issues, aiming for few-shot channel estimation in HMIMO systems. Simulation results show that the normalized mean squared error performance of the proposed CMamba outperforms existing channel estimation schemes under various signal-to-noise ratios. In particular, for the case of limited channel labeled samples, the proposed masking-based CMamba model can effectively overcome overfiting and outperforms the state-of-the-art schemes by over 3 dB in average NMSE with approximately half of the complexity.
31 October 2025 in IEEE Transactions on Vehicular Technology
This paper presents a covert beamforming framework for a RIS-NOMA-assisted integrated sensing and communication (ISAC) system to enhance both sensing performance and covert signal transmission. Specifically, the dual-functional base station simultaneously transmits public and covert signals to users using the non-orthogonal multiple access (NOMA) protocol while performing sensing functions and avoiding covert signal detection by eavesdroppers. The reconfigurable intelligent surface (RIS) is employed to enhance both covert transmission and sensing performance. We aim to maximize the covert signal transmission rate by jointly designing the transmit beamforming vectors and the RIS reflection coefficients subject to the covert requirement, radar receive signal-to-noise ratio (SNR) constraint, public signal communication rate constraint, and the total power budget constraint. The coupling of variables and the high power of the radar SNR constraint make the problem challenging to solve. We propose an alternative algorithm to decouple the variables and divide the problem into two subproblems. The penalty-based method optimizes the transmit beamforming, while the majorization-minimization method transforms the quartic radar SNR constraint into a quadratic constraint. In addition, the imperfect eavesdropper's channel state information (ECSI) scenario is considered, and the proposed algorithm can also be employed to deal with this case after tackle the semi-infinite constraints caused by the imperfect ECSI. Simulation results demonstrate the advantages of our proposed covert transmission design in the RIS-NOMA-assisted ISAC system.