Mobile Edge Computing (MEC) enables the delegation of computing tasks from Consumer Electronics (CEs) to edge servers. This offloading process significantly reduces the latency and energy consumption associated with CEs. Nonetheless, Deep Reinforcement Learning (DRL)-based offloading techniques often encounter challenges in reaching optimal solutions within a confined number of iterations due to the inherent complexity of the task. In light of this challenge, this paper introduces an approach that integrates DRL with Hyper-dimensional Networks (HDN) for task offloading, aiming to improve the efficiency of MEC systems. First, we establish a dynamic model of the MEC system and formulate the task-offloading problem to minimize the cumulative cost incurred by the MEC. Subsequently, we advance an offloading algorithm grounded in HDN principles. The experimental findings demonstrate that DRL with HDN leads to a marked reduction in the computational overhead of MEC systems when contrasted with alternative methodologies. Compared to the baseline algorithm, the proposed HDN-enhanced DRL reduces energy consumption, latency, and system consumption by 10.3%, 14.5%, and 10%, respectively.

With the deep integration of Sixth-generation (6G) networks and the Industrial Internet of Things (IIoT), security threats are becoming increasingly sophisticated. Recently, Large language models (LLMs) have demonstrated unique potential in handling massive heterogeneous data and unknown attack patterns within 6G-enabled IIoT environments. However, challenges remain in terms of addressing fragmented information and complicated and high-dimensional spatio-temporal correlation in industrial contexts. To this end, this article proposes an intrusion detection framework that integrates Knowledge Graphs (KGs) with LLMs. We first systematically review the key technologies of 6G IIoT and LLMs, and analyze in depth the application of LLMs to intrusion detection in 6G IIoT as well as the associated challenges. The framework encodes domain knowledge into structured KGs and fuses them with LLMs, thereby enhancing the model’s capability to understand and reason about attack behaviors in complex industrial scenarios. Extensive experiments demonstrate that the proposed KG-LLM achieves superior performance in intrusion detection, outperforming traditional machine learning models such as XGBoost, Random Forest, and Logistic Regression in accuracy, precision, recall, and F1-score. The model particularly excels in detecting attacks, showing robust performance even under noisy conditions. Finally, we discuss promising future research directions.

This paper explores the issue of collision-free optimal tracking control of autonomous aerial vehicle (AAV) and autonomous ground vehicle (AGV) formation under Byzantine attacks. Based on collision-free margin and collision risk angle, a collision avoidance scheme for AAV-AGV swarms is proposed, which takes into account both external and internal aspects of the formation. Meanwhile, considering that the AAV-AGV swarm is susceptible to the propagation of incorrect neighbors' information and false input signals (called Byzantine attacks), which is effectively reinterpreted as the management of unknown variables within the control inputs. Then, a barrier function and a control force direction function are introduced in the optimal performance index for collision avoiding with each system's radius, and a collision-free optimal control strategy under the reinforcement learning (RL) algorithm is investigated. Furthermore, the neural network is utilized to model the Byzantine attacks and appropriate unknown factors arising from Hamilton-Jacobi-Bellman (HJB) equation, the actor and critic adaptive laws are presented in actor-critic-identifier architecture. Subsequently, a collision-free optimal tracking control scheme is proposed to ensure safe collaborative moving of the AAV-AGV formation under Byzantine attacks. Finally, simulations are performed to validate the effectiveness of the proposed approach.

With the rapid development of consumer autonomous intelligent systems, ensuring the security of Internet of Vehicles (IoV) has become a critical challenge. Machine Learning (ML) and Deep Learning (DL) techniques are widely used to develop classifier-based Intrusion Detection Systems (IDS). However, traditional methods often suffer from high computational complexity, limiting their effectiveness in real-time detection. Hyperdimensional Computing (HDC), a brain-inspired machine learning paradigm, offers a compelling combination of high precision with exceptional robustness and training efficiency. In this paper, we present Hyperdimensional Intrusion Detection in IoV (HIDIV), a lightweight and efficient framework designed explicitly for IoV security. HIDIV introduces a dynamic error update module, enabling faster convergence and higher accuracy than conventional HDC methods. Experimental results show that HIDIV significantly accelerates training and inference, achieving speedups of approximately 43.4% and 33.2% over state-of-the-art machine learning methods while maintaining comparable accuracy. Furthermore, the proposed framework also surpasses traditional HDC in terms of accuracy.

