Edge-based Internet of Things (IoT) emerges to facilitate low-latency data sharing in a decentralized way. However, the intrinsic separation of data ownership and control inflicts serious illegal access and loss of trust on the data sharing. Although Ciphertext Policy Attribute-Based Encryption (CP-ABE) enables confidentiality and fine-grained access control, it is susceptible to impersonation attack from distributed and dynamic users, and its high overhead brings heavy burden to devices, which poses a huge challenge to its application in IoT data sharing. Moreover, most of existing schemes are insufficient to exert privacy-preserving access control on time-sensitive data. Furthermore, semi-trusted edge servers may compromise data integrity, which is seldomly considered. In this work, a Blockchain-assisted Revocable and Trustworthy Data Sharing (BRTDS) scheme is proposed for time sensitive data. It combines blockchain and zero knowledge proof with revocable CP-ABE to ensure distributed data access control across dynamic users and resist malicious users with impersonated identities and attributes. By designing a time-sensitive and privacy-preserving access policy tree, privacy-preserving access control is achieved. Furthermore, a distributed data auditing mechanism is devised through smart contracts to guarantee data integrity in access control process without a centralized auditor. Efficient encryption and verifiable data decryption is achieved with a better performance. Finally, the security of our proposal and its practicability and efficiency is demonstrated through extensive performance evaluation with simulations.
Security risks including false‑information dissemination were introduced with the rapid advancement of AI‑generated image technologies. Existing detection approaches were found to possess insufficient generalization capability under unseen‑generator scenarios. For detection methods relying on pre‑trained visual representations, a balance between semantic representation stability of real images and cross‑generator transferability of forgery features was hardly achieved in feature optimization. To tackle this problem, DeepfakeCLIP, an asymmetric semantic‑anchoring detection framework, was proposed in this work. The frozen real‑image encoding branch was adopted as a stable semantic reference. A gated bottleneck adapter was embedded only into the generated‑image branch, and multi‑template semantic fusion together with class‑specific calibration was leveraged to boost discrimination stability. Superior performance against competing methods on multiple benchmark datasets was observed, and detection generalization against unknown generators was effectively enhanced by the presented approach. This work confirms that DeepfakeCLIP improves the generalization performance for AI‑generated image detection under unseen generators.
Existing token-level multi-model parallel collaboration can reduce collaboration overhead through dynamic routing or participation control, but its execution stage is still constrained by token-by-token synchronization. To address this execution-efficiency bottleneck, we propose CrossFlow, an efficient token-level parallel collaborative inference mechanism. CrossFlow designs an asynchronous block-wise collaborative calibration algorithm to replace token-by-token synchronous waiting with base-model draft block generation and asynchronous calibration by collaborative models. It also introduces a KV-cache incremental reuse algorithm that re-encodes shared text, checks prefix consistency, and reuses historical KV caches within each collaborative model, transforming repeated full-prefix computation into incremental computation over newly added tokens. Experiments on multiple heterogeneous open-source model combinations and benchmark datasets show that CrossFlow maintains accuracy largely consistent with corresponding token-by-token synchronous parallel methods, while improving the average token generation rate to 1.10–2.88 times that of these methods.
To address the structural mismatch between strict operational constraints of the Industrial Internet and the conventional bolt-on, static-perimeter security paradigm, research on deterministic endogenous security defense was conducted. A deterministic endogenous security defense paradigm was proposed based on four categories of intrinsic determinism: entity, behavior, structure, and environment. Five key enabling technologies were systematically surveyed: endogenous identity management, task-behavior assurance, unknown-attack inference, active security defense, and security effectiveness evaluation. These were used to construct an endogenous security defense architecture covering prerequisites, system models, and four-dimensional technical implementation. The common bottlenecks of existing endogenous security practices in cross-path coordination and scenario adaptation were revealed through comparative analysis. Scenario analyses in intelligent manufacturing and industrial 5G illustrate that the proposed architecture can support the translation from theoretical mechanisms into engineering protection by leveraging intrinsic system determinism to counter threat uncertainty, thereby providing a viable path to addressing the cross-domain coordination and dynamic-evolution bottlenecks of industrial Internet security.
