Cyberattacks continue to pose significant threats to interconnected systems. While deep neural networks offer high accuracy for intrusion detection, their high computational cost often leads to latency and missed detections on resource-constrained edge devices. To address this challenge, we present IDNet, an Instance-adaptive Dynamic Network enhanced with adversarial training to jointly improve detection accuracy and inference efficiency. IDNet incorporates a semantic-aware traffic instantiation module that converts heterogeneous network flows into unified grayscale matrices, enabling consistent processing of traffic data with different feature dimensionalities and extraction methodologies. Based on this unified representation, adversarial training is used to stabilize feature extraction under instance-adaptive dynamic inference. A generator is constrained by structural similarity to ensure that synthetic samples preserve label fidelity and traffic-specific characteristics. The discriminator consists of a residual detection network and a lightweight policy network; the latter generates routing vectors to dynamically activate relevant paths on the former based on traffic instance complexity. We introduce a self-optimized pre-training strategy that combines adversarial and curriculum learning to initialize the detection and policy networks. An augmentation-enhanced joint fine-tuning strategy is further applied to align detection behavior with learned block-discarding patterns, with improved generalization on underrepresented attack classes. Extensive experiments across multiple authoritative, heterogeneous datasets show that IDNet significantly reduces inference costs compared to full-scale residual networks and conventional dynamic neural networks, while maintaining competitive detection accuracy against known and unknown attacks. Evaluation on resource-limited platforms further demonstrates IDNet’s practical efficiency, outperforming all baseline methods in inference speed.

This paper presents MobiFormer, a split-federated transfer learning framework with multi-head attention designed for distributed drone Radio Access Network (RAN) slicing. The objective is to optimize slice performance isolation and training costs. Based on a flexible service metric, the problem of maximizing slice performance isolation quality is formulated as a joint optimization of slice windowing and resource allocation. For single-drone autonomous operations, we construct an "unconstrained mobility and sustainable fine-tuning" paradigm, enabling drones to adapt previously trained resource slicing models to new environments with the assistance of multiple target-domain terrestrial Base Stations (BSs). This adaptation is facilitated by a Source-free Multi-target-domain Transfer Learning (SMTL) approach, where Transformer-based multi-head attention is employed on drones to integrate fine-tuned models from multiple target-domain BSs. Building on SMTL and continuing its scenario, a Clustered Split-Federated Learning (CSFL) approach is developed to support multi-drone collaborative training, where BSs serve as cluster heads to aggregate parameters from member drones. To save energy, part of the onboard models are migrated to BSs while local iterations occur through gradient exchanges. Unlike SMTL, the Transformer encoder-decoder is deployed at BSs to enhance the global model's adaptability and generalization. Extensive simulations demonstrate that MobiFormer outperforms benchmark approaches in terms of performance isolation, energy consumption, and online decision-making efficiency in distributed learning settings.

Consumer drones are increasingly used for crowd-counting in complex environments; however, their deployment faces challenges from adverse external conditions such as low illumination and inclement weather, as well as inherent limitations like constrained onboard computational resources. To address these constraints, we present MA-DyNN (Modal-Adaptive Dynamic Neural Network), a lightweight and robust framework that dynamically adapts to varying modality conditions for accurate crowd counting. This framework employs an efficient single-stream architecture with specialized modal extractors to capture and integrate complementary information from both visible and thermal infrared (TIR) inputs. Based on the extracted modal features, we design a modality-adaptive gating mechanism to dynamically select the optimal modality based on environmental conditions, favoring visible imagery for inference efficiency in well-lit scenarios and leveraging TIR as auxiliary support under low-light or degraded conditions. To enhance robustness against sensor failure or missing modalities, we develop a density-aware modality converter that adds crowd density constraints to a cycle-consistent generative adversarial learning framework to generate high-fidelity TIR images. This enables consistent performance by aligning synthetic and real TIR-based counting outcomes through adversarial learning. Extensive experiments on DroneRGBT and RGBT datasets show that MA-DyNN achieves superior accuracy, generalization, and real-time performance compared to state-of-the-art multimodal baselines. Its inference acceleration performance approaches single-modality models without compromising the accuracy gains provided by multimodal learning

