In the industrial Internet environment, Due to its high-value attribute, data imbalance in network activity becomes a pressing concern. Rare attack-such as zero-day exploits and Advanced Persistent Threats (APTs)-pose significant risks, but their samples are often far fewer than those of normal or more common attack types. This imbalance can severely undermine the performance of intrusion detection systems, causing them to prioritize frequent attacks while overlooking rare, high-impact threats, thus weakening overall defense capabilities and compromising security in industrial environments. To address this challenge, we propose a dual-module intrusion detection system, CNN-WSG-IDS, specifically designed to handle data imbalance and significantly enhance the detection accuracy of minority-class threats in industrial Internet systems. The system consists of two core modules: the CNN-WSG module, which mitigates data imbalance by generating more balanced, high-quality data and enriching the diversity of training samples, and the IDS module, which optimizes detection performance during classification, improving both accuracy and reducing false-positive rates. To validate the system's effectiveness, we conducted exten-sive evaluations on the NSL-KDD and UNSW-NB15 datasets. The results demonstrate that, compared to traditional intrusion detection methods, CNN-WSG-IDS not only effectively mitigates the negative impact of data imbalance but also substantially improves the detection rate of minority attacks, highlighting its utility in safeguarding industrial Internet systems against both common and rare threats.

In the industrial Internet environment, attack types are increasingly diverse and stealthy. This requires not only responding to evolving external threats but also detecting complex internal anomalies, making behavior-based intrusion detection crucial. However, the industrial Internet faces challenges such as dynamic network changes, complex architectures, high real-time demands, and limited data and resources. Thus, intrusion detection systems must achieve efficient, stable attack detection with minimal resources and small sample sizes. Few-shot incremental learning methods can continuously update models in resource-constrained environments, adapting to dynamic network conditions. However, incremental learning is prone to catastrophic forgetting, which can degrade model performance. To address this, we propose an incremental learning method combining Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory Networks (Bi-LSTM) with a focus on feature retention. This method leverages CNN and attention mechanisms to efficiently extract spatial features and uses Bi-LSTM to capture temporal information. It enhances the model's adaptability and detection accuracy under data scarcity, significantly mitigating catastrophic forgetting and allowing continuous adaptation to new attacks. Experimental results show that the CBL-ISL method outperforms traditional machine learning algorithms on the UNSW-NB15 and NSL-KDD datasets, with real-time performance and robustness validated through time analysis and error testing. Notably, when faced with evolving attack patterns, the method can update in real-time, maintaining effective protection capabilities.

Detecting Ethereum phishing scams is extremely urgent. In this paper, we propose a novel Hybrid Attention Model for Ethereum phishing scams detection called HATTM to solve the problem of irregular transaction series in Ethereum, fully extract account features and then improve detection performance. Specifically, we take a novel perspective by regarding each transaction of an account as a separate amount-time point to handle irregular data. In the hybrid attention model, we capture intra-account and inter-account trading features through intra-account attention of EPS-FORMER and inter-account attention of EPSGAT, respectively. We further extract Intra-account and Inter-account statistical features to enrich the account representation. The complete representation of accounts is composed of the above four types of features to detect phishing accounts. Experimental results on the real-world Ethereum dataset show that HATTM outperforms existing models and is far ahead in the recall, which indicates that our model can effectively detect Ethereum phishing scams.

As a representative of the public blockchain, Ethereum has been applied in various industries. However, the vast number of transactions on the platform has also brought a number of illegal activities, such as phishing scams, which have caused significant damage to the Ethereum ecosystem. Due to anonymity of the blockchain, it is difficult for detectors to extract features that can be directly applied to phishing scams detection. Existing studies mainly model Ethereum transaction records as a network and mine key information from them to identify phishing addresses. However, these methods usually employ traditional feature engineering or network embedding, ignoring the fine-grained features in the transaction network. In addition, since the original network is too large to make learning difficult, existing work usually uses random walk (RW) to sample a part of nodes for training, thus ignoring the multiplicity of the network. To address these issues, in this article, we propose a three-stream feature fusion (TSFF) approach to enhance the feature representation of nodes. Specifically, we construct node states to guide RW sampling, and manually extracted 8-D features from the resulting dataset as basic features. Temporal features are jointly learned through long short-term memory network and contrastive learning. We combine residual blocks and graph convolutional network to extract fine-grained structural features from transactional networks. Finally, we fuse these three types of features and input them into a downstream classifier. Experiments show that our TSFF (85.3% Precision) outperforms the state-of-the-art methods, and the effectiveness of each feature is demonstrated.