Spatio-Temporal (ST) data is pervasive on the various aspects in our daily lives. By mining the ST information from the data, we are able to predict trends in numerous domains. The Transformer and one of its more recent enhancements, foundation models, have achieved a remarkable success in such ST prediction. In this article, we first survey the state of the art of Transformers-related work, then introduce the network architecture of the Transformer and summarize the improvements to adapt to the ST prediction Transformer and foundation models, including module enhancement and adjustment. Subsequently, we categorize the ST Transformer and foundation models in selected applications in some relevant domains, mainly urban transportation, climate monitoring, and motion prediction. Next, we propose an evaluation method in the ST prediction with Transformers and foundation models, list the most relevant open-source datasets, evaluation metrics and performance analysis. Finally, we discuss some future directions on the task of ST prediction with Transformer and foundation models. Relevant papers and open-source resources have been collated and are continuously updated at: https://github.com/cyhforlight/Spatio-Temporal-Prediction-Transformer-Review.

The emerging 6G-Internet of Things (IoT) paradigm is driving communication toward intelligent services, semantic communication enables efficient semantic sharing via artificial intelligence (AI), significantly boosting communication efficiency. However, current semantic systems suffer from closed-box decision-making, while existing explainable AI (XAI) methods face two key challenges: explicability granularity mismatch and closed-loop optimization gap. To address these, we propose a semantic communication framework integrated with XAI (XAI-SCS). Specifically, we first design an explainable semantic codec architecture enhanced by Kolmogorov-Arnold networks (KAN), where traditional fixed activation functions are replaced with learnable and parameterized ones, enabling function-level visualization to improve model explainability. Second, we develop an explainable semantic transmission module driven by contrastive learning that enhances the robustness of semantic transmission, and incorporating a semantic separability metric to quantify channel impacts on semantic integrity. Third, we introduced a KAN-enhanced causal semantic decoder, which integrates counterfactual interventions to generate pixel-level difference maps. We also propose a contrastive explanation consistency metric to evaluate the sensitivity of key features, enhancing the quality of the reconstruction. The experimental results show that our approaches enhance explainability across the entire decision process, achieve a significant accuracy improvement of up to 55% on the CIFAR-10 dataset with a bandwidth compression ratio of 1/25, and also obtain competitive image reconstruction quality in increasing compression levels. The source code is publicly available at: https://github.com/guyuangui/XAI-SCS.git.

Unmanned aerial vehicles (UAVs) play a crucial role in tasks such as target detection due to their speed and high degree of freedom. However, the limited computing power of UAVs and the complexity of their communication environments pose significant challenges for real-time tasks. To ensure accuracy and reduce the amount of transmitted data, we propose a novel multi-task joint cross-modal semantic communication system, named “CMSCDet”, for object detection and image transmission. In this system, the UAV captures both visible and infrared images, which are fused in the mid-term and transmitted to a server using joint source-channel coding (JSCC). The system employs multi-level feature extraction, fusion, and reconstruction techniques to perform both object detection and image reconstruction tasks simultaneously. Simulation results demonstrate that the proposed method achieves performance improvement over 10

Multi-agent federated learning requires all agents to stay online until the global model converges, and its multi-round training incurs substantial on-agent computing. However, most edge mobile agents with limited computational capacity and battery life may drop out of FL training prematurely due to low battery, which can degrade the convergence speed and accuracy of the global model. Prior works assign heterogeneous models to different agents to reduce on-agent computational energy, but neglect the changes in data distributions and computational capacities on agents. This makes it difficult to save on-agent computational energy and to support global aggregation. To this end, we propose a hierarchical Federated learning framework based on Adaptive Local Training (FedALT). Specifically, FedALT can adaptively select the most suitable local model for each agent based on its computational capacity and data distribution to reduce on-agent computational energy. To achieve finer-grained savings in computational energy, FedALT dynamically adjusts the number of local epochs for each agent during local training. Furthermore, FedALT conducts edge aggregation through model layer sharing and knowledge distillation to improve the convergence speed and accuracy of the global model. We compare FedALT with three baselines by training six models on two Non-IID datasets. Experiments on both the simulation platform and the hardware platform demonstrate that FedALT can achieve the lowest computational energy, the highest average global model accuracy, and the fastest global model convergence.

