Table of Content

30 September 2026, Volume 44 Issue 5

Previous Issue   

Blockchain

Topic Mining and Evolution Analysis of Blockchain Open-Source Projects Integrating Large Language Model

RAO Ruirui, ZHU Li, YAN Jiaqi

2026, 44(5):  701-715.  doi:10.3969/j.issn.0255-8297.2026.05.001

Asbtract ( 44 )   PDF (3643KB) ( 20 )  

References | Related Articles | Metrics

Based on GitHub blockchain open-source projects from 2021 to 2024, this paper proposed a topic identification framework for open-source projects that integrates a large language model with bidirectional encoder representations for topic (BERTopic). The framework first applied large language model prompt engineering to perform specific data preprocessing and text refinement for open-source projects. It then constructed a BERTopic model through word embedding, dimensionality reduction using uniform manifold approximation and projection (UMAP), and hierarchical density-based spatial clustering of applications with noise (HDBSCAN) to identify blockchain topics. Finally, topic labels for blockchain open-source projects were generated by combining “topic representation and naming using a large language model” with “expert evaluation and optimization”. Based on the topic identification results, a three-dimensional index system comprising “topic intensity-technological influence-secondary development activity” was constructed to evaluate the identified topics and analyze their evolutionary characteristics. The research shows that the blockchain open-source ecosystem presents an evolutionary path of “foundation consolidation-application expansion-intelligent integration”. In addition to identifying common technical topics around smart contracts, decentralized finance (DeFi), non-fungible tokens (NFTs), etc., the study also finds that “AI + blockchain”-related technical topics are becoming new growth points. The multidimensional index evaluation shows that the developer community continues to strengthen security foundations and toolchain development while keeping pace with technological and market trends.

LauHunter: A Fast and Efficient Tool for Money Laundering Detection

GUO Li, ZHANG Wenjie, PAN Heng, LU Siqi, XING Ying, LI Kunyang, ZHANG Bowei, SI Xueming

2026, 44(5):  716-730.  doi:10.3969/j.issn.0255-8297.2026.05.002

Asbtract ( 26 )   PDF (954KB) ( 8 )  

References | Related Articles | Metrics

Existing blockchain risk control methods often struggle to effectively identify suspicious accounts and transaction pathways due to the massive scale of Ethereum network data, sparse label information, and high computational complexity. To address this problem, this paper proposes LauHunter, a novel Ethereum money laundering detection tool that combines reverse clustering analysis with fine-grained behavioral pattern analysis. Experimental results demonstrate that LauHunter significantly outperforms existing mainstream detection methods in both detection accuracy and efficiency, particularly when handling complex fund flow pathways, thus assisting regulatory authorities in tracking illicit fund flows.

Blockchain-Based Protection Framework for Personalized Federated Edge Learning

YUN Jian, LIU Tao, GAO Xinru, ZHANG Haiyu

2026, 44(5):  731-751.  doi:10.3969/j.issn.0255-8297.2026.05.003

Asbtract ( 27 )   PDF (1417KB) ( 14 )  

References | Related Articles | Metrics

Federated learning in edge scenarios faces resource constraints, data heterogeneity, and security threats. This paper proposes a blockchain-based protection framework for personalized federated edge learning. The framework combines a decentralized consensus mechanism, secure compression, and personalized fine-tuning. These components jointly improve robustness to malicious clients and non-IID data. A coordination mechanism integrating secure compression and personalized fine-tuning is designed to enhance local model adaptability, reduce communication overhead, and ensure robust aggregation. A dual incentive mechanism combining contribution and trust is further proposed. Based on the Raft protocol, a verifiable random function is integrated to construct a consensus algorithm called VRF-VGRaft, enabling candidate node screening, random and verifiable leader election, and multi-party verification group construction. Simulation experiments were conducted on the FEMNIST and CIFAR10 datasets. The results show that, compared with existing methods, the framework significantly improves the security and efficiency of federated learning when handling non-IID data and malicious nodes.