Automatic modulation classification, a core technique for electromagnetic spectrum sensing, serves as a key sensing task in applications such as spectrum monitoring, yet its adversarial vulnerability poses a severe security threat. To address the limitations that existing adversarial example detection methods only analyze static features while ignoring the dynamic cumulative effects of adversarial perturbations during forward propagation, an adversarial example detection method based on inter-layer activation evolution consistency was proposed. Activation representations from multiple key network layers were extracted, and autoencoders were trained to learn their activation manifolds. Distance matrices between samples and class-average activations were constructed in the reconstruction space, and adversarial examples were detected based on their inter-layer evolution consistency. Experimental results show that an average AUC exceeding 90% was achieved across two datasets, five typical attacks, and multiple signal-to-noise ratios, effectively enhancing the security of deep-learning-based spectrum sensing against adversarial attacks.
International video coding standards have historically undergone major iterations approximately every decade. The explosive growth of emerging applications, such as ultra-high-definition and immersive media, has significantly accelerated the pace of standardization. Joint Video Experts Team (JVET), the international video coding standard organization, continuously focuses on high-efficiency coding algorithms after the standardization of H.266/Versatile Video Coding (VVC). The research is conducted along two parallel tracks: Enhanced Compression Beyond VVC capability (Beyond VVC) and Neural Network-based Video Coding (NNVC). With the beginning of the next-generation video coding standardization process, this paper systematically reviews the design principles of advanced coding algorithms, explores the mechanisms for coding efficiency improvement, measures the coding gain of these efficient algorithms, and discusses potential directions for future efficient video coding. In-depth understanding and analysis of these advanced coding algorithms facilitate the exploration of the next generation video coding standard.
The rapid development of the low-altitude economy has accelerated the large-scale deployment of unmanned aerial vehicles (UAVs). However, unauthorized UAV activities are increasingly threatening urban public safety, giving rise to severe potential hazards including mid-air collisions, privacy infringements, and non-compliant harassment or attacks against critical infrastructure and public events. Addressing this emerging airspace environment, this paper presents a systematic review of research on anti-UAV technologies. First, the paper outlines the unique challenges faced by urban low-altitude security, such as complex sensing environments, difficult target identification, and limited countermeasure options. Subsequently, centered on three stages—sensing, identification, and countermeasure—the current mainstream technical means are comprehensively reviewed. At the sensing level, detection technologies including radar, radio frequency (RF), optical and infrared, acoustics, and multimodal fusion are covered; at the identification level, identification technologies based on identity and behavior are analyzed; at the countermeasure level, neutralization technologies such as jamming suppression, deceptive hijacking, physical capture, and kinetic damage are summarized. Meanwhile, the core evaluation metrics and experimental environment settings of each stage are concluded, with a comparison made between the differences in academia and industry. Finally, the paper forecasts the future trends of anti-UAV technologies, namely, networked sensing, intelligent identification, and collaborative countermeasure.
Aiming at the problems of high communication cost and large verification overhead in existing ring signature schemes under scenarios of large-scale ring members and multi-signature verification, this paper designs a novel aggregate ring signature scheme based on homomorphic vector commitments and compressed ∑ protocols. The scheme first generates individual ring signatures based on BLS signatures,(ℤq,𝔾1)vector commitments, and the compressed ∑ protocol for (ℤq,𝔾1) vector commitments. Furthermore, based on ElGamal-like vector commitments and the ∑ protocol for ElGamal-like vectors, it aggregates k individual ring signatures into one aggregated ring signature for transmission and verification. Theoretical analysis and experimental verification demonstrate that the proposed scheme does not require a trusted third party and satisfies security properties such as anonymity and unforgeability. Moreover, the total signature length and verification time for k individual signatures are both O(k), independent of the ring size, and the efficiency is significantly superior to similar schemes.
