最新刊期
- 摘要：Low Earth Orbit (LEO) satellite constellations feature low latency and high bandwidth, and have become the core carrier for ground terminals to access the integrated space-terrestrial network. Nevertheless, LEO satellite channels are vulnerable to illegal eavesdropping and jamming, and communication links frequently suffer from intermittent outages. Accordingly, enabling trusted access of user terminals to LEO satellite networks has become an urgent challenge to address. This paper proposes a terminal trusted access authentication and key agreement protocol for integrated space-terrestrial networks, which achieves secure and efficient mutual authentication and key agreement between user terminals and LEO satellites. The protocol adopts lightweight cryptographic primitives, and offloads most computational tasks to the ground control center to reduce the computing overhead of satellites. Pseudo-identities and identity passwords are designed to enhance the privacy of user terminals. In addition, the authentication parameters generated during the initial access are reused to support rapid terminal re-authentication after link disconnection, so as to improve the efficiency of re-authentication during reconnection. Formal security analysis and verification are conducted via the Real-Or-Random (ROR) model and the Tamarin prover. The results demonstrate that the proposed protocol achieves improved security and privacy protection. Performance evaluation shows that the protocol reduces communication and computational overhead and obtains a significant improvement in communication efficiency compared with existing schemes, which well satisfies the lightweight requirements of integrated space-terrestrial networks.关键词：Space-air-ground integrated network;access authentication;re-authentication;key agreement;privacy preservation3|0|0更新时间：2026-07-17
-
Multivariate Time Series Anomaly Detection Based on Frequency-Domain Channel Interaction Enhancement AI导读
摘要：To address the insufficient characterization of local frequency variations and channel representations in existing multivariate time series anomaly detection methods, this paper proposes TFDG-AD, a frequency-domain channel interaction enhancement method for multivariate time series anomaly detection. The proposed method first applies the fast Fourier transform to the input sequence and partitions the spectrum into sub-bands, where the real and imaginary components are combined to construct sub-band representations for describing local anomaly features in different frequency ranges. Then, channel relationships are adaptively learned from the current sub-band representations to build sparse normalized channel graphs, on which residual graph propagation is performed to aggregate information from related channels and obtain enhanced sub-band representations. Finally, the enhanced representations are reconstructed back to the time domain, and anomaly scores are calculated by combining time-domain reconstruction errors with frequency-domain errors. Experiments on multiple public datasets show that TFDG-AD effectively improves detection performance and outperforms several baselines.关键词：multivariate time series;anomaly detection;frequency-domain reconstruction;time-frequency joint scoring2|1|0更新时间：2026-07-16 - 摘要：To solve the problem that flight safety and operating rules were difficult to balance simultaneously in multi-UAV cooperative path planning in low-altitude integrated airspace, a multi-UAV cooperative path planning method integrating the asymmetric conflict region (ACR) and the soft actor-critic (SAC) algorithm was proposed. The main idea of the method was that a direction-biased hexagonal conflict region was constructed in the body frame of each UAV, and its geometric asymmetry was exploited to produce differentiated overlap penalties for different passing directions, by which the right-side avoidance, right-side overtaking and clockwise circulation rules were implicitly embedded into the reward signal. A centralized training with decentralized execution framework was adopted for policy learning and deployment. Simulation results show that, in the three typical conflict scenarios, the proposed method achieves a task success rate above 99.83% with the average number of conflicts approaching zero, and induces an orderly passing pattern of right-side avoidance, right-side overtaking, and clockwise circulation, thereby reducing the randomness of passing directions.关键词：low-altitude integrated airspace;multi-UAV;cooperative path planning;deep reinforcement learning;asymmetric conflict region4|2|0更新时间：2026-07-14
-
SSCE-YOLO: Spectral-Spatial Collaborative Enhancement for Infrared Remote-Sensing Object Detection AI导读