In recent years, phishing scams have seriously threatened Ethereum's ecological security and caused massive economic losses. Moreover, the significant disparity between the number of normal addresses and phishing addresses on Ethereum poses a challenge for detecting phishing scams. Existing studies primarily employ methods such as oversampling, filtering rules, and traditional machine learning models to resolve the Ethereum data imbalance problem. However, these methods disregard topological structure features of the transaction network and the link relationship between nodes. In this paper, we propose an Ethereum phishing scams detection model based on Generative Adversarial Graph Networks called EGAGN to alleviate imbalanced data, enhance node representation, and then improve detection performance. Specifically, the graph generator and discriminator play with each other to generate synthetic nodes that satisfy the real nodes distribution to balance Ethereum data and extract effective network structural features. We further extract statistical features from the transaction network and aggregate transaction records based on time series to obtain trading features. The complete representation of nodes is composed of the above three types of features to detect phishing nodes. Experimental results on the real-world Ethereum dataset show that EGAGN outperforms existing models and is far ahead in recall, which indicates that our model can effectively detect Ethereum phishing scams.

The anonymity of blockchain and its inadequate supervision make it difficult to investigate criminal activities on the blockchain. In recent years, criminals have profited from deploying Ponzi schemes on the Ethereum blockchain through smart contracts, resulting in substantial economic losses and adverse impacts, seriously impeding the development of the blockchain community and technology. However, despite some research on identifying Ponzi schemes on Ethereum, existing methods face certain difficulties in data acquisition, complex feature construction, and insufficient exploration of opcode data features. To address these issues, this paper proposes a method that only relies on smart contract opcodes to verify whether a contract is a Ponzi scheme. Specifically, Word2vec word embedding technology is first used to train the data, obtaining opcode word vector representations through the training process. Subsequently, by passing the word vectors into the Convolutional Neural Network (CNN) and Bidirectional Gated Recurrent Unit Network (BiGRU) models, spatial and semantic features are extracted to better capture semantic information at different levels of smart contract opcodes. The attention mechanism allocates different weights to various features, highlighting key attributes. The results of the experiment show that the proposed method displays a strong detection performance, indicating specific improvements in precision, recall, and F1 score in contrast to previous methods.

Federated learning is a novel machine learning paradigm in which the model is trained on local data by distributed clients. Most of the current research on federated learning assumes that clients are unconditionally providing data and training models, and little consideration has been given to how to incentivize clients with high-quality data to participate in the model training task. Therefore, this paper proposes a blockchain-based federated learning incentive mechanism combining data quality verification and reverse auction. Firstly, by verifying the quality of client data, the client with high-quality data that meets the task requirements is selected, and then the client sends its bid for the task to the reverse auction smart contract. Secondly, some clients with the best performance in different training phases are selected to participate in the task training using smart contracts, and a certain number of clients are selected to form a committee among the unselected participants. The committee members are responsible for validating the local model parameters of the clients while receiving validation rewards. Finally, we conduct simulation experiments on two datasets separately, and the experimental results demonstrate the effectiveness of our proposal.

Online Judge (OJ) systems provide a valuable platform for independent programming skill improvement. However, the sheer number of exercises available often overwhelms users when selecting practice problems. To address this issue, we present the Attention-based Deep Graph Reinforcement Learning algorithm (ADGRL) for exercise recommendation. ADGRL employs graph convolutional neural networks to effectively model user-problem relationships and incorporates an attention mechanism for enhanced personalized recommendations. The model is trained using reinforcement learning to better understand user behavior and provide precise suggestions. Our model is evaluated and trained on a dataset with 14,818 users, 18,171 problems, and 2,887,011 interactions. Comparative experiments with recent graph-based recommendation models demonstrate that ADGRL outperforms them in terms of recall and normalized discounted cumulative gain.

While federated learning (FL) has made significant strides in addressing data privacy concerns, the challenges of heterogeneous data and unfair performance among participants remain substantial. Existing solutions confront challenges such as high computational costs, difficulty in balancing performance with fairness, and poor convergence in partially data heterogeneous environments. The alternating direction method of multipliers (ADMM) is a highly promising approach that effectively addresses issues related to data heterogeneity by imposing constraints on local client updates through dual variables. In this paper, we propose a novel FL framework, named FAT (tilted FL with ADMM), designed to address the issue of data heterogeneity while reducing bias and unfair treatment towards different clients, and it provides a better trade-off between accuracy and fairness. We conducted experiments on two real-world datasets, and the results demonstrate that, compared to existing methods, FAT significantly improves fairness while maintaining accuracy. Our experiments demonstrate that FAT significantly outperforms existing state-of-the-art methods in both accuracy and fairness, offering a superior trade-off between these crucial aspects.