Deep semantic communication (Deep JSCC) alleviates the cliff effect of conventional VLC-OFDM under harsh channels, but lacks physical-layer constraints, resulting in high PAPR and LED nonlinearity. To address these issues, a deep learning-based three-dimensional semantic index modulation system (3D-SIM-OFDM) was proposed. Image semantics were extracted and divided into core and secondary parts, which were quantized into index and symbol bits and mapped to subcarrier activation patterns and dual 3D symbols, respectively. Joint detection and semantic reconstruction were then performed. At a CCDF of 10−3, the proposed system reduced PAPR by 3-4 dB compared with Deep JSCC and mitigated LED nonlinear distortion. Under low SNR and extremely low compression ratios, it achieved higher PSNR and SSIM than conventional OFDM and Deep JSCC while accurately reconstructing core semantics. The results demonstrate the strong robustness of the proposed semantic hierarchical mapping scheme.
Traditional shortwave channel models are mainly aimed at single-input single-output or local compact array scenarios, making it difficult to characterize the time-space non-stationarity, path state evolution, polarization characteristics, and spatial correlation among array elements of shortwave wide-area cooperative transmission distributed SIMO (Single-Input Multiple-Output) channels. This limits the effective implementation of tasks such as capacity estimation and waveform design for shortwave integrated access network multi-link wide-area cooperative transmission. To address this issue, a distributed SIMO skywave channel modeling method for shortwave wide-area cooperative transmission is proposed. First, based on the Watterson model, a distributed SIMO channel delay-domain model is constructed, establishing the dynamic geometric relationship between transmitting/receiving array elements and ionospheric reflection clusters, deriving the time-varying delay expressions of propagation paths, and using a Markov process to describe path state changes caused by the birth and death of reflection clusters. Second, an autoregressive model is introduced to achieve time-varying amplitude modeling, constructing a distributed SIMO channel Doppler-domain model. Furthermore, antenna polarization characteristics are introduced to construct a distributed SIMO channel polarization angle power spectrum model. Finally, the polarization angle power spectrum model is coupled into the delay-domain and Doppler-domain models, proposing a unified skywave channel model for distributed SIMO channels in the delay-Doppler-polarization angle domain. The experimental results show that on three long-distance integrated links, the maximum Doppler spread error of this model is 0.28 Hz, the multipath delay spread error is 0.16 ms, the Doppler spread error of the Watterson model is 0.48 Hz, the Doppler spread error of the ITS model is 0.52 Hz, and the multipath delay spread error is 0.36 ms. This model demonstrates good practical applicability.
To address the performance degradation of malware detection models based on deep learning caused by concept drift, a malicious code concept drift detection and adaptation method based on active learning is proposed. This method designs a category sensitive dual threshold drift detection mechanism, setting distance and uncertainty thresholds for each category to detect drift samples in real-time; Adopting an iterative sampling strategy based on embedding space cosine similarity, balancing uncertainty and diversity under limited annotation budget, and sampling high-value samples for annotation; By efficiently fine-tuning parameters through low rank adaptation, the model is incrementally updated to alleviate catastrophic forgetting and reduce computational overhead. The experimental results on the BODMAS dataset show that the proposed method outperforms existing methods in concept drift scenarios, with an accuracy improvement of 0.32%, F1 value improvement of 0.31%, FPR reduction of 0.25%, and FNR reduction of 0.29%, effectively alleviating the model performance degradation caused by concept drift.
Node mobility causes propagation delays to drift within a scheduling cycle. Time division (TD) scheduling and timing advance (TA) based on static delays may lead to transmission-reception conflicts, making deterministic latency difficult to guarantee. To address this issue, a scheduling mechanism tolerant of delay uncertainty was proposed. First, a “TA envelope” locking mechanism was introduced, and redundant time slots were reserved to enhance the robustness of the scheduling table against delay uncertainty. Second, a worst-case fractional interference coefficient for dynamic scenarios was derived to evaluate physical-layer interference. Finally, a time-expanded conflict graph was constructed, and the joint traffic-admission and resource-allocation problem was formulated as a mixed-integer nonlinear programming problem. Delay-constraint generation and state-space pruning strategies were further designed to accelerate the solution. Simulations show that, under the evaluated parameters and a 10 s control period, the proposed mechanism achieves 100% on-time delivery for admitted flows in half-duplex mobile wireless networks with limited redundant-slot overhead. Compared with a scheme using a 20% guard period (GP), slot utilization is about 30% lower, while the average end-to-end delay is reduced by 28.9%; the tolerable node speed is on the order of 10² m/s.