The 5th Generation (5G) Mobile Network, coupled with the Internet of Things (IoT), is a heterogeneous environment prone to various security vulnerabilities and frequent attacks. Therefore, analyzing attackers’ intrusion intentions and penetration behaviors is crucial for guiding network security defenses. However, existing penetration path constructions mostly rely on a single agent or consider a single type of network, which cannot comprehensively assess the impact of group attack behaviors and system vulnerability combinations on security. To address this issue, a multi-agent reinforcement learning approach is presented for collaborative path penetration. This solution leverages network situational information to guide policy learning, improving the overall path penetration quality. Agents adopt a hierarchical structure of primary and subordinate roles. The primary agent can observe the entire environmental state and formulate top-level collaborative strategies for the group, while subordinate agents learn individual policies based on assigned tasks and team rewards and interact with the environment to learn. Through this mode, agents coordinate with each other to select and continuously improve their strategies in penetration. Experimental results show that compared with single-path penetration under a single agent, the proposed method generates multiple penetration paths with higher attack efficiency, decision stability, and task completion rate.

Multi-object detection in autonomous driving faces challenges due to multi-scale entities, diverse streetscapes, and limited computational resources. To address these challenges, we present MT-DyNN, a Multi-Teacher knowledge-distilled Dynamic Neural Network framework for instance-adaptive detection, optimizing detection accuracy and inference cost in autonomous driving. The framework’s student network comprises a customizable multi-branch residual detection network and a lightweight policy network. The former efficiently extracts multi-scale features in parallel without altering receptive fields, while the latter, depending on curriculum learning, captures task-relevant features and dynamically generates routing vectors to guide the activation or deactivation of residual blocks according to image instance complexity. The framework’s teacher network employs a soft-voting strategy to consolidate knowledge from multiple pre-trained teacher models, providing consistent guidance to the student. Within this distillation paradigm, the policy network’s routing search space is gradually refined, and the policy and detection networks are jointly fine-tuned to optimize the alignment between routing decisions and feature extraction. Experimental results on CIFAR and ImageNet demonstrate that compared to early exiting and stochastic depth methods, MT-DyNN achieves higher accuracy at the same inference cost and reduces the cost by 50% and 59% at comparable accuracy levels. The generated routing maintains channel sparsity across diverse scenarios.

Most existing user-collaborative privacy protection mechanisms in location-based services assume that collaborative group members are trustworthy and can strictly enforce collaboration rules. Such assumptions do not match reality and reduce the usability of the schemes. In this paper, a Monero covert communication-enhanced collaborative privacy-preserving framework is presented for mobile query services. Built on blockchain networks, this solution enables secure in-group interaction and behavior collection, enabling group members without trust to collaborate efficiently. A lightweight key negotiation is developed to support in-group Monero covert communication on blockchain networks to prevent in-group eavesdropping and selfish behavior. Blockchainassisted behavior collection and evidence recording ensure efficient and secure execution of in-group collaboration. Security analysis and simulation results demonstrate the proposed scheme achieves high security and privacy levels with low communication and computational costs in queries.

In this paper, a blockchain-assisted cross-silo graph federated learning (B-CGFL) framework is presented for large-scale network intrusion detection, aiming to break down barriers among different organizations and achieve a secure and transparent multi-party collaboration ecosystem. The network scenario is divided into multiple regions. Organizations in each region leverage graph neural networks to analyze local network flow topology information and identify traffic types accurately. With cross-silo graph federated learning, coordinators and organizations collaboratively complete the training and updating of the global model. Multiple coordinators jointly maintain a chain for model storage security. Oracles bridge the off-chain data provider and on-chain smart contracts, enabling Secure and trusted data access in off-chain model testing. For fair competition, a reputation-aware incentive mechanism is designed to boost global model quality. Security analysis confirms that B-CGFL can defend against model plagiarism and tampering with model test results. Experiments on three challenging datasets ToN-IoT, CSE-CIC-IDS2018, and BoT-IoT demonstrate that compared with benchmark methods, B-CGFL exhibits superior performance.