The rise of Ethereum in various economic and social domains has made it a prime target for illegal activities, particularly phishing scams, which has caused substantial financial losses. Existing methods mainly model transaction records as networks and classify nodes. However, a particular challenge is that not all transactions involved in phishing nodes are illegal, which makes phishing detection very difficult. In order to address the problem, in this paper, we propose a Curriculum Learning-based method for Ethereum phishing detection. We incorporate the Local Outlier Factor to measure the difficulty of nodes, considering the significant feature differences among nodes of the same class. By assigning lower difficulty values to easily identifiable nodes and higher difficulty values to nodes involved in mixed-class transactions, we ensure an effective difficulty measure. Then we gradually increase the number of training nodes input into Graph Convolutional Network in each epoch based on a certain ratio determined by the sorted difficulty scores. Finally, we employ LightGBM as the classifier for identifying phishing nodes. Experimental evaluations on a real-world Ethereum phishing scam dataset demonstrate the superiority of our method over baseline approaches, as evidenced by several evaluation metrics.

Phishing scams have become one of the primary frauds on Ethereum, leading to substantial financial losses for users. Therefore, developing an effective phishing detection method for Ethereum holds significant importance for the blockchain ecosystem. However, the heterogeneity of Ethereum poses challenges when it comes to detecting phishing scams. In this paper, we propose a Curriculum Learning-based approach (CLPSD) for Ethereum phishing detection. We collect transaction records to build a graph. In order to distinguish the varying degrees of contribution among nodes, we design a difficulty measurer combining Local Outlier Factor and information entropy. Thus, CLPSD utilizes Graph Convolutional Network to initially learn from easy samples and progressively advance to more complex ones. The experimental results demonstrate that CLPSD outperforms existing methods, highlighting the superior performance of our approach and its significant relevance in enhancing the Ethereum ecosystem.

Blockchain technology has found widespread use in various fields, including finance, healthcare, the Internet of Things, and supply chain management, due to its ability to meet technical and non-technical requirements. Therefore, it is crucial for system designers to evaluate the performance and characteristics of a given blockchain platform before deploying it. One way to achieve this is through the use of blockchain simulators. However, existing simulators have limitations in areas such as model scaling, event scheduling, time control, object model evaluation, and simulation magnitude. To address these limitations, we propose a Blockchain Simulator based on Event-Layered Architecture (BSELA) in this paper. BSELA comprises multiple functional modules with effective model scalability, with the core being our proposed Event-Layered Architecture. This architecture introduces a new event-driven scheduling mechanism for discrete-event simulation, which improves the efficiency, stability, and maintainability of event scheduling during blockchain simulation. Furthermore, we propose a time advancement mechanism for subsequent events based on event rounds, which improves the accuracy of time advancement, simulation complexity, and real-time demand handling capability during the simulation. We begin by validating the efficiency of our simulator by comparing experiments with those of existing blockchain simulators. We then validate the accuracy of BSELA through simulation experiments on blockchain network connectivity, block propagation latency, and the number of INV messages propagated per hour. We validated the accuracy of the experimental results by conducting a comparison with Bitcoin data. Finally, in order to address the research gap in the performance of blockchain systems at the underlying peer-to-peer network level and in other fields, we designed and experimentally validated the node-trust network construction mechanism by adjusting the simulator’s blockchain object model. Our experiments show that this tuning mechanism improves data transfer efficiency by 30%–40%, and also improves security. The design and experiments of the BSELA tuning mechanism validate the scalability and flexibility of our simulator for blockchain research.

The demand for effective, safe, and privacy-preserving machine learning methods has increased due to the rapid growth of large pre-trained models in recent years. In large-scale AI applications, federated learning (FL) has emerged as a cutting-edge method for addressing privacy and data silos issues. However, FL systems are vulnerable to poisoning attacks, and centralized master-slave architectures have reliability, fairness, and security limitations. We propose a secure and efficient decentralized FL framework called ABFL to address these challenges. The framework tightly integrates FL with blockchain technology to strengthen data ownership guarantees and significantly lessen the negative impact of malicious nodes on the global model. Using historical data stored on the blockchain, ABFL enables model update prediction and identifies malicious nodes by verifying consistency. In addition, we present a novel agent consensus mechanism to lower the expense of model cross-validation and increase consensus efficiency. The ABFL framework’s robustness to various sophisticated poisoning attacks while maintaining high model performance and increasing consensus efficiency is demonstrated in a comprehensive analysis of three benchmark datasets.