Advanced metering infrastructure in smart grids faces critical security challenges from distributed denial of service (DDoS) attacks, while traditional traceback methods suffer from resource constraints that limit their mitigation effectiveness. To address this, we propose a reverse auction-driven fog resource allocation strategy that innovatively employs a reverse auction mechanism to coordinate distributed fog resources for collaborative DDoS attack mitigation. Specifically, the framework is formulated as a 0-1 integer programming model, for which we develop a second-price sealed-bid auction (SPSA) algorithm that reduces computational complexity from factorial to polynomial time while maintaining allocation optimality. Experimental results demonstrate that the SPSA algorithm outperforms existing baseline methods, significantly improving both the efficiency and security of DDoS defenses in smart grid environments.

With the development of real-time video conferences, interactive multimedia services have proliferated, leading to a surge in traffic. Interactivity becomes one of the main features on future multimedia services, which brings a new challenge to Computer Vision (CV) for communications. In addition, many directions for CV in video, like recognition, understanding, saliency segmentation, coding, and so on, do not satisfy the demands of the multiple tasks of interactivity without integration. Meanwhile, with the rapid development of the foundation models, we apply task-oriented semantic communications to handle them. Therefore, we propose a novel framework, called Real-Time Video Conference with Foundation Model (RTVCFM), to satisfy the requirement of interactivity in the multimedia service. Firstly, at the transmitter, we perform the causal understanding and spatiotemporal decoupling on interactive videos, with the Video Time-Aware Large Language Model (VTimeLLM), Iterated Integrated Attributions (IIA) and Segment Anything Model 2 (SAM2), to accomplish the video semantic segmentation. Secondly, in the transmission, we propose a two-stage semantic transmission optimization driven by Channel State Information (CSI), which is also suitable for the weights of asymmetric semantic information in real-time video, so that we achieve a low bit rate and high semantic fidelity in the video transmission. Thirdly, at the receiver, RTVCFM provides multidimensional fusion with the whole semantic segmentation by using the Diffusion Model for Foreground Background Fusion (DMFBF), and then we reconstruct the video streams. Finally, the simulation result demonstrates that RTVCFM can achieve a compression ratio as high as 95.6%, while it guarantees high semantic similarity of 98.73% in Multi-Scale Structural Similarity Index Measure (MS-SSIM) and 98.35% in Structural Similarity (SSIM), which shows that the reconstructed video is relatively similar to the original video.

In recent years, AI Agents centered on large language models (LLMs) have flourished, providing a prerequisite for the robots with embodied intelligence in the new era. Using LLM as the brain of AI agents can endow robots with powerful abilities of understanding, reasoning and decision- making. Moreover, semantic communication is considered as a key technology to drive the development of multi-robot collaborative communication, which provides a novel solution for multi-robot systems. However, current multi-robot systems suffer from several problems when dealing with complex tasks, including chaotic agent cooperation, semantic relaying distortion, and inefficient topology, combined with the powerful autonomy of LLMs. To address these issues, we propose a task-driven semantic autonomous collaboration framework with AI Agent in multi-robot system. First, we introduce the different roles to achieve an efficient autonomous decomposition by AI Agents facing complex tasks through the dynamic iterative leadership election. Second, we propose a cooperative forwarding scheme according to the relay AI agent based on the prediction of robot behavior patterns and hybrid communication mechanisms. Third, a dynamic task-driven topology optimization in multi-robot systems by graph neural network (GNN) is proposed to achieve fine-tuning of the semantic communication topology through sparse regularization. Finally, the results of the simulation experiment show that the task execution accuracy of our method on the GSM8K dataset is 6\% higher than that of LLM-Debate, verifying the effectiveness and feasibility of this framework in multi-robot cooperative communication for complex tasks in the future.

