The problem of training federated learning(FL) algorithms which is used to perform intrusion detection and protect data privacy in actual wireless networks is studied. Specifically, unmanned aerial vehicles(UAVs) can move at high speeds that results the doppler effect, which further deteriorates the communicational quality of the wireless network interactions with the base station. Since all training parameters are transmitted through the wireless network, the quality of training is affected by many factors such as packet errors and the number of wireless resources. The federated learning and wireless communication is formulated as an optimization problem, whose goal is to minimize the FL loss function that captures the performance of the algorithm. To solve the problems, we present the PVBO: transmit power, velocity and resource block allocation joint optimization algorithm. Based on the expected convergence rate of the FL algorithm, the optimal transmit power and velocity for each UAV are derived, under a given uplink resource block allocation scheme. Then, the uplink resource block allocation is optimized so as to minimize the FL loss function. The simulation results show that it can averagely improve the accuracy of intrusion detection by 4.53 % compared to the randomized policy.
Semantic communication, as a next-generation communication technology, has proven to be highly effective in different tasks. However, the inherently open nature of wireless channels makes semantic communication systems particularly vulnerable to security risks, with eavesdropping being a primary concern. Despite these risks, security issues in semantic communication have not received sufficient attention. To fill this gap, we propose a novel Proactive Security Mechanism (PSM) within the semantic communication network to prevent eavesdroppers from obtaining private data. The main idea of the PSM is to ensure that legitimate users can communicate normally while deceiving eavesdroppers without arousing suspicion. To achieve this, the PSM actively adds controlled perturbations into the transmitted semantic information. To ensure that legitimate receivers can accurately eliminate these perturbations and maintain communication quality, we introduce Communication Identification Codes (CICs) in the PSM, which are assigned by the edge servers of the network. Furthermore, we build a semantic communication system based on the generative model and implement the PSM to verify the effectiveness of the PSM. The simulation results demonstrate that the PSM effectively ensures the privacy and confidentiality of data, thereby safeguarding the security of semantic communication.
This study aims to identify the most efficient and cost-effective approach to addressing the computational challenges associated with EEG-based biometric systems that use machine learning models. Traditional brain signal biometric analysis methods often extract features from a single EEG channel or a randomly selected subset of channels from the same brain region. However, given that essential physiological information is functionally distributed across multiple brain regions, this study systematically explores various combinations of EEG channels both within and across different brain regions. The biometric distinctiveness of EEG signals is evaluated under different emotional simulation conditions. Addressing the channel selection problem is critical, as it can lead to improved identification of optimal EEG sensor locations. EEG data were collected using all 32 available channels and the experimental results revealed that an optimally selected subset of eight channels outperformed the complete set of 32 channels in classification performance.
Backlit photography encounters challenges such as poor contrast, reduced visibility, and high noise levels, which adversely impact both human perception and computer vision systems. Current deep learning-based enhancement methods typically rely on paired images for training, yet the difficulty in acquiring such pairs gives rise to authenticity issues and artifacts in the enhanced images. Moreover, defining a standard for the ideal enhanced image is problematic due to the inherently subjective nature of lighting conditions. Although previous methods have successfully addressed unsupervised backlit image enhancement, the training process remains complex, with significant potential for improvement in both methodology and performance. In this regard, we propose a novel unsupervised approach that leverages the visual-semantic priors of the CLIP model to guide the enhancement process. Specifically, we begin by fine-tuning the CLIP image encoder using unpaired image data, refining its visual priors for backlit image enhancement. Subsequently, we achieve effective unsupervised backlit image enhancement through the use of the specialized CLIP image encoder tailored for this task, coupled with the CLIP text encoder, which provides integrated visual and semantic supervision. Our approach outperforms existing methods on the BAID dataset across multiple quality assessment metrics, establishing a new state-of-the-art for unsupervised backlit image enhancement.
