Semantic communications (SemCom) provide efficient transmission for mobile edge computing (MEC) services by extracting critical semantics from raw information. Although widely adopted in various scenarios, existing single-modal SemCom systems struggle to efficiently support edge multimodal data transmission. Additionally, mobile end users have varying communication requirements across different modalities. However, existing work lacks the ability to generate personalized communication policies tailored to diverse intents (Typically, communication policies include bandwidth allocation and modulation and coding schemes, etc.). In this paper, we propose an edge Cognitive SemCom Agent (CSCA) to facilitate edge multimodal SemCom. Specifically, CSCA leverages an edge Large AI Model (LAM) to realize modality alignment and natural language intent understanding. Moreover, we develop a communication planning module to realize the planning capability, which generates personalized wireless communication policies based on LAM’s environment and intent cognition. Particularly, to assess the efficiency of communication policies in multimodal SemCom and capture intent competition, we present a novel indicator named cognitive SemCom quality indicator (CSCQI). Then, we use the denoising diffusion probabilistic model to optimize the generation policy. Extensive experimental results demonstrate that CSCA achieves an average improvement in intent satisfaction rate and semantic accuracy by 42.19% and 29.75% respectively, while reducing communication delay by 33.40% .

Transmitting multimodal data through semantic communication offers a promising way to enhance the quality of experiences. However, existing single-modal semantic communication systems struggle to efficiently support multimodal data transmission. Additionally, users have different communication requirements for different modalities, while existing work lacks the capability to generate personalized communication schemes tailored to diverse requirements. In this paper, we propose a cognitive multimodal semantic communication system. At its core is a cognitive semantic communication agent (CSCA) powered by edge large AI model (LAM), enabling low-latency modality alignment and natural language intent understanding. The CSCA integrates a cognitive communication planning algorithm that leverages intent cognition and environment cognition to create personalized communication schemes for users. Experimental results demonstrate that our system outperforms baseline systems in terms of semantic accuracy, intent satisfaction rate and communication latency.

Blockchain’s anonymity and decentralization improve financial efficiency but also facilitate illicit activities like money laundering and fraud. In the field of blockchain illicit transaction detection, researchers are confronted with several challenges, including the difficulty of analyzing high-dimensional data features, the imbalance in the quantity of training data, and the absence of certain features in the training data. This paper proposes a reinforcement learning-enhanced temporal generative adversarial networks (DRL-TGAN) model for detecting illicit transactions in blockchain. Firstly, this model adopts a dynamic feature selection strategy to achieve feature compression, saving computing resources without affecting accuracy. Secondly, we tackle the issue of training data imbalance by introducing noise during the generator training process to synthesize illicit data. Moreover, we propose a sliding window method based on TCN to capture the dynamic changes and long-term dependencies between transactions at different timesteps. We conducted the experiment using two Elliptic dataset, DRL-TGAN achieved a precision of 96.4%, outperforming existing methods in handling imbalanced and incomplete data.

With the rise of AI-Generated Content (AIGC) services in wireless networks, efficient and high-quality distributed training of Large Language Models (LLMs) has become essential for enabling the large-scale application of next generation AI technologies. However, the extensive parameters of LLMs impose significant demands on memory, computing power and communication resources in heterogeneous networks. To efficiently utilize the dispersed network resources, this paper presents a First-Pipeline- Then-Federated Learning (FPTFL) approach with a hybrid parallel scheduling strategy to facilitate the training of Transformer-based LLMs. We propose a block-wise splitting mechanism to partition the Transformer's encoder into distinct segments, which are deployed cross individual devices. The encoder parameters and intermediate smashed data are uploaded to the edge server, where the whole model is updated through federated aggregation. Particularly, we develop a fine-grained computation-efficient method based on pipeline parallelism, enabling the segments to cooperatively train the entire encoder. An optimization problem is formulated to determine the LLM segments and the number of micro-batches under network resource constraints, with the goal of minimizing the total latency of LLM training services. Simulation results demonstrate that our approach enables Transformer-based model training on resource-constrained devices, preserves model performance, and reduces waiting time.

In Internet of Things, V-OTDR distributed fiber optic sensors are often used to collect and monitor event signals. Recently, deep learning models have been widely used in V -OTDR event recognition systems. However, traditional event signal classification models only can recognize known events, and they identify known events by constructing class boundaries without intra-class compactness. A signal feature space method for recognizing V -OTDR events is proposed in this paper. This method can not only effectively construct class boundaries, but also reduce intra-class distances. The more accurate class boundaries will help to reject the unknown class samples, which will inversely improve classification accuracy of unknown classes. The experimental results show that the proposed method has an accuracy of 91.851% in recognizing known events and 90.314 in recognizing unknown events, with an average recognition accuracy improvement of 13.6% compared to traditional methods.