The rapid development of intelligent video services has imposed increasingly stringent requirements on the efficiency, robustness, and semantic fidelity of wireless video transmission. However, existing video transmission systems still face several challenges, including substantial semantic redundancy, difficulties in generative reconstruction, and vulnerability to data distribution shifts. To address these challenges, a multi-task video semantic communication system empowered by a vision-language model (VLM) and a world model (WM) was proposed. At the transmitter, the original video was compressed by the VLM into keyframes and textual semantic descriptions, such that transmission redundancy was reduced while essential semantic content was preserved. At the receiver, high-quality videos were reconstructed by the WM based on the received keyframes and textual semantic descriptions. To mitigate task-specific data distribution shifts in dynamic environments, a continual semantic adaptation mechanism was further developed, through which the semantic and channel encoders and decoders were stably updated using cross-task optimization to preserve semantic understanding across different tasks. High semantic consistency, stability, and generalization capability under different video tasks and channel conditions were demonstrated by the experimental results.
To reduce the performance overhead caused by frequent interval intersection and comparison when conventional interval decision diagrams process interval constraints in access control policies, an efficient policy evaluation and conflict detection method based on half-interval decision diagrams (HIDDs) was proposed. The method uses sequences of half-interval boundaries arranged in ascending order of their right endpoints to implicitly represent attribute-value segments. By recording only the critical points at which policy decisions change, it avoids the redundant storage of complete intervals. Furthermore, a double-pointer scanning-based HIDD composition algorithm is designed to traverse two ordered boundary sequences simultaneously, thereby reducing pairwise interval operations and supporting efficient policy-to-HIDD conversion, access request evaluation, and policy conflict detection. Experimental results show that the construction time of HIDD is up to 72.8% lower than that of conventional interval decision diagrams. Compared with existing approaches, the proposed HIDD method reduces policy evaluation time by up to 48.0% and conflict detection time by up to 55.7%.
In integrated sensing and communication (ISAC) systems, information leakage risks may be increased when sensing targets are also regarded as eavesdroppers. In most existing studies, only a single ISAC signal mode has been considered, and thus the tradeoffs among covert, communication, and sensing performance under different signal structures cannot be directly compared. Therefore, beamforming designs for three ISAC signal modes were investigated in multi-eavesdropper scenarios with non-cooperative and cooperative eavesdropping. First, an imperfect channel model based on angular uncertainty was established, and the echo signal-to-interference-plus-noise ratio in multi-target sensing scenarios was derived. The covert performance under multiple eavesdroppers was then analyzed. On this basis, optimization problems were formulated to maximize the covert rate subject to the transmit power limit, the minimum sensing performance requirement, and the covertness constraint. An iterative algorithm combining a penalty function, successive convex approximation, and semidefinite relaxation was proposed to jointly optimize the communication beamformer and sensing covariance matrix. It was shown through simulations that, over the investigated parameter ranges, higher covert rates were achieved by the collaborative ISAC mode than by the separated and dual-functional ISAC modes. Meanwhile, a better communication–sensing performance tradeoff was achieved while the sensing and covertness constraints were satisfied. Compared with non-cooperative eavesdropping, lower achievable covert rates and detection error probabilities were obtained under cooperative eavesdropping.
Due to the limited coverage of charging service providers, electric vehicle users often need to switch between multiple providers. The lack of interoperability between providers’ authentication systems not only forces users to register multiple platforms. but also lead to cross-platform behavioral linkage and a lack of cross-domain trust. Therefore, a privacy-preserving mechanism for cross domain vehicle charging is proposed. First, a distributed registration and dynamic pseudonym mechanism is designed to achieve cross-domain identity anonymity and behavior unlinkability. Second, a hybrid cross-domain access control model is proposed, in which the consortium blockchain smart contract handles global access decisions and the charging service provider handle local dynamic authorization. In addition, a trusted billing and auditing mechanism is introduced to ensure data integrity and non-repudiation of charging records. Security analysis and experiment results demonstrate that our approach can provide comprehensive charging services for vehicles. It outperforms existing solutions in terms of authentication efficiency, privacy protection, and system scalability, effectively mitigating trust and privacy leakage risks in cross-domain charging scenarios.