摘要：To address noise interference and complex-background disturbance in infrared remote-sensing object detection, this paper proposes SSCE-YOLO, a spectral–spatial collaborative enhancement-based detector, where SSCE-YOLO denotes Spectral-Spatial Collaborative Enhancement YOLO. Unlike single attention or frequency-domain enhancement modules, SSCE-YOLO builds a hierarchical mechanism integrating channel pre-selection, frequency-domain modulation, deep gated propagation, and global stability constraints. Specifically, the Frequency-Spatial Collaborative Filter (FSC) suppresses noise-related frequency bands while preserving target structures; the SpectralGNConv (SGNC) injects FSC-enhanced features into multi-order channel interactions to continuously propagate denoising priors; the Channel-aware Global Context (CGC) uses mean–variance statistics to suppress background-induced high-fluctuation channels; the Efficient Channel Pre-selection Attention (ECPA) controls computational overhead. Experiments on VEDAI-IR and HIT-UAV show that SSCE-YOLO improves mAP@0.5 from 59.2 to 63.7 and from 79.1 to 82.2, respectively, achieving a better accuracy–efficiency balance under noisy and complex-background conditions.关键词：infrared remote sensing;object detection;YOLO;spectral-spatial collaboration9|2|0更新时间：2026-07-14 - 摘要：The rapid expansion of artificial intelligence and emerging applications has triggered a surge in computing demand, making the energy footprint of data centers an increasingly critical concern. Against the backdrop of high renewable energy penetration and the strategic "dual-carbon" targets, the long-standing spatiotemporal mismatch and siloed operation of computing and power systems have constrained both energy efficiency and grid regulation capabilities. Computing-Electricity Collaboration (CEC) has emerged as a vital pathway for sustainable development. By integrating information and energy flows, CEC transforms the inherent mobility of computing workloads into a flexibility resource for the power grid. This paper provides a comprehensive review of the state-of-the-art in CEC research. We first delineate the core concepts and fundamental connotations of CEC, followed by the proposal of a six-layer reference architecture. The technical landscape is then analyzed across four key dimensions: basic theory, perception and migration, analytical computation, and closed-loop control, with a comparative assessment of the applicability, strengths, limitations, and evolution of major technical routes. Furthermore, the study categorizes two dominant operational paradigms—“Computing Following Power” and “Power Following Computing”—and explores the paths for ecosystem development from the perspectives of power utilities, computing providers, platform operators, and end-users. Finally, the paper identifies technical bottlenecks and envisions future trajectories, including intelligent scheduling, green microgrids, and standardization. This work is intended to serve as a systematic reference for both theoretical advancements and engineering implementations in the field of CEC.关键词：computing power networks;computing-electricity collaboration;green collaborative computing;key technologies5|2|0更新时间：2026-07-14
-
Optimization method for edge computing service deployment based on dueling double deep Q-learning AI导读
摘要：To address the issues of uneven deployment and limited resource coverage of edge computing server for Vehicle-to-Everything (V2X), this paper proposes an intelligent deployment method for edge computing servers based on a residual graph attention network and dueling double deep Q-learning. First, the roadside unit (RSU) network is transformed into structured graph data through a V2X environment modeling method that utilizes a dual adjacency matrix and multi-dimensional feature encoding. Second, a graph neural network integrating residual connection and attention mechanism is constructed to extract complex spatial dependencies among nodes. Third, a sequential decision-making framework based on dueling double deep Q-learning is designed, which improves the accuracy of Q-value estimation through value function decomposition and a dual-network update mechanism, while guiding policy convergence using a multi-objective reward function. Experimental results show that compared to existing methods, the proposed method improves service coverage by 5.1% and load balancing by 1.4%, and reduces transmission latency by 17%.关键词：Vehicle-to-Everything;edge computing;attention mechanism;graph neural network;dueling double deep Q-learning3|4|0更新时间：2026-07-13 -
Fed-DCR: A Backdoor Defense Scheme for Federated Learning Based on Dual-Domain Synergistic Detection AI导读
摘要：Federated learning was shown to be vulnerable to backdoor attacks, while existing defense methods were mainly dependent on single statistical features in the parameter space. Under non-independent and identically distributed (Non-IID) conditions, the overlap between benign and malicious update features often led to degraded detection performance, and dedicated defense schemes for low-rank fine-tuning scenarios were still insufficiently investigated. To address these issues, a backdoor defense scheme for federated learning based on dual-domain synergistic detection was proposed. In the frequency domain, discrete cosine transform (DCT) was used to decouple client updates, and semantic alignment together with spectral density clustering was introduced to improve the separability between benign and malicious updates. In the parameter domain, Low-Rank Adaptation (LoRA) spectral consistency detection was adopted to identify anomalous deviations in the low-rank subspace, and causal probing was integrated to actively expose hidden backdoors. Experimental results showed that the proposed scheme effectively decreased the attack success rate of multiple backdoor attacks on both the CIFAR-10 vision task and the Llama-2-7B-based federated LoRA fine-tuning task. Fed-DCR provides a favorable balance between backdoor mitigation and main-task performance in most experimental settings.