Large Language Models (LLMs) can stimulate intelligent functions on mobile devices in the applications of consumer electronics. However, it is challenging to deploy LLMs on mobile devices due to limited memory and computational resources that fluctuate as devices execute multiple AI tasks concurrently. Quantization is a widely adopted compression technique, but faces two major difficulties in practical deployment: 1) Existing quantization methods are unable to select suitable models for deployment according to the resource limitations of mobile devices; 2) Mixed-precision quantization-based methods utilize reinforcement learning and element-level quantization, which incurs substantial computational and memory costs on mobile devices. To this end, we propose Resource Awareness-Based Salience-Weighted Mixed-Precision Quantization (RAS-MQ). Specifically, RAS-MQ constructs a memory and latency computation incorporating model parameters and device constraints, then RAS-MQ employs integer linear programming for optimal per-layer bit-width allocation. Weight matrices are partitioned into groups with bit-widths assigned based on average saliency ranking. Furthermore, RAS-MQ applies training-free quantization range clipping to salient weights, which can improve representational capacity while reducing quantization error. Experiments on five zero-shot inference datasets, for the 2-bit quantization of LLaMA2-13B, RAS-MQ outperforms the state-of-the-art baseline (SqueezeLLM) by 0.85% in accuracy and reduces weight memory consumption from 24.5 GB to 4.1 GB, achieving a $5.98\times $ compression ratio.

Hierarchical Federated Learning (HFL) offers a scalable, privacy-preserving framework for training models on massive End Devices (EDs). However, frequent model parameter exchanges cause excessive communication energy consumption on energy-constrained EDs. Premature dropouts of EDs due to battery depletion slow down convergence, degrade model accuracy, and reduce participant diversity, undermining the integrity and robustness of the federated system. Existing methods save device-side communication energy by adjusting the HFL architecture, but fail to achieve a satisfactory trade-off between energy consumption and global model accuracy, a balance crucial for maintaining the long-term device participation and thus the system’s overall integrity. To address this challenge, we propose HFELAG, a synergistic hierarchical federated learning framework that, for the first time, co-designs inter-group device association and intra-group aggregation to enhance system robustness. HFELAG introduces a device metric measurement mechanism to capture both communication capabilities and data distribution impacts. Based on these metrics, it implements an elastic grouping strategy on the cloud server, guided by an adaptive weight to navigate the critical trade-off between communication efficiency and model accuracy. This is complemented by an intra-group hierarchical aggregation mechanism on edge servers, designed to fully leverage the optimized groupings for local energy savings. This integrated approach allows HFELAG to achieve dynamic joint optimization between device communication energy consumption and global model accuracy. Extensive experiments show that HFELAG significantly reduces communication energy consumption while improving model accuracy compared to five strong baselines, demonstrating a practical pathway to bolster HFL’s security through energy-aware optimizations.

With the advancement of 6G, future mobile networks are witnessing a seamless integration between terrestrial and non-terrestrial networks to establish a global air-space-ground-maritime network. The dynamic channel and vast geographic coverage in these networks require efficient transmission of the multi-modal data. Meanwhile, semantic communication is a promising solution for this issue since it focuses on meaning rather than symbol-level reconstruction. However, due to wireless transmission and additional vulnerabilities introduced by semantic processing, these networks face critical security challenges and open issues in both conventional communication and the semantic level. Therefore, we propose a secure multimodal semantic communication framework leveraging multimodal large language models (MLLMs) to address these challenges. We provide a systematic analysis of three security aspects: security in system and model, security in data and semantics, and security in communication and transmission. In addition, we have developed a novel framework that integrates semantic representation with joint source-channel coding (JSCC) for semantic information transmission. Then, a comprehensive protection mechanism is proposed that jointly optimizes security at both physical and semantic layers through artificial noise. Finally, experimental results demonstrate that our approach enhances the efficiency and security of communication. It shows that the achievable secrecy rate is improved by up to 40 percent while maintaining a gap exceeding 10 dB between the legitimate receiver and eavesdropper.