Vehicle-to-vehicle (V2V) video content sharing is promising for connected and autonomous vehicles. However, various security, trust, and privacy risks, coupled with conflicts of interest among different operators, hinder the large-scale promotion of such an application. To address these issues, a consortium blockchain-based framework for cross-operator V2V video content distribution is presented, breaking operators’ barriers and achieving a fair value transfer. First, a two-tier consortium blockchain architecture is designed for cross-operator transaction data management. Ledger maintenance and consensus verification are performed within subchains, while user information storage and smart contract execution are conducted between the mainchain and subchains. A multi-constrained vehicle-group selection algorithm is then designed, which minimizes service fees under vehicle reputation, connection duration, and transmission rate constraints. Lastly, an incentive mechanism is developed to promote the rapid upload of transaction data for cross-operator contract invocation. Extensive security analysis and simulation results demonstrate the proposed scheme’s feasibility, superiority, and effectiveness.

A slicing-based collaborative task offloading framework for space-air-ground integrated vehicular networks is proposed in this study, which can provide differentiated quality-of-service (QoS) guarantees for task offloading for high-speed vehicles while maximizing the number of completed tasks. A service-oriented radio access network (RAN) slicing framework is presented that supports slicing window adaptation, spectrum and computing resource orchestration, and collaboration among heterogeneous base stations. Based on the queuing model, the collaborative decision-making of RAN slicing and task offloading is modeled as a problem of maximizing the number of long-term task completions, which consists of three subproblems-slicing window division, resource slicing, and task scheduling-which are solved by a multi-access edge computing (MEC)-enabled controller, forming a closed loop with the slicing window as the period. When a new slicing window arrives, the controller determines its duration according to task traffic fluctuations and allocates resources to RAN slices through an optimization method. A double deep Q-learning network (DDQN)-based algorithm is developed for scheduling workflow on small time scales within a slicing window. Simulation results demonstrate that the proposed scheme performs better than existing approaches in terms of adaptability, task completion rate, and control overhead.

In this paper, a live layered video multicast framework based on unmanned aerial vehicle (UAV) relays is presented for on-orthogonal multiple access (NOMA) networks, aiming to maximize the aggregated video reception quality for cell-edge users. A visualizable graph model is constructed to characterize the coupling of the deployment of UAVs and their association with multicast groups with interference cancellation. Based on this graph model, the problem of maximizing multicast group video reception quality is modeled as a clique-based nonlinear integer programming and decoupled into UAV placement and UAV-group association subproblem and subchannel allocation subproblem. For mathematical traceability, the former is transformed into a unique maximum weight clique problem under the available UAV number constraint, solved by an improved branch-and-bound algorithm. For the latter, a heuristic resource-matching strategy is designed to obtain a near-optimal subchannel allocation with low computational complexity. Simulation results demonstrate that the proposed scheme outperforms mainstream benchmarks in terms of aggregate peak signal-to-noise ratio (PSNR), spectrum utilization, network coverage, and adaptability.

Deep reinforcement learning (DRL) has become prevalent for decision-making task assignments in mobile crowdsensing (MCS). However, when facing sensing scenarios with varying numbers of workers or task attributes, existing DRL-based task assignment schemes fail to generate matching policies continuously and are susceptible to environmental fluctuations. To overcome these issues, a twin-delayed deep stochastic policy gradient (TDDS) approach is presented for balanced and low-latency MCS task decomposition and parallel subtask allocation. A masked attention mechanism is incorporated into the policy network to enable TDDS to adapt to task-attribute and subtask variations. To enhance environmental adaptability, an off-policy DRL algorithm incorporating experience replay is developed to eliminate sample correlation during training. Gumbel-Softmax sampling is integrated into the twin-delayed deep deterministic policy gradient (TD3) to support discrete action space decisions and a customized reward strategy to reduce task completion delay and balance workloads. Extensive simulation results confirm that the proposed scheme outperforms mainstream DRL baselines in terms of environmental adaptability, task completion delay, and workload balancing.