关键词：federated learning security;backdoor defense;dual-domain synergistic detection;robust aggregation4|6|0更新时间：2026-07-13 - 摘要：To meet the requirement of doctors’ comprehensive diagnosis and treatment across multiple servers in smart healthcare, and to address the lack of efficient switching mechanisms in existing post-quantum authentication schemes, a post-quantum password authentication and fast-switching key agreement protocol for multi-server smart healthcare environments was proposed. With ML-KEM as the core cryptographic primitive, a mechanism combining initial authentication and fast switching authentication was designed. In the initial authentication phase, full authentication was performed, and a master session key and a service session key were established. During multi-server access, fast switching authentication was performed, and a new service session key was independently negotiated with the target server. In terms of security, the session key semantic security of the proposed protocol was proved under the ROR model, and resistance to common attacks was analyzed. In terms of performance, when a user continuously accesses five servers, compared with the repeated full-authentication approach in Wen’s and Chen’s post-quantum schemes, the proposed protocol reduces the cumulative computational overhead by more than 70.34% and the cumulative communication overhead by more than 27.53%. The performance gap increases as the number of servers grows, which shows that the proposed protocol has significant efficiency advantages and good scalability in large-scale deployment scenarios.关键词：post-quantum cryptography;ML-KEM;password authentication;multi-server authentication;fast switching authentication5|5|0更新时间：2026-07-13
- 摘要：With the widespread adoption of fully homomorphic encryption (FHE) in privacy-preserving machine learning, homomorphic Transformer inference based on the CKKS scheme has become an active research topic. However, LayerNorm—an indispensable component in Transformers—requires computing the inverse square root , which is conventionally implemented via Goldschmidt iterations at a cost of 6–8 additional multiplicative depth levels, constituting a criticalbottleneck for inference efficiency. This paper proposes a fused homomorphic LayerNorm computation scheme based on CKKS functional bootstrapping. The key idea is to replace the arcsine recovery polynomial in the EvalMod stage of bootstrapping with a composite polynomial that directly outputs the inverse square root, thereby obtaining the inverse square root "for free" during the bootstrapping process. Under the default setting, in which the input satisfies the calibrated bound and the optional soft clamping is not triggered, this incurs zero additional multiplicative depth. To address the periodic aliasing risk of the sine function, we propose a low-depth range guard consisting of pre-scaling and statistical bounding, complemented by an optional soft-clamping fallback for out-of-distribution inputs, and rigorously prove that the scaled input variance is guaranteed to fall within the monotonic interval of the sine function, fundamentally ensuring computational correctness. The polynomial fitting employs the Remez algorithm to solve for a degree-31 minimax approximation on the Chebyshev basis, achieving an approximation error on the order of 10-11, well below the inherent noise of CKKS bootstrapping. Experimental results demonstrate that the proposed scheme saves 6 levels of multiplicative depth compared to Goldschmidt iteration, achieves a 1.26×speedup for single LayerNorm computation, and reduces peak memory usage by 16.7% relative to Goldschmidt(k=3). On the SST-2 dataset with BERT-Base, this paper attains a homomorphic classification accuracy of 92.1%, only 0.4% below plaintext inference (92.5%). The results confirm that the proposed scheme effectively eliminates the additional depth overhead of inverse square root computation without sacrificing inference accuracy, significantly outperforming traditional Goldschmidt iteration in both performance and resource efficiency. This work provides a novel approach for efficient normalization computation in the encrypted domain and offers key support for the practical deployment of homomorphic Transformer inference, carrying significant theoretical and practical value.关键词：fully homomorphic encryption;CKKS;LayerNorm;Functional Bootstrapping;Transformer;privacy-preserving inference4|5|0更新时间：2026-07-13