In the industrial Internet of Things (IIoT), blockchain technology has been employed to ensure the trustworthiness of federated learning (FL) services. However, the existing framework that combines blockchain and FL suffers from poor training performance and high computational overhead due to the complex consensus mechanism. Although recent studies have explored architectures that integrate Directed Acyclic Graph (DAG) with FL, the aggregation process in DAG-based multi-branch structures still faces significant challenges due to strong statistical heterogeneity across branches. To accommodate the heterogeneity, this paper proposes a DAG-based asynchronous aggregation framework for decentralized FL services. In this framework, the local models are aggregated with global models in the DAG ledger to form a new transaction block (TB). The verified TB becomes the subsequent node of the tail node in the DAG multi-branch structure. Additionally, an FL model delivery mechanism based on improved ensemble distillation is designed. This mechanism merges the models in verified TBs from multiple branches of the DAG, enhancing the accuracy of the final delivery model without compromising system training efficiency. Extensive ablation and comparative experiments demonstrate that our proposed scheme enhances the training efficiency and accuracy of DAG-FL systems while ensuring the security and trustworthiness.

To mitigate global climate change, the replacement of conventional coal-fired power generation with clean energy sources such as photovoltaics (PV) has become a key strategy. However, solar power output is highly variable because it depends on the amount of sunlight reaching the ground, referred to as Downward Shortwave Radiation (DSR). Accurately forecasting DSR in the short term is therefore critical for the stable integration of large-scale PV systems into urban power grids. Existing methods typically adopt a retrieve-then-forecast paradigm (first deriving physical products, then forecasting them), which performs poorly under rapidly varying atmospheric conditions. We propose a new forecast-then-retrieve framework for short-term DSR prediction based on geostationary satellite observations. Unlike conventional approaches, our method first forecasts the full-spectrum L1B radiances from Himawari-8 for the next 3 hours, and then retrieves DSR values from the predicted radiances. To support this framework, we design AtmoNet, a multimodal network that takes the past 3 hours of Himawari-8 L1B radiances as input and captures spatiotemporal patterns of reflectance, water vapor, and longwave emission. Experiments show that this approach improves DSR prediction accuracy by up to 6.8% in the 1–3 hour forecast window. Moreover, in direct L1B forecasting tasks, AtmoNet outperforms leading models, including the state-of-the-art Swin Transformer, particularly in capturing complex, moisture-driven phenomena such as localized convection and evaporation. By enabling accurate and scalable DSR forecasting, this work supports the stable integration of renewable energy into power grids and has the potential to contribute to global efforts to reduce carbon emissions.

Large-scale models, also referred to as pretrained models, have attracted significant attention due to their outstanding performance and robust generalization. However, the high demands for data quality, stringent data privacy and security requirements, and limited computational and communication resources have imposed restrictions on the further development of large-scale models. Federated Learning (FL) is a popular machine learning framework that can effectively address this issue. However, when collaboratively training large-scale models using FL, local training and the transmission of large-scale parameters impose significant computational and communication burdens on mobile devices. This paper proposes a light-weight and high-efficiency federated learning framework (FedLH) for large-scale models. This framework divides large-scale models into semantic block-based submodels, allowing clients to transmit these submodels to the server for heterogeneous aggregation. This approach enables both communication and computational efficiency. With the proposed framework, each device can learn personalized, structured sparse models that can efficiently run on terminal devices. Experimental results demonstrate that FedLH outperforms other baseline algorithms by significantly reducing the number of training parameters and transmitted data. It also exhibits strong generalization and scalability.

Foundation models (FMs), known as pre-trained models, have garnered significant interest in Industrial Internet due to their remarkable performance and robust generalization capabilities in downstream tasks. However, with the increasing requirements of computing infrastructure and data privacy protection for large foundation models, existing learning frameworks face challenges such as data privacy leakage, poor scalability, and deployment difficulties. To address these issues, this paper proposes a novel collaborative Transformer Block (TB)-wise training framework based on Federated Learning (FL), which consists of three stages: pre-training, graph regularization, and personalized training. To tackle the challenge of statistical heterogeneity in distributed data, we design a Graph Convolutional Network (GCN)-based update operator that captures local training representations. Besides, we conduct an analysis based on feature similarity to enhance the interpretability of our algorithm. We choose popular vision Transformer models for the experiments, extensive results demonstrate that our framework can jointly train multiple clients to build a foundation model while improving the single client's personalized performance. The proposed method outperforms state-of-the-art frameworks under various data distributions and system heterogeneity settings, highlighting its robust performance.