Shi Jin

Task-oriented communication leverages semantic information for efficient transmission by focusing on meanings relevant to downstream tasks rather than exact symbol recovery. In multimodal scenarios, most existing approaches adopt static fusion, where all modalities are always processed jointly. However, this often leads to redundant communication overhead, since some samples can be accurately inferred from a single modality, whereas more ambiguous cases require additional modalities. To address this issue, we propose a multimodal task-oriented communication system based on dynamic modality selection, where conditional mutual information (CMI) is used to assess the incremental value of each modality. A gating network leverages these estimates to select task-relevant modality combinations, achieving adaptive resource allocation. Experiments on movie genre classification and sentiment analysis show that the proposed system sustains competitive accuracy while reducing both computation and transmission costs by over 30%, providing an interpretable and efficient solution for resource-constrained task-oriented communication.

WiFi channel state information (CSI) enables privacy-preserving, device-free continuous authentication on commodity hardware. However, CSI is highly sensitive to room layout and walking routes. When several people walk at the same time, their signals overlap and create strong inter-person interference. Many existing gait authentication systems degrade substantially under realistic cross-room, cross-route, and multi-person settings. As a result, we present AdaGait, a domain-adaptive multi-person gait authentication system built on a pair of WiFi devices. AdaGait targets cross-domain deployment, where training and testing differ in rooms, walking routes, and the number of persons. AdaGait first stabilizes CSI measurements via bandpass filtering, wavelet denoising, and conjugate multiplication. To better use the multi-subcarrier structure and improve sample efficiency, AdaGait constructs a subcarrier-frequency map and uses window-slicing data augmentation to expand training instances without extra data collection. For classification, AdaGait adopts a CNN-Transformer backbone together with a weakly supervised asymmetric tri-training scheme. This scheme adapts from labeled single-person source domains to weakly labeled multi-person target domains by injecting set-level label composition into pseudo-label screening. We implement AdaGait in multiple indoor rooms, walking routes, and crowd sizes. Extensive experiments show that AdaGait achieves over 90% authentication accuracy for one- and two-person cases and above 80% for three and four persons. AdaGait also consistently outperforms state-of-the-art CSI-based methods in all multi-person and cross-domain scenarios, demonstrating strong domain-adaptive performance.

The rapid proliferation of heterogeneous wireless devices in sixth-generation (6G) networks raises significant security challenges, necessitating continuous and zero-trust authentication. Radio frequency fingerprint identification (RFFI) is a promising solution, however, its performance deteriorates markedly under dynamic and unpredictable wireless environments. To address this issue, we propose DrffNet-X, a cross-modal RFF distillation framework to train a lightweight and self-robust RFFI model within a cloud-edge collaborative agentic artificial intelligence (AI) paradigm. In the proposed framework, wired and wireless signals are used to train a large teacher network and a small student network, respectively, enabling the student to learn channel-robust RFF representations. Specifically, we first develop the DrffNet-ResNet, which is based on the ResNet backbone and the spatial attention mechanism. The complexity of the proposed DrffNet-ResNet is analyzed, where the student network achieves a 27% reduction in parameters, but has the similar performance as the teacher network. Next, to prove the model-agnostic of the proposed framework, we also develop the DrffNet-MobileNet and evaluate the performance via the experiments. The experiment results demonstrate that the developed DrffNet-ResNet and DrffNet-MobileNet achieve an accuracy of above 95% and 93% on Long Range (LoRa) radio frequency (RF) datasets, respectively, under different channel effects and various signal-to-noise ratios (SNRs), and also achieve an outstanding performance on unmanned aerial vehicle (UAV) RF datasets. The developed networks have significant advantages over the state-of-the-art deep learning (DL)-based RFFI methods in terms of accuracy, computation amount and universal applicability. Extensive system-level evaluations verify that the proposed cloud–edge collaborative RFFI framework enables efficient deployment, low-latency edge inference, and stable performance under continuous model updates.

Environment-aware 6G wireless networks demand the deep integration of wireless and multimodal data. However, most existing datasets are confined to 2D terrestrial far-field scenarios, lacking the 3D spatial context and near-field characteristics crucial for low-altitude extremely large-scale multiple-input multiple-output (XL-MIMO) systems. To bridge this gap, this letter introduces Multimodal-NF, a wireless-centric, large-scale multimodal dataset and its generation framework. Operating in the upper mid-band, it synchronizes high-fidelity near-field channel state information (CSI) and precise wireless labels (e.g., Top-5 beam indices and a binary LoS indicator) with comprehensive sensory modalities (RGB images, LiDAR point clouds, and GPS). Crucially, these multimodal priors provide spatial semantics that help reduce the near-field search space and thereby lower the overhead of wireless sensing and communication tasks. Finally, we validate the dataset through representative case studies, demonstrating its utility and effectiveness. The open-source generator and dataset are available at https://github.com/Lmyxxn/Multimodal-NF.