- 摘要：To address the trustworthiness challenges faced by agents during autonomous decision-making, environmental interaction, and task execution, including insufficient decision reliability, security risks, loss of behavioral control, and difficulty in accountability, research progress on trustworthy agents was systematically reviewed. First, the fundamental components and representative agent frameworks were reviewed, and the distinctions between agents and large language models were compared. Then, the trustworthiness challenges of agents were systematically defined from seven dimensions, including correctness and reliability, robustness, governance, and controllability. Existing trustworthiness enhancement techniques were further reviewed and analyzed with respect to the fundamental components of agents, trustworthy verification, and multi-agent collaboration. In addition, to evaluate the trustworthiness of agents across different dimensions, existing benchmarks and evaluation metrics are summarized, and the limitations of current benchmarks were analyzed. Finally, future research directions for trustworthy agents are discussed and envisioned.关键词：large language model;agent;artificial intelligence;trustworthiness;trustworthiness enhancement6|4|0更新时间：2026-07-13
- 摘要：With the rapid development of the low-altitude economy, the communication support role of cell-free massive multiple-input multiple-output (MIMO) systems has become increasingly important. To address the limitations of conventional clustering methods based on instantaneous locations, which were subject to delayed updates, and fixed power control schemes, which were unable to adapt to dynamic position variations, the joint optimization problem of unmanned aerial vehicle (UAV) –access point association and uplink power control was formulated as a Markov decision process. A two-stage alternating optimization algorithm based on deep deterministic policy gradient was then proposed, in which the large-scale fading sequence over the entire UAV flight cycle was taken as global prior input to enable joint planning-oriented learning of clustering and power control sequences throughout the full flight period. Simulation results showed that the proposed algorithm achieved higher and more stable system performance than conventional schemes and exhibited good adaptability under various UAV spatial topology distributions. These results demonstrate that the proposed method provides an effective solution for joint UAV–access point association and uplink power control in dynamic low-altitude communication scenarios.关键词：cell-free massive MIMO;UAV;reinforcement learning;Association;power control6|8|0更新时间：2026-07-09
- 摘要：Current multimodal misinformation detection methods lack propagation structure constraints on cross-modal semantic interaction, and fixed fusion strategies fail to accommodate sample-level variations in modal quality. To address these issues, a Structure-Guided Multimodal Adaptive Fusion method (SGMAF) was proposed. First, topological features including node influence, propagation timing, edge types, and node degrees were extracted from the information propagation graph to construct a structural bias matrix, which was injected into the text-image cross-modal attention computation to constrain and reweight the attention distribution. Second, a sample-level adaptive multimodal fusion method was designed to dynamically generate modality fusion weights based on the information quality of text, images, and propagation structures within each sample, thereby mitigating the interference of low-quality modalities. Experimental evaluations were conducted on two public datasets, Pheme and Weibo. The results showed that SGMAF outperformed mainstream baseline models across Accuracy, Precision, Recall, and F1-score, achieving 90.91% and 96.27% accuracy on Pheme and Weibo, respectively. These results validate the effectiveness of the structure-guided mechanism and the dynamic fusion strategy.关键词：misinformation detection;multimodal learning;structure-guided mechanism;adaptive fusion;information propagation network7|3|0更新时间：2026-07-09
- 摘要：Local Differential Privacy (LDP) is vulnerable to data poisoning attacks, yet the robustness of memory-constrained streaming data protocols with domain compression remains understudied. Four mainstream LDP streaming frequent item identification protocols (BGR, DSR, BDR, and CNR) were investigated. First, an attack-driven evaluation framework was proposed, and a data poisoning attack maximizing frequency gain was designed. Evaluated on three datasets using Attack Success Rate (ASR), DSR was found to be the most robust protocol, and storage size was shown to significantly impact robustness. Furthermore, a protocol-agnostic attack detection scheme based on distribution consistency was proposed, utilizing data frequency reconstruction to address missing global frequencies. Experimental results demonstrate that this scheme achieves significant detection performance under low privacy budgets or high poisoning ratios..关键词：local differential privacy;robustness evaluation;data stream mining;memory bounded;data poisoning11|5|0更新时间：2026-07-06