A Multi-dimensional Evaluation and Intelligent Dynamic Reconfiguration Scheme for Sharded Blockchain Nodes Based on Deep Reinforcement Learning

MENG Lingyang, LI Peng, SUN Lijun, DU Jian

2026, 44(5):  752-770.  doi:10.3969/j.issn.0255-8297.2026.05.004

Asbtract ( 33 )   PDF (2664KB) ( 13 )  

References | Related Articles | Metrics

Blockchain suffers from low throughput and poor scalability, thus making it difficult to achieve large-scale practical applications. Although sharding technology can improve performance through parallel processing, traditional sharding schemes have limitations including insufficient reconfiguration flexibility and a single-dimensional node evaluation metric, and cannot balance system throughput and security. To this end, this paper introduced a multi-dimensional node evaluation and intelligent dynamic reconfiguration scheme for sharded blockchain nodes based on deep reinforcement learning (DRL), aiming to balance throughput performance and system security during sharding reconfiguration. Specifically, it designed a multi-dimensional evaluation method to construct more comprehensive dynamic profiles for nodes and quantitatively analyze shard throughput and security. Meanwhile, the DRL algorithm was introduced to identify optimal sharding strategies responsive to network conditions based on system states. Simultaneously, node allocation was dynamically balanced, ultimately achieving intelligent and adaptive adjustment of sharding reconfiguration. The experimental results show that compared with the traditional scheme, the proposed method can effectively improve throughput and reduce sharding risk. Additionally, it achieves a long-term dynamic balance between throughput and security, providing a feasible solution for the application of blockchain in large-scale scenarios.

Trusted Data Quality Evaluation Based on Zero-Knowledge Proof

LI Xi, WEN Bin, WANG Yibei

2026, 44(5):  771-786.  doi:10.3969/j.issn.0255-8297.2026.05.005

Asbtract ( 33 )   PDF (1702KB) ( 16 )  

References | Related Articles | Metrics

To address the common data security problems in data quality evaluation, such as privacy leakage, tampering of evaluation results, and lack of reliable verification mechanisms in data transactions, this paper proposed a trusted data quality evaluation method based on zero-knowledge proof (ZKP). By applying zk-SNARK technology, data providers can prove to the verifiers that the data quality is up to the standard without disclosing the original data. Meanwhile, this paper introduced the Pedersen commitment mechanism to ensure the integrity of data content and prevent data tampering. Additionally, the evaluation process and results were stored on the blockchain to further improve the transparency and credibility of the system. The experimental results show that the proposed method can accurately evaluate the quality of multiple datasets. Furthermore, it can smoothly complete the verification by ZKP while protecting privacy. After designing and deploying the smart contract on the Ethereum chain, the system runs stably, the Gas consumption is reduced, and the computational efficiency is improved. The proposed method can effectively protect user privacy while ensuring the credibility of data quality evaluation, and is applicable to scenarios such as data trading, federated learning, and data sharing that require high privacy protection.

Intelligent Information Processing

Multi-exposure Image Fusion Based on RAW Domain

SHU Yong, SHEN Liquan

2026, 44(5):  787-799.  doi:10.3969/j.issn.0255-8297.2026.05.006

Asbtract ( 27 )   PDF (1114KB) ( 7 )  

References | Related Articles | Metrics

Existing multi-exposure image fusion methods in the sRGB domain are prone to generating ghost artifacts when tackling scenes with large motion amplitude; the fused high dynamic range (HDR) images have poor quality, and the algorithm efficiency is low. To this end, this paper proposed a multi-exposure image fusion model based on multi-scale alignment and fusion to directly perform multi-exposure image fusion in the RAW domain. The model designed a multi-scale deformable alignment module (MS-DAM) to implicitly align features of different exposures at multiple scales, thus achieving more accurate alignment and greatly reducing ghost artifacts. Meanwhile, to better perform multi-exposure feature fusion, this paper designed a progressive fusion module (PFM) to perform multiexposure feature fusion at multiple scales and predict the HDR images at the corresponding scales, with features beneficial to subsequent multi-exposure fusion gradually enhanced. Considering the lack of multi-exposure fusion datasets in the RAW domain, this paper constructed a benchmark dataset for multi-exposure image fusion based on the RAW domain. Experimental results show that the proposed method outperforms previous methods for multi-exposure image fusion based on the RAW domain.