Most user-collaborative location privacy protection mechanisms assume that collaborative group members are trustworthy and can strictly enforce collaboration rules. Such assumptions do not match reality and reduce the usability of the schemes. In this paper, we propose Invisible Man, a blockchain-enabled framework for peer-to-peer collaborative privacy games for location-based services. The framework enables users to protect their privacy in extreme environments and allows members of a collaborative group without trust to collaborate efficiently. To defend against inference attacks, a user-collaborative privacy game model is constructed, and a dual-verification mechanism with Chainlink and Witnet oracles is developed to provide security guarantees for model generation. Members can conduct cooperative games under the guidance of the model. To realize secure and efficient in-group collaboration, a blockchain-based reward and punishment mechanism for collaboration is designed, integrating token incentives and a blacklisting mechanism to ensure the verifiability and audibility of user behaviors. Security analysis and extensive simulation results demonstrate that the proposed scheme achieves high security and privacy with low costs.

We propose an altitude-adaptive vehicle counting method with an attention mechanism and multiscale receptive fields that optimizes the measurement accuracy and inference latency of unmanned aerial vehicle (UAV) images. An attention mechanism is used to aggregate horizontal and vertical feature weights to enhance spatial information and suppress background noise. The UAV flight altitude and shooting depression angle are considered for scale division and image segmentation to avoid acquiring distance measurements. Based on the dilation rate, we introduce a receptive field selection strategy for the trained model to exhibit scale generalization without redundant calcula-tions. A distribution-aware block loss is optimized via k roots to balance the loss of sparse and crowded regions by dividing the density map. Experiments on three authoritative datasets demonstrate that compared with CSRNet, the proposed method improves the mean absolute error by 29.4%–54.0% and mean squared error by 28.6%–41.2% while reducing the inference latency. The pro-posed method exhibits higher counting accuracy than lightweight models including MCNN and MobileCount.

Vehicle-to-vehicle (V2V) video content dissemination is fundamental for many emerging services (e.g., on-road entertainment and roadside surveillance). However, many security, privacy, and trust issues hinder the large-scale promotion of this service. Malicious entities may spread incomplete and illegal content or defraud the requester to obtain rewards. Moreover, resource consumption and possible privacy leakage issues decrease the participation of many vehicle users. To address these issues, we propose a blockchain-based trust and scalable V2V video content dissemination framework to create an excellent content caching and sharing ecosystem with resistance to diverse attacks and adaptation to high mobility. Under a decentralized framework, a fair reward strategy is designed to price each video content transaction as a V2V service fee. A vehicle-group selection policy for many-to-one video content delivery is developed to minimize service fees subject to the constraints of the number of video layers, vehicle reputation, V2V connection duration, and achievable transmission rate. Smart contracts are designed to regulate transaction triggering, verification, and rewards, with a traceable punishment mechanism against malicious behaviors. We provide the proposed scheme’s security analysis and simulation results, demonstrating our scheme’s security, feasibility, and superiority.

In recent years, with the rapid development of various video and short video platforms, online video traffic accounts for an increasing proportion of the total global wireless network traffic.Mobile edge computing may cache the video content that users possibly request in the network edge in advance, thus effectively reducing the backbone network load and improving the response time of video services.On the other hand, the popularization and application of Unmanned Aerial Vehicle（UAV）technology has brought new opportunities for mobile edge computing applications.This paper, aiming at emerging application scenarios such as short video, proposes a video content caching strategy（DC-DRL）,which combines the prediction of video file content popularity with UAV scheduling, by multi-factor and multi-objective joint optimization to improve the average cache hit rate of the edge node cache system.In terms of cache replacement strategy, the use of long and short-term memory neural networks that are sensitive to timing features can perform file replacement based on popularity; In terms of the trajectory optimization of the UAV base station, our system model combines the actor-critic algorithm and uses the mobility of the UAV to assist in optimizing the average cache hit rate of the system.Finally, the performance of our strategy is analyzed and evaluated by a series of simulation experiments.

The drone-small-cell-assisted air-ground integrated network is a promising architecture for enabling diverse vehicle applications. This paper presents a joint resource slicing and vehicle association framework for drone-assisted vehicular networks, which facilitates spectrum sharing among heterogeneous base stations (BSs) and achieves dynamic resource provisioning in the presence of network load dynamics. We formulate the network utility maximization problem as mixed-integer nonlinear programming, considering traffic statistics, quality-of-service (QoS) constraints, varying vehicle locations, load conditions in each cell, and interdrone interference. The original maximization problem is transformed into a biconcave optimization problem to ensure mathematical tractability. An alternate concave search algorithm is then designed to iteratively solve vehicle association patterns and spectrum partitioning among heterogeneous BSs until convergence. Simulation results show that the proposed scheme achieves a significant performance improvement in throughput and spectrum utilization compared with two other baseline schemes.