- 摘要：To address the severe multipath-induced inter-symbol interference in underwater acoustic (UWA) channels and the limited equalization capability of deep learning methods, this paper proposes a UWA channel equalization network (RoMAFNet) that integrates rotary position embedding (RoPE) and a multi-scale adaptive fusion mechanism. First, a RoPE-based Transformer encoder multiplicatively embeds inter-symbol time-delay information, enhancing the network's position awareness and time-varying adaptability. Second, a multi-scale adaptive multipath equalization module uses parallel branches with increasing dilation rates to characterize multipath components at different delay scales, then dynamically weights them via a global context-driven gating mechanism. Furthermore, under Lipschitz smoothness and related assumptions, a closed-form error bound for the proposed fusion mechanism is derived, and its online fine-tuning stability is analyzed accordingly. RoMAFNet adopts an "offline pre-training and online fine-tuning" paradigm: on simulated channels, continual per-frame updates with only about 50 training symbols suffice to adapt to new channels, while on measured South China Sea channels it also utilizes incremental training symbols far more efficiently than the comparison methods.关键词：underwater acoustic communication;channel equalization;Rotary position embedding;Multi-scale adaptive fusion;Online fine-tuning deployment8|6|0更新时间：2026-07-05
- 摘要：To address the high T-depth issue in SM3 hash function quantum circuits caused by in-place computation, this paper proposed a low T-depth architecture. By introducing Wallace tree-based quantum carry-save adders and low-depth full/half adders, we compressed the multi-operand modular addition path to a constant level. Furthermore, allocating extra qubits to store intermediate states eliminated the high T-depth operations required for uncomputing ancilla qubits. Experiments demonstrated that using 5453 qubits reduced the circuit T-depth from 4,528 to 1,712 and decreased the total T-gate count from 965,632 to 314,882. This scheme provides an efficient approach for evaluating the post-quantum security of symmetric cryptography.关键词：quantum circuit;SM3;Quantum Carry-Save Adder;T-depth optimization13|6|0更新时间：2026-07-05
- 摘要：Ciphertext-policy attribute-based encryption (CP-ABE) schemes are widely utilized for the secure transmission and sharing of data in the medical Internet of things (IoMT). However, conventional schemes exhibit deficiencies in resisting collusion attacks and providing data integrity auditing. To address these issues, a secure IoMT data scheme supporting cloud auditing and device security detection was proposed. In this scheme, data user keys were generated using secure two-party computation to resist user collusion attacks under a corrupted authority. Furthermore, device fingerprints were embedded by computing a homomorphic privacy-preserving message authentication code within ciphertext-aggregated polynomials, thereby enabling efficient cloud data auditing and secure medical device detection. Security analysis indicates that the proposed scheme ensures cloud data integrity auditing, auditing privacy and reliability, as well as the security status detection of medical devices. It also achieves indistinguishability under chosen-plaintext attacks (CPA) against user collusion even in the presence of a corrupted authority. Simulation results demonstrate that, compared with similar schemes, the proposed scheme reduces communication overhead by approximately 38.7% and computational overhead by approximately 68%. This scheme significantly improves the security and efficiency of medical data sharing.关键词：cloud auditing;attribute-based encryption;Physical unclonable funcation;Homomorphic hash function;secure two-party computation72|50|0更新时间：2026-07-03
- 摘要：To reduce the computational cost of digital signature algorithms on resource-constrained embedded platforms, a RISC-V cryptographic coprocessor was designed and implemented. A hardware-software co-design method was adopted. The key finite-field operations in SM2 scalar multiplication and the SM3 expansion-compression process were mapped to custom coprocessor instructions. A modular add/sub unit, a dedicated prime-field modular multiplication unit, a general Montgomery modular multiplication unit, and an SM3 expansion-compression unit were designed to accelerate the core computations. Experimental results showed that a minimum scalar multiplication time of 1.14 ms was achieved, with an AT value of 2.69. A single message block was processed by the SM3 expansion-compression unit in only 64 clock cycles, and a throughput of 1240 Mbit/s was achieved. Complete digital signature generation and verification were completed in 2.92 ms and 3.40 ms, respectively, with corresponding AT values of 11.77 and 13.70. The proposed approach improves the execution efficiency of SM2 scalar multiplication, SM3 expansion-compression, and digital signature generation and verification, and provides a practical reference for the engineering implementation of Chinese national cryptographic digital signature algorithms in resource-constrained scenarios.关键词：RISC-V;digital signature;SM2 algorithm;SM3 algorithm;coprocessor14|11|0更新时间：2026-07-01