Blockchain has facilitated the growth of cryptocurrencies but has also provided new ideas for illegals to commit fraud. Research on malicious accounts detection shows that the number of malicious accounts is much smaller than that of benign accounts, leading to imbalanced dataset samples. Most researchers adopt the under-sampling method to help deal with this issue, but this method does not correspond to the actual scale. So, we propose an anomaly detection method based on community discovery. Firstly, we use the transaction information in the Ethereum public chain to build a transaction network and use the Louvain algorithm to divide the transaction network into communities. Secondly, we use the LightGBM algorithm to classify the community. Finally, based on the classification results, we use HBOS, LOF, K-Means, KNN and iForest algorithms as benchmark algorithms for anomaly detection and compare the experimental results using the methods in this paper with the results of anomaly detection using the original transaction network. Experimental show that our method can reduce the amount of data by 35.53% and increase the AUC values of the five algorithms by 7.52%, 8.41%, 14.88%, 0.83% and 27.95%.

In light of the vulnerabilities in traditional medical information systems, including the susceptibility to malicious tampering of patients’ medical records, opaque access decisions, and challenges in data sharing, this paper introduces a fine-grained security sharing model for medical information based on blockchain. Firstly, it leverages blockchain technology in conjunction with attribute-based access control to automate multi-authority fine-grained access control through smart contracts. Secondly, it categorizes sensitive attributes of different users and electronic medical records into distinct levels, implementing personalized privacy protection within the transaction workflow using local differential privacy policies and varying privacy budgets. Finally, a prototype system, built upon a consortium blockchain, was developed and tested using a dataset of 100,000 real-world electronic medical records. Experimental results demonstrate the model’s capability to facilitate stable and reliable transactions between electronic medical record publishers and requestors while ensuring privacy protection.

Applications in the private Ethereum blockchain environment are required by many enterprises. Although, un-like public Ethereum platforms, private Ethereum platforms allow some specific parameters to be configured at setup time to increase throughput, its throughput performance is still not sufficient for enterprise-level applications due to the low scalability of the Ethereum blockchain. This paper analyzes and tests the throughput of the private Ethereum blockchain at the hardware level, and according to the analysis results, proposes a parallelized dynamic load balancing scheme based on network sharding. On this basis, various charts are created to visually show the impact of changing hardware parameters on performance, and experiments are conducted to verify that the model can further improve the performance of private Ethereum blockchains through effective load balancing.

With the development of blockchain technology, the data on the blockchain represented by Bitcoin and Ethereum is exploding. Data represents people's activities, indicating that blockchain has been paid more and more attention. However, many problems are hidden behind so much data, such as security and privacy. Analyzing the data can help to find out the problems and propose methods to improve blockchain. Therefore, in order to make the blockchain well applied to various walks of life, data analysis is quite essential. In this paper, we collected the relevant literature in the field of blockchain data analysis, summarized the current analysis methods, and analyzed the research status from four aspects: security, privacy, performance and prediction of price. Finally, we discussed those research work and forecasted the future trend and challenges in this field, providing some reference for related researchers. We thought that applying machine learning technology to blockchain data analysis will become mainstream in the future.

Named Data Networking (NDN) is a new revolutionary network architecture and a research hotspot of next generation Internet. As a key candidate technology for 5G, D2D communication technology has attracted much attention. Data sharing is the main application of D2D communication, and data transmission is the core of data sharing. At present, there are few researches on D2D data transmission in NDN data sharing. The researches mainly focus on dataset synchronization and distributed file transmission. In addition, the existing data sharing protocol in NDN shows lower efficiency when directly applied to D2D data transmission. According to the main factors that affect the efficiency of data transmission in D-ChronoChat running in NDN, we design three optimized D2D data transmission protocols, and implement the prototype system. In order to evaluate the performance of each protocol and the main factors affecting the performance, we test the transmission delay of each protocol. The test results show that the Multi Threads Batch Packets (MTBP) protocol shows the best performance among the three optimized data transmission protocols. In MTBP, the data transmission rate can reach 6Mbps, nearly 7.4 times higher than the Single Thread Individual Packet protocol (STIP) in D-ChronoChat.