Reconfigurable Intelligent Surface (RIS) has emerged as a promising enabling technology for wireless communications, which significantly enhances system performance through real-time manipulation of electromagnetic wave reflection characteristics. In RIS-assisted communication systems, existing deep learning-based channel state information (CSI) feedback methods often suffer from excessive parameter requirements and high computational complexity. To address this challenge, this paper proposes LwCSI-Net, a lightweight autoencoder network specifically designed for RIS-assisted multiple-input single-output (MISO) systems, aiming to achieve efficient and low-complexity CSI feedback. The core contribution of this work lies in an innovative lightweight feedback architecture that deeply integrates multi-layer convolutional neural networks (CNNs) with attention mechanisms. Specifically, the network employs 1D convolutional operations with unidirectional kernel sliding, which effectively reduces trainable parameters while maintaining robust feature-extraction capabilities. Furthermore, by incorporating an efficient channel attention (ECA) mechanism, the model dynamically allocates weights to different feature channels, thereby enhancing the capture of critical features. This approach not only improves network representational efficiency but also reduces redundant computations, leading to optimized computational complexity. Additionally, the proposed cross-channel residual block (CRBlock) establishes inter-channel information-exchange paths, strengthening feature fusion and ensuring outstanding stability and robustness under high compression ratio (CR) conditions. Our experimental results show that for CRs of 16, 32, and 64, LwCSI-Net significantly improves CSI reconstruction performance while maintaining fewer parameters and lower computational complexity, achieving an average complexity reduction of 35.63% compared to state-of-the-art (SOTA) CSI feedback autoencoder architectures.
In aerial images of prairies, rat holes occupy an extremely small proportion, thus belonging to the small-target detection task. Rat holes blend so well with the surrounding environment that information like their texture and boundaries is often obscured by the complex background. As a result, rat-hole detection can be classified as small and weak target detection. Additionally, the frequently - occurring ground shadows in the prairie environment interfere with rat - hole detection. This interference tends to generate pre - selected bounding boxes with high Intersection over Union (IoU) values during the small and weak target detection process, leading to a high false - detection rate. In this paper, we distinguish shadows by exploring the depth - of - field characteristics of rat holes. We enhance the anti - interference ability of positive samples by fusing the depth - information channel with the RGB channel. Moreover, we improve the backbone network of Faster R - CNN to further extract detailed features. We selected rat holes in two prairie regions of Inner Mongolia with frequent rodent - infestation as the research subjects and constructed an aerial - photography dataset of prairie rat holes named RatHoles. Through in - depth analysis of the target features in the rat - hole dataset, we proposed a small and weak target detection model for prairie rat holes, DEFA, which integrates depth - estimation information. The Average Precision (AP) value of DEFA is 4.2
Driven by technological innovation and modern industrial transformation, the Industrial Internet of Things (IIoT) has emerged as a pivotal force in promoting the new intelligent manufacturing production paradigm. Within the paradigm, production tasks generated from heterogeneous manufacturing devices (MDs) possess intricate dependency relationships, encompassing internal dependencies (IDs) based on directed acyclic graphs (DAGs) and external dependencies (EDs) among MDs. Therefore, accurately characterizing such dependencies is crucial for reflecting the realities of modern industrial production. However, there is currently a lack of precise representations of the dependencies, as existing studies typically focus solely on IDs, failing to present the EDs that interconnect different MDs. To bridge this gap, this paper innovatively designs the multi-coupled DAGs model capable of accommodating various task dependency relationships. Moreover, by considering the aspects of dependent task offloading and transmission power allocation for data in EDs, a joint optimization problem is formulated with the objective of minimizing the system’s Energy-Time Cost (ETC). Finally, to solve this problem, we design the proximal policy optimization (PPO) based multi-coupled directed acyclic graphs for resource allocation and tasks offloading (MDRAO) algorithm. Simulation results indicate that the MDRAO algorithm outperforms its counterparts in terms of convergence performance and adaptability across various scenarios.
To improve the throughput of Ad Hoc such as Wireless Sensor Networks (WSN) and Internet of Vehicles (IoV), we proposed a Dynamic Frame Slotted ALOHA (DFSA). DFSA combines FSA (Frame Slotted ALOHA) with Q-learning to achieve an optimal way to select time slot and dynamically change the frame length. The Q-learning framework is implemented to dynamically optimize slot Q-values through feedback mechanisms and memory retention systems. Agents maintain two critical memory states: 1) the cumulative count of sequential collision or successful transmission events within the current slot, and 2) the temporal distribution pattern of idle slots observed in the preceding frame. To expedite convergence efficiency, a truncated binary exponential increment algorithm is integrated into the Q-value update process. The simulation results show that the average convergence time and collision number of this algorithm are significantly lower than other ALOHA algorithms, the throughput are higher than others.