- 摘要：Decentralized Identifiers (DID), as a novel user-controlled identity system, offer an innovative solution to the identity management challenges posed by large-scale, dynamic, and cross-domain collaboration of agents. This paper systematically explores DID technologies for agents. It first analyzes the bottlenecks encountered by the traditional Internet identity paradigm and centralized authentication in agent scenarios, including scalability, mobility, cross-domain mutual trust, and privacy protection. Second, the paper traces the evolution of digital identity technologies from centralization to decentralization, and conducts an in-depth examination of the blockchain ledger-based DID architecture, highlighting its core value in enhancing communication security, enabling fine-grained dynamic permission management, and safeguarding data sovereignty and privacy compliance. The paper further analyzes the current challenges of DID in terms of performance scalability, cross-chain interoperability, and legal compliance governance, and looks forward to future development directions such as building agent reputation systems, realizing cross-domain trust transfer, and designing regulation-oriented identity architectures. This paper aims to provide theoretical references and practical guidance for building secure, trusted, and autonomous network identity infrastructures for Agents.关键词：Decentralized Identifier;Verifiable Credentials;agent;Identity System;blockchain180|7|0更新时间：2026-07-01
-
Privacy-Preserving Task Offloading Strategy for Digital Twin System in Satellite Edge Computing AI导读
摘要：To address the high privacy leakage risks, strict task dependency constraints in dynamic networks, and the inability of traditional algorithms to effectively perceive complex dependency topologies during task offloading for digital twin systems in satellite edge computing environments, we proposed a computation offloading strategy based on priority graph deep reinforcement learning. First, we constructed a privacy-preserving satellite-terrestrial digital twin task offloading architecture to collaboratively optimize delay, energy consumption, and privacy protection. Second, we designed a priority sorting mechanism to satisfy dependency constraints and simplify the solution space, resolving the high scheduling complexity of dependent tasks in dynamic networks. Finally, we developed a priority-based graph deep reinforcement learning algorithm that extracted topological features through a graph attention network to determine the optimal offloading strategy for the sorted tasks. Experimental results show that the strategy achieves multi-objective synergistic optimization. Compared with classic algorithms, system delay decreases by 4.52% to 46.16%, satellite energy consumption reduces by 3.42% to 32.31%, and the number of compromised tasks drops by 2.69% to 57.71%.关键词：Satellite edge computing;Digital twin system;Priority scheduling;graph attention network;deep reinforcement learning26|15|0更新时间：2026-06-18 - 摘要：When multimedia content such as images was transmitted via mobile broadband, it was prone to interception and tampering, which posed a threat to the authenticity of communication content. To address frequent tampering operations in real-world scenarios, such as forgery and splicing, as well as the limitations of existing methods in cross-dataset generalization and in accurately detecting and localizing tampered regions of different scales, a highly generalizable detection and localization method based on multi-granularity feature decoupling and collaboration was proposed. First, a multi-granularity feature extraction network was constructed to capture multi-scale tampering cues at the pixel, region, and image levels through a hierarchical architecture. Second, an explicit decoupling module was designed to decompose feature maps into content-stable and tampering-sensitive components, thereby suppressing interference from irrelevant information such as background textures. Furthermore, a gated adaptive fusion module was introduced to enhance consistency and complementarity across different granularities through feature interaction. Experimental results showed that excellent comprehensive performance was achieved on public datasets such as CASIA_V1, Columbia, NIST16, and Coverage. Strong generalization ability and high localization accuracy were also demonstrated under cross-dataset conditions. The proposed method provides an effective solution for robust image tampering detection and localization in complex real-world scenarios.关键词：Mobile broadband;image tampering detection and localization;multi-granularity feature;explicit decoupling;gate-controlled adaptive fusion18|22|0更新时间：2026-06-17
- Technical support is provided by Beijing Founder electronics co., LTD 京ICP备09064830号-19
京公网安备11010802024621 - It is recommended to read the content of this site in Chrome&IE9+. Please switch to extreme mode in browser 360.
- Cookies We use cookies to help provide and enhance our service and tailor content. By continuing, you agree to the use of cookies.
- Total Visits：766239 Visits Today：587
0