GAN-Generated Face Anti-forensic Method Based on Triplet Feature Distribution Loss

SHANG Yuting, FAN Chunnian, CHEN Beijing

2026, 44(5):  800-815.  doi:10.3969/j.issn.0255-8297.2026.05.007

Asbtract ( 19 )   PDF (5961KB) ( 4 )  

References | Related Articles | Metrics

To address the limited transferability of existing anti-forensic methods for generative adversarial network (GAN)-generated faces, this study designed a triplet feature distribution loss (TriFDL) and proposed a TriFDL-based anti-forensic method for GANgenerated faces. The proposed method optimized TriFDL to train the perturbation generator in the anti-forensic model, enabling the anti-forensic faces generated by the perturbation generator to fall, as much as possible, within the common real-face decision region shared by all forensic models, thus deceiving multiple forensic models and improving transferability. In addition, to enhance robustness to JPEG compression, a JPEG compression simulation module was introduced during the training of the anti-forensic model. Experiments conducted on a dataset of StyleGAN-generated faces show that, compared with seven existing methods, the proposed method achieves the best average attack performance against black-box models and is the least affected by JPEG compression, demonstrating strong transferability and robustness.

Secure Query for User Location Privacy Based on Homomorphic Encryption and Hashing

HUANG Xinyi, LE Yanfen

2026, 44(5):  816-826.  doi:10.3969/j.issn.0255-8297.2026.05.008

Asbtract ( 17 )   PDF (658KB) ( 9 )  

References | Related Articles | Metrics

With the continuous proliferation of location-based service applications, indoor positioning technology has become an important foundational technology for location-based services. When users request positioning services from a service provider, protecting users’ location privacy becomes an important issue that needs to be addressed. For scenarios where only specific personnel are allowed to enter and service providers need to collect user access statistics for specific areas in an anonymous state, this paper proposed a three-party protocol that combines an improved spatial Bloom filter with the Paillier encryption algorithm to achieve secure and efficient anonymous statistics on users accessing specific areas. In the protocol, the service provider used multiple Hash functions to generate Bloom filters for each specific area, which were then encrypted and sent to the cloud server. By leveraging the homomorphic properties of the encryption scheme, the cloud server performed the matching operation between the user’s location and the Bloom filter of a specific area, and the service provider obtained secure user access statistics after decryption. Theoretical analysis and experimental results show that, compared with existing secure query schemes, the proposed protocol achieves a lower false positive rate and reduces computational and communication overhead.

Application of an Attention-Enhanced Convolutional Framework in Alzheimer’s Disease Classification

ZHANG Kuo, HU Zhongyi

2026, 44(5):  827-837.  doi:10.3969/j.issn.0255-8297.2026.05.009

Asbtract ( 19 )   PDF (3156KB) ( 6 )  

References | Related Articles | Metrics

The auxiliary diagnosis of Alzheimer’ s disease (AD) has long been an important research topic in brain science. This paper proposed a novel deep learning model termed AttentionConvFusion (ACF). By introducing a self-attention mechanism that dynamically optimizes convolutional kernel weights, the model can more effectively capture critical features in neuroimaging data. Experimental results show that the proposed ACF model achieves satisfactory performance across different AD classification tasks. Specifically, in the classification of AD versus normal cognition, the ACF model achieves an accuracy of 81.5%. For AD versus mild cognitive impairment, it achieves an accuracy of 80.1%. In the multi-class classification task, it achieves an accuracy of 77.8%. These results indicate that the proposed ACF model has significant advantages in the diagnosis and classification of AD.