To solve the problems of insufficient interpretability and fragmented defense strategy generation in the reasoning process of existing cyber threat intelligence question-answering systems, a cyber threat intelligence question-answering framework integrating a knowledge graph and multi-agent collaboration was proposed. The framework was composed of three major functional modules: the ATT&CK knowledge graph, multi-agent collaboration, and model inference enhancement. A bidirectionally reversible cyber threat knowledge graph was constructed from ATT&CK and STIX 2.1 data. Collaboration among multiple types of agents was implemented using the blackboard architecture and LangGraph technology. Model inference enhancement was achieved by integrating SP-CoT and graph retrieval, whereby attack attribution analysis and defense strategy generation were enabled. Experiments were carried out on the AthenaBench benchmark, and prominent performance improvements over comparison models were obtained. Experimental results demonstrate that the proposed framework delivers superior capabilities in multi-hop reasoning and defense strategy generation.
As manufacturing moves toward flexibility, intelligence, and autonomy, industrial intelligence is shifting from data-centric processing to industrial embodied intelligence with real-time physical closed loops. Addressing the lack of focus on the synergistic coupling of communication, sensing, computing, and control in such scenarios, this paper takes their co-evolution as the main line. It abstracts industrial embodied intelligence as a closed-loop system of information transmission, state sensing, environmental cognition, and physical execution, systematically reviewing key enabling technologies. It reveals how deterministic networking supports distributed sensing and real-time control, how integrated sensing and communication enables continuous semantic mapping, how sensing–computing fusion allows near-source multimodal cognition, and how the computing–control loop ensures stable mapping from digital decisions to physical execution. Key challenges in cross-layer standardization, data alignment, multi-agent collaboration, and security are further summarized, and future research directions are outlined.
To address the transparency limitations of black-box endpoint detection and response (EDR) evaluations, a protection strength evaluation method based on multi-feature learning to rank was proposed. A cross-version kernel scanning framework and a dual-path verification algorithm were developed to locate and classify protection functions. Multidimensional protection-effectiveness data were collected through hot-patching ablation experiments, and a multi-feature fusion model was constructed to rank EDR products across capability groups and technical scenarios. Experimental results showed that the capability-group-level ranking model achieved an average normalized discounted cumulative gain (NDCG) of 0.904 and a Top-1 accuracy of 0.778, while the technical-scenario-level ranking model achieved an average NDCG of 0.975 and a Top-1 accuracy of 0.933. The proposed method effectively distinguishes protection strength differences among different EDR products.
To enhance the security and confidentiality of wireless communication signals, a novel two dimension extended weighted fractional Fourier transform (2D-EWFRFT) communication method based on hyperchaos-driven dual quaternion scrambling for 3D constellation encryption was proposed. A hyperchaotic map with superior dynamic performance was constructed by employing the sine function as the core nonlinear transformation and introducing exponential terms to strengthen the fluctuation of iterative values. On the basis of dual quaternion theory, a 3D constellation encryption algorithm with joint rotation and translation was designed, which was collaboratively controlled by seven independent parameters. The proposed algorithm improved the information entropy of the constellation while effectively reducing computational complexity. The 2D-EWFRFT was constructed to process signals in both row and column dimensions, and the interaction among multiple transform parameters further diversifies the signal structure. Furthermore, a cascade encryption strategy controlled by hyperchaos was presented, in which the 3D modulated signal was first encrypted via constellation scrambling and then processed by 2D-EWFRFT. The limitations of insufficient scrambling in single constellation encryption and weak parameter sensitivity of the conventional EWFRFT were overcomed. Simulation results demonstrate that the proposed method can effectively disrupt the distribution characteristics of modulated signals, strengthen the anti-interception and anti-detection performance of wireless signals, achieve confusion and diffusion of transmitted information, and improve the ability of the communication system to resist typical attacks, such as statistical attacks and brute-force attacks.