Large models (LMs) exhibit transformative power in 6G Mobile-Edge Intelligence (MEI) scenarios, which can effectively address key challenges regarding network complexity, resource optimization, intelligent decision-making, and user experience. In this article, we first introduce the basics of LM and explain the potential applications and existing solutions for integrating LMs into 6G MEI. Afterward, we conduct a case study oriented to LM-assisted 6G MEI scenarios, focusing on how to use LMs to predict the size of traffic. Specifically, we employ two attention modules named Dartboard Spatial Multi-head Self-Attention (DS-MSA) and Causal Temporal Multi-head Self-Attention (CT-MSA) to extract spatio-temporal features. A Large Language Model (LLM) named GPT-2 is utilized as the backbone model. To fit time-series data, the original token encoding layer is replaced by the convolutional encoding layer. Additionally, to handle multi-scale temporal information, we adopt an additional temporal encoding layer. During training, the majority of the pre-trained parameters are frozen, with only the input/output layers and the fully connected layer involved in parameter updates. The simulation results demonstrate that, compared to the traditional AI-based baselines, the proposed LM-based framework acquires a lower training loss and converges faster. Moreover, it achieves better prediction accuracy with less resource consumption.

The future communication networks refers to large-scale mass-connected networks consisting of billions of cloud, edge, and end devices, which are expected to support the ever-growing communication demands. In the future communication networks, billions of end devices generate massive amount of data that needs to be processed and analyzed (e.g., model training). Artificial Intelligence of Things (AIoT) is a groundbreaking technology that leverages artificial intelligence models to process and analyze data generated by a large number of internet of things devices. As an emerging AIoT method, personalized Federated Learning (pFL) has emerged prominently in distributed model training using massive data from the future communication networks. However, it is challenging to accomplish high-performance and communication-efficient model training by existing pFL methods in the future communication networks, due to the following limitations. a) Dynamic role differences in each layer of a multi-layer model are neglected, leading to poor accuracy in customized models deployed on end devices. b) Owing to numerous end devices in the future communication networks, the communication frequency between a cloud server and end devices is extremely high in each communication round, resulting in expensive communication cost. To solve these two limitations, this paper presents a novel pFL framework for distributed model training in the future communication networks, called Adaptive Layer-wise personalized Federated Learning via Dual Delay Update (ALpFLDDU). First, in end devices, a layer-wise aggregation scheme based on an adaptive weight calculation mechanism is designed to capture the dynamic role differences of model layers. Second, in each communication round, we develop a dual delay update strategy to reduce communication frequency between a cloud server and end devices while ensuring model performance. Simulation experiments on text and image classification datasets are conducted. The experimental results show that ALpFLDDU realizes higher classification precision and lower communication cost than advanced pFL benchmarks on various classification tasks.

Autonomous Aerial Vehicle (AAV)-assisted Internet of Things (IoT) represents a collaborative architecture in which AAV allocate resources over 6G links to jointly enhance user-intent interpretation and overall network performance. Owing to this mutual dependence, improvements in intent inference and policy decisions on one component reinforce the efficiency of others, making highly reliable intent prediction and low-latency action execution essential. Although numerous approaches can model intent relationships, they encounter severe obstacles when scaling to high-dimensional action sequences and managing intensive on-board computation. We propose an Intent-Driven Framework for Autonomous Network Optimization comprising prediction and decision modules. First, implicit intent modeling is adopted to mitigate inaccuracies arising from ambiguous user expressions. For prediction, we introduce Hyperdimensional Transformer (HDT), which embeds data into a Hyperdimensional space via Hyperdimensional vector encoding and replaces standard matrix and attention operations with symbolic Hyperdimensional computations. For decision-making, where AAV must respond to user intent while planning trajectories, we design Double Actions based Multi-Agent Proximal Policy Optimization (DA-MAPPO). Building upon MAPPO, it samples actions through two independently parameterized networks and cascades the user-intent network into the trajectory network to maintain action dependencies. We evaluate our framework on a real IoT action dataset with authentic wireless data. Experimental results demonstrate that HDT and DA-MAPPO achieve superior performance across diverse scenarios.