Steganalytic Method Based on Training Sample Optimization for Cover Selection Steganography

CAO Haiteng, WANG Zichi, ZHANG Xinpeng

2026, 44(5):  838-845.  doi:10.3969/j.issn.0255-8297.2026.05.010

Asbtract ( 15 )   PDF (1100KB) ( 8 )  

References | Related Articles | Metrics

To address the challenge posed by cover selection steganography to traditional steganalysis, this paper proposed an improved steganalytic method based on training sample optimization. Traditional handcrafted feature-based steganalysis usually trains ensemble classifiers using paired clean-stego image pairs during the training phase. Given the characteristic that steganographers deliberately select suitable images for embedding in cover selection steganography, this paper optimized the method of training sample construction by replacing the traditional paired samples with disjoint clean and stego image subsets, making the training data more consistent with real-world steganographic scenarios. Therefore, an ensemble classifier more sensitive to cover selection behavior was trained. Experimental results have shown that the proposed method can effectively improve the detection accuracy of cover selection steganography.

Long-Tail-Aware Point of Interest Recommendation Using Hypergraph and HiLo Attention

HU Junyin, LU Jing

2026, 44(5):  846-860.  doi:10.3969/j.issn.0255-8297.2026.05.011

Asbtract ( 24 )   PDF (808KB) ( 7 )  

References | Related Articles | Metrics

This paper proposes a long-tail-aware POI recommendation method using a global frequency hypergraph network and a multi-span HiLo (high/low) attention mechanism. First, a global frequency hypergraph network is constructed to separately model the subsequent trajectories of head and tail POIs, thus mining higher-order frequency signals across sessions and capturing complex interaction features. Then, a HiLo attention mechanism with multiple temporal spans is used to capture high- and low-frequency features within the hidden sequences. Finally, a long-tail enhancement module is introduced to improve the propagation ability of tail POIs through adaptive edge deletion and enhanced graph construction. Experimental results on multiple real-world datasets show that the proposed method outperforms existing baseline models and demonstrates significant advantages, particularly in long-tail POI recommendation tasks.

Artiflcial Intelligence Technology and Applications

One-Shot Federated Clustering Method for Cluster Distribution Heterogeneity

MA Xiaofeng, ZHANG Cheng, SHI Wenze, LU Lin

2026, 44(5):  861-874.  doi:10.3969/j.issn.0255-8297.2026.05.012

Asbtract ( 33 )   PDF (1462KB) ( 11 )  

References | Related Articles | Metrics

Federated clustering is vulnerable to data heterogeneity under non-independent and non-identically distributed (Non-IID) conditions, making it difficult to maintain robust model performance. Traditional multi-round iterative clustering methods incur high communication overhead and privacy risks, making them unsuitable for resource-constrained environments. To address this issue, this paper proposes federated Bayesian weighted clustering (FedBWC), a one-shot federated clustering method designed to handle heterogeneity in cluster distributions. On the client side, a Bayesian Gaussian mixture model (BGM) with variational inference is applied to local data to adaptively identify effective clusters and extract their centers along with corresponding confidence weights. On the server side, weighted K-means aggregation is performed with only a single communication round, where dynamic weight assignment enhances the representativeness of global centers. Systematic experiments were conducted on six benchmark datasets (Thyroid, Balance Scale, Digits, Waveform, USPS, and PenDigits). The results show that under strong heterogeneity conditions (Dirichlet parameter = 0.3), FedBWC achieves average improvements of 21.6% in V-measure, 28.7% in ARI, and 22.6% in AMI over the baseline methods, while retaining a performance advantage of more than 5% under moderate heterogeneity conditions (Dirichlet parameter = 0.5). Overall, FedBWC significantly outperforms existing methods and provides an efficient and robust solution for federated clustering in complex heterogeneous environments.