A joint resource-optimization scheme is investigated for nonorthogonal multiple access (NOMA)-enhanced scalable video coding (SVC) multicast in unmanned aerial vehicle (UAV)-assisted radio-access networks (RANs). This scheme allows a ground base station and UAVs to simultaneously multicast successive video layers in SVC with successive interference cancellation in NOMA. A video quality-maximization problem is formulated as a mixed-integer nonlinear programming problem to determine the UAV deployment and association, RAN spectrum allocation for multicast groups, and UAV transmit power. The optimization problem is decoupled into the UAV deployment-association, spectrum-partition, and UAV transmit-power-control subproblems. A heuristic strategy is designed to determine the UAV deployment and association patterns. An upgraded knapsack algorithm is developed to solve spectrum partition, followed by fast UAV power fine-tuning to further boost the performance. The simulation results confirm that the proposed scheme improves the average peak signal-to-noise ratio, aggregate video-reception rate, and spectrum utilization over various baselines.

The fifth generation mobile communication system（5G） uses non-orthogonal multiple access（NOMA） technology for non-orthogonal multiplexing of wireless communication resources, which improves the spectrum utilization efficiency and system capacity by the way of overload.The NOMA system uses the grant-free mode to reduce the system flow and signaling overhead, but the receiver needs to perform multi-user detection.Based on the sparse characteristics of active users, the base station uses the compressed sensing（CS） reconstruction algorithm to recover the mixed sparse vectors of active users, and realizes efficient multi-user detection.The dense deployment of base stations in 5G network enhances the interferences among neighboring cells that increases the difficulty of CS-based detection and reduces the accuracy of detection.Aiming at the problem of interference in multi-user detection in the grant-free NOMA system, an anti-interference multiuser detection algorithm based on variable step size adaptive matching pursuit is proposed.Unknowing the sparse degree, the anti-interference active user detection can be realized by the adaptive variable step size way, in which the sparse degree is fast approached with large step size and accurately approximated with small step size.Simulation results show that, under different overload rates, the bit error rates of the proposed algorithm are lower than that of traditional multi-user detection algorithms based on OMP,gOMP and SAMP.

In consideration of the characteristics of multi-dimensional resource consumption fluctuating with time in the Internet of Vehicles system and users’ demands for efficient computing services and data privacy and security,this paper proposes a me-thod of multi-dimensional resource allocation for Internet of Vehicles based on federated learning.On the one hand,the allocation of computing,cache and bandwidth resources is considered comprehensively to ensure the completion rate of computing tasks and avoid the redundant allocation of multidimensional resources.For this purpose,a deep learning algorithm is designed to predict the consumption of various resources through the data collected by edge servers.On the other hand,considering the data island problem caused by users’ data privacy and security requirements,federated learning architecture is adopted to obtain a neural network model with better generalization.The proposed algorithm can not only adjust the allocation of multi-dimensional resources over time,but also meet the resource requirements that change over time,and ensure the efficient completion of computing tasks in the Internet of Vehicles system.Experimental results show that the algorithm has the characteristics of fast convergence and good model generalization,and can complete the aggregation of federated learning with fewer communication rounds.

5G车联网差异化的服务质量（Quality of Service，QoS）需求需要网络切片技术的辅助和多运营商协作提供资源。然而，多方参与下的车联网切片资源提供往往伴随着不信任、欺骗性攻击等问题。针对这些问题，本文提出一种基于区块链的5G车联网切片资源交易框架，目的是构建安全可信、可追溯的车联网资源交易生态。首先，设计一种基于智能合约的分布式资源交易方法，保证交易的公平性和隐私安全。然后，探索一种基于VCG拍卖的信誉评估机制，在资源交易者选择时作为参考。最后，开发一种信用感知的实用拜占庭式容错（Practical Byzantine Fault Tolerance，PBFT）共识机制。安全性分析和仿真结果表明，提出的方法可以实现安全可信的资源交易，提高节点运行效率，并且降低恶意节点参与共识的概率。