Federated learning (FL) enables distributed devices to collaboratively train a machine learning model while keeping their sensitive data private, but it remains vulnerable to Byzantine attacks and privacy threats. While differential privacy (DP) has emerged as a promising technique for safeguarding privacy in FL, its noise introduction weakens defenses against Byzantine attacks, making FL susceptible to adversarial manipulation. Existing solutions attempt to address both privacy and security concerns but often involve modifying DP mechanisms or imposing restrictive assumptions, limiting their practical applicability. To overcome this challenge, we introduce DGShield, a dual-phase group-wise aggregation approach designed to address the dilemma at the group level. First, we propose a deviation-based group formulation method that groups updates based on magnitude deviation. Building on this, we design a group-wise similarity filtering mechanism to reduce noise by leveraging group-level metrics, thereby providing defense against Byzantine attacks at both the magnitude and direction of group-level updates. Experimental results demonstrate that DGShield effectively defends against Byzantine attacks in FL with DP, while also improving accuracy by up to 13
Reliability issues arising from random failures and malicious attacks in the Internet of Things (IoT) are increasingly concerning, making it essential to enhance node failure tolerance for system stability. Most existing studies employ centralized optimization methods to reconfigure network communication link layouts for improved topology robustness. However, these methods depend on global information, pose challenges in distributed IoT scenarios, and often fail to balance computational resources with robustness gains. To address these limitations, we propose a Distributed Robustness Optimization scheme for IoT Topology (DROIT) based on local information. Our scheme employs a multi-agent graph reinforcement learning approach to achieve distributed topology optimization through local information collection and collaborative decision-making among agents. Specifically, agents utilize local observation data to determine optimal actions based on the graph reinforcement learning module and are guided to select actions that enhance overall robustness through a decentralized heuristic reward mechanism. Experimental results indicate that while DROIT is marginally less effective than advanced centralized algorithms in improving topology robustness, it achieves 10 to 100 times more efficient robustness enhancement per unit time compared to other algorithms.
In response to challenges such as relation overlap complexity, semantic understanding difficulty, and class imbalance in relation extraction, this work introduces the adaptive feature fusion enhanced cascade pointer network (AFFCPN). The method leverages the Chinese pre-trained language model RoBERTa (Robustly Optimized Bidirectional Encoder Representation from Transformers) to enhance deep semantic understanding in Chinese text with lengthy and intricate structures. It also utilizes entity boundary information from the named entity recognition model to precisely identify entities. By employing the adaptive feature fusion module, it extracts crucial features from the start and end positions of subject entities, effectively utilizing key information to improve judgment regarding object entity boundaries. Moreover, a cascade pointer network structure is adopted to manage overlapping relation structures efficiently, thereby enhancing the model’s capability to extract complex overlapping relations. Additionally, the model incorporates the Focal Loss function to enhance extraction performance on minority class relations by adjusting class weights more directly and effectively. Experimental results across three diverse datasets showcase that the model’s performance surpasses that of existing mainstream relation extraction models. This not only validates its effectiveness in handling overlapping and imbalanced relations, but also demonstrates its applicability in general domain relation extraction tasks.
The graph neural network model has a wide range of application value in the intelligent transportation system. However, the data design pattern based on graph structure cannot solve the directional and hierarchical problems of node information transmission. This makes the deep models represented by graph convolutional networks lack a certain predictive ability in node distribution scenarios. In this study, a tree spatial-temporal model with tree structure as the sample space is designed for the traffic node distribution scenarios. Firstly, road nodes and spatial relationships are abstracted according to the graph structure, so as to realize the preliminary spatial distribution relationship of nodes. Secondly, different nodes are used as the root nodes of the tree to construct the plane tree structure and plane tree matrix to complete the conversion process from the graph structure to the tree structure. Finally, the plane tree matrices of all nodes are fused into a spatial tree matrix representing the spatial global relationship of the nodes. This study designs the deep tree traffic forecast model based on tree structure, which converts the graph structure of small-scale aggregated nodes into tree structure. The deep tree traffic forecast model realizes the mining and prediction tasks of various traffic measurements based on the spatial tree convolution module and the temporal convolution module. This study demonstrates the excellent predictive ability of the deep tree traffic forecast model in traffic node distribution scenarios by comparing with multiple existing baselines on real datasets.