With the rapid advancement of artificial intelligence (AI) technology, the Internet of Vehicles (IoV) is becoming increasingly important in intelligent transportation systems (ITSs). At the same time, large language models (LLMs) and generative AI (GenAI) are gradually playing significant roles in the IoV. Multimodal LLMs (MLLMs) are models capable of integrating multiple modalities of information, enhancing the environmental perception capabilities of the IoV. GenAI can generate highly complex virtual driving scenarios, which are used to test and optimize intelligent driving algorithms, reducing the risks and costs of real-world road testing. Embodied AI can interact with the environment and make decisions, and can learn in the simulated environments that are provided by GenAI. To address issues like low correlation of multimodal data, high-uncertainty of driving environment, and insufficient intelligence in the IoV, this article proposes a perception and decision-making system empowered by embodied AI. First, we utilize a transformer to achieve multimodal data fusion integrated with vehicle-to-vehicle (V2V) data. Second, the MLLM analyzes the situation based on driving intentions, communication conditions, and the perceived environmental expressions. Finally, the MLLM makes the optimal decisions and adjustments based on the analysis. The experimental results show that the integration of V2V enables the IoV to more comprehensively perceive a complex driving environment. The MLLM can empower the IoV to make accurate decisions in high-uncertainty scenarios, significantly improving route completion rates.

Cyber-physical social systems (CPSS) integrate computation, communication, perception, and actuation capabilities, enabling dynamic interaction with humans. One building block for implementing CPSS is group conversations, which refer to multiparty interactions where participants dynamically exchange information and respond to others' inputs while fulfilling different conversational roles. Such group conversations are characterized by the interconnected but unique nature of each participant. To enhance the CPSS's ability to select an addressee within group conversations, we propose the speaker interaction self-attentional recurrent neural network (SISRNN). This innovative model is capable of accurately predicting the goals of group interactions. Specifically, we improve recurrent neural networks (RNNs) with gated recurrent units (GRUs) to optimize performance. To capture dynamic shifts in conversational roles (i.e., senders, responders, and onlookers), we implement three distinct GRU units to monitor these roles throughout conversation turns. An attention mechanism is then integrated to prioritize key information within the conversation, improving predictive accuracy during GRU updates by producing an augmented hidden state. Leveraging the nuances of group conversations, we adopt a distributed learning framework for network training, utilizing the parameter server architecture for efficient parameter dissemination and synchronization across diverse settings. Our analysis encompasses a wide range of datasets, evaluating five proposals against three criteria: conversation context length, total conversation participants, and responder count per turn. The SISRNN excels in predictive accuracy in these metrics.

Accurate PM25 prediction plays an important role in climate change mitigation and environmental protection. As an emerging artificial intelligence technique, Large Language Models (LLMs) have exhibited powerful data processing and adaptive feature learning capabilities, thus being widely applied in Time Series Prediction (TSP). Unfortunately, the current LLM-based TSP models pose difficulties in accurate PM25 prediction due to two factors: 1) they neglect or fail to fully extract spatio-temporal dependencies, and 2) these LLM-based TSP models solely rely on the supervised training, failing to learn real data's distribution. Jointly considering these factors, in this paper, we present a novel PM25 prediction framework namely Spatio-Temporal LLM Generative Adversarial Network (STLLM-GAN). In detail, to capture spatio-temporal dependencies, Spatio-Temporal Large Language Model (STLLM) is first developed, containing a Spatio-Temporal Module (STM) and an LLM-enabled Inference Module (LLMIM). For the purpose of optimizing STLLM's training, we design an adversarial training scheme using Generative Adversarial Network. The scheme incorporates un-and supervised training through a jointly beneficial manner. The unsupervised training seeks to learn real data's distribution, while the supervised training strives to align the actual values with estimations by minimizing Mean Squared Error (MSE) function. We carry out extensive experiments on two real-world air pollutant concentration datasets, covering Shanghai city and Beijing city, respectively. The experimental results prove that STLLM-GAN is superior to advanced benchmarks in prediction performance.