The robustness of a network reflects its ability to maintain functionality in the face of attacks or failures. However, current methods for calculating robustness metrics rely on simulating attacks, which is computationally complex and severely time-consuming, resulting in poor scalability. To address this problem, this paper proposes a topology robustness prediction method based on hierarchical feature fusion and progressive learning, which utilizes the multi-scale features of the network structure to achieve an efficient and accurate evaluation of robustness. Specifically, we first design a node multi-scale feature extraction module to comprehensively capture the multi-level structural characteristics of the network topology; second, we propose a progressive input strategy, which enables the model to focus on different levels of structural information in phases by progressively introducing features of different scales, thus improving the prediction accuracy of the model. Finally, the model is extensively evaluated by testing it on several different types of networks as well as real-world networks. Experimental results show that we significantly improve the robustness computational efficiency compared to the simulated attack method, while outperforming other methods in robustness prediction performance.
In industrial IoT, multi-motor control systems are crucial for meeting the complex demands of modern manufacturing. To guarantee the stability of control performance while minimizing the induced network costs, e.g., energy consumption, the co-design of control and network systems is essential. Moreover, multi-motor systems face the challenge of torque synchronization, where the torque deviation among motors can significantly affect system and control error, potentially leading to the damage of industrial IoT equipment. In this paper, we study the networked multi-motor control with torque synchronization in industrial IoT under stringent energy consumption requirements. We formulate a joint optimization problem to minimize the overall system costs including both the control error and energy consumption, and then further propose a novel hierarchical deep reinforcement learning (DRL) based control algorithm, namely, HDRL-NMC-TS, which solves the problem. Specifically, the optimization problem is decoupled into two sub-problems, each sub-problem is formulated as a Markov decision process, and respectively solved by a designed HDRL-based algorithm, for optimizing the sampling period and the voltage compensation value. Extensive simulations show our HDRL-NMC-TS algorithm is effective and outperforms counterparts.
Non-safety applications such as autonomous driving and crowd sensing require substantial data transmission, exacerbating the scarcity of spectrum resources in the Internet of Vehicles (IoV). To mitigate this issue, cognitive radio-assisted IoV (CIoV) has attracted increasing attention. Despite extensive research addressing various challenges in CIoV networks, securing additional spectrum opportunities continues to be a key obstacle. This paper explores whether the signal-blocking effect of viaducts can provide additional spectrum opportunities for vehicular secondary users (SUs) in CIoV networks. First, we classify primary user (PU) regions into three types based on signal reception characteristics and introduce a novel parametric geometric model to estimate the geometric boundaries of each region. Next, we use a dual-path propagation model to figure out the reflection and penetration coefficients for both vertical and horizontal polarizations. This helps us figure out how much interference power there is between the SU transmitter (SU-Tx) and the PU receiver (PU-Rx). Finally, we use real-world parameters to conduct two case studies, showing that SUs’ transmissions do not cause harmful interference to PUs when operating at specific heights. We also provide a comprehensive analysis of how various parameters influence the interference from SU-Tx to PU-Rx, using detailed numerical results.
As a low-latency real-time communication technology, WebRTC is widely used in real-time applications. However, its default congestion control algorithm, GCC, struggles to handle complex network fluctuations, responding slowly to bandwidth changes and adjusting the bitrate inefficiently, which limits high-quality communication. Many improved algorithms focus on increasing bitrate but overlook other factors that affect users’ Quality of Experience (QoE), such as latency and frame rate. To address this issue, we propose LST M and R einforcement Learning for C ongestion C ontrol, which optimizes throughput while jointly considering latency and frame rate. MRCC consists of a Throughput Trend Predictor that forecasts bandwidth variations and a Learning-Based Bitrate Adjuster that dynamically modifies the bitrate based on historical network conditions. This design enhances adaptability, reduces transmission delays, stabilizes frame rates, and improves overall QoE, providing an optimized solution for real-time communication scenarios.
Adapting general-purpose large language models (LLMs) to domain-specific tasks remains challenging, especially under limited computational resources and strict privacy constraints. In-context learning (ICL) offers a lightweight adaptation method but suffers from sensitivity to example selection and limited generalization. A promising strategy is to introduce small auxiliary models that assess task relevance and enhance ICL inputs. However, privacy concerns and non-IID data distributions hinder centralized training of such models. To address these challenges, this paper presents FedBridgeICL, a framework that leverages federated learning (FL) and in-context learning (ICL) to bridge small language models (SLMs) and LLMs, thereby enhancing domain adaptation under privacy constraints. The framework introduces three mechanisms: (1) federated SLM fine-tuning for global feature extraction; (2) embedding SLM predictions with confidence scores into LLM prompts to enable privacy-aware knowledge transfer; (3) dynamic confidence-based reasoning for optimized LLM decisions. Evaluations on the GLUE benchmark demonstrate FedBridgeICL’s consistent outperformance over conventional ICL and centralized baselines in heterogeneous federated settings, confirming its viability for privacy-sensitive deployments.
Clustered Federated Learning (CFL) effectively addresses the challenge of data heterogeneity in Federated Learning (FL), where clients often hold Non-IID (Non-Independent and Non-Identically Distributed) data, typically limited to a few categories. However, the updates of the cluster models in CFL inadvertently expose additional information, rendering it vulnerable to Category Inference Attack (CIA), where the attacker exploits this exposure to infer sensitive category information from these updates. In our experiments on the image classification datasets, the attacker consistently achieves the F1-score exceeding 90% across various scenarios, highlighting CFL’s vulnerability to CIA and the urgent need for robust privacy protections. To defend against this attack, we propose an adaptive local differential privacy (LDP) strategy for CFL, named AFC-CFL (Adaptive Fisher Information and Dynamic Clipping Threshold in CFL). AFC-CFL adopts adaptive Fisher information to adjust the privacy budget and dynamically modifies the clipping threshold during model training, mitigating the noise’s effect on model performance while ensuring strong privacy protection. Experiments demonstrate that AFC-CFL significantly reduces the impact of noise on model accuracy, achieving a maximum accuracy improvement of 32.8% compared to common LDP method. Additionally, AFC-CFL reduces the attacker’s F1-score by up to 24.3% , achieving a superior trade-off between model performance and privacy protection, making it highly suitable for deployment in privacy-sensitive CFL scenarios.
Gait phase classification plays a crucial role in diagnosing walking impairments and guiding rehabilitation training. In recent years, various deep learning models have been proposed to classify gait phases by automatically learning features from temporal data. Many of these models rely on recurrent neural networks (RNNs) to capture the temporal continuity inherent in gait data. However, existing methods often struggle to model long-term dependencies and the subtle phase transition features that are essential for capturing the dynamic variations of gait. To address these limitations, we propose a novel multilayer context network (MLCNet), which integrates a multi-head contextual learning module to extract long-term dependency features and a squeeze-and-excitation (SE) module to capture adjacent phase transition features. This hybrid design enhances the model's ability to represent the dynamic characteristics of gait. We evaluate the performance of the proposed model on the publicly available gait dataset GEDS. The experimental results show that our method achieves an overall accuracy of 89.6
Modern AI accelerators face significant challenges in balancing memory bandwidth, capacity, and cost requirements, particularly for large model inference tasks. Traditional solutions often rely on expensive high-bandwidth memory or struggle with limited memory capacity and bandwidth utilization. This paper presents a novel AI accelerator architecture that effectively addresses these challenges through a DDR-based approach. Our design features a hybrid multi-channel DDR memory system with dynamic interleaving modes, coupled with a high-performance Controller CPU for sophisticated task scheduling. Through careful hardware-software co-design, our architecture achieves efficient utilization of both memory bandwidth and compute resources while maintaining programming simplicity. The memory system supports flexible data movement patterns and enables efficient handling of diverse AI workloads. The results of silicon implementation demonstrate high performance in memory bandwidth utilization, computational efficiency, and model inference tasks, validating the effectiveness of our approach in providing high memory bandwidth and large memory capacity. Our work establishes a practical paradigm for designing efficient AI accelerators using DDR-based memory systems.
