Financial transaction data security management based on blockchain technology

In order to construct a high-quality Ethernet account portrait, this chapter first models Ethernet transfer transaction data, contract creation transaction data and contract invocation transaction data as three kinds of graph structure data, namely, transfer transaction graph, contract creation graph and contract invocation graph, respectively, and then constructs the account portrait from four aspects, namely, manual features of transaction behaviors, cascade features, centrality, and deep account portrait based on graph embedding algorithm, respectively, with the overall framework The overall framework is shown in Figure 1.

Figure 1.

Ethernet account profile technology overall framework

Among them, the manual features of transaction behavior comprehensively analyze the behavioral information of the accounts contained in the transfer transaction network in three kinds of transactions: transfer transaction, contract creation transaction and contract invocation transaction. Based on the manual features, the cascade feature analysis analyzes the behavioral information of the neighboring accounts in the transfer transaction network in the three types of transactions: transfer transaction, contract creation transaction, and contract invocation transaction. Centrality describes the degree of importance of an account, and in-depth account profiling based on graph embedding algorithms captures information about the spatial structure of the transaction network.

Ada-Katz centrality measurement algorithm is generally divided into two parts, the first part is the node centrality computation model, and the second part is the node centrality online updating algorithm, the following will be the main line of these two parts, and gradually analyze Ada-Katz centrality measurement algorithm. 1)

Ada-Katz node centrality calculation model

Ada-Katz centrality is KatzIndex class centrality, the base model of this kind of centrality is centrality measurement based on path weights, which is essentially to count the sum of the weights of all paths pointing to the node. In the base KatzIndex algorithm, when calculating the Katz centrality of a node, the length of the paths pointing to that node is decayed exponentially, and then summed to obtain the Katz centrality. The formula for the base KatzIndex is: (1)Katz→=∑k=1∞βkAk$$\overrightarrow {Katz} = \sum\limits_{k = 1}^\infty {{\beta ^k}} {A^k}$$

where Katz→$$\overrightarrow {Katz}$$ is the vector representation of Katz centrality, A denotes the adjacency matrix of the directed graph, k is the length of the path (number of hops), β is the path length decay factor, and 0 < β < 1. For a given node u, its Katz centrality is computed as: (2)Katz→(u)=∑n∑k=1∞βk|{pathsoflengthkfromntou}|$$\overrightarrow {Katz} (u) = \sum\limits_n {\sum\limits_{k = 1}^\infty {{\beta ^k}} } \left| {\left\{ {paths{\text{ }}of{\text{ }}length{\text{ }}k{\text{ }}from{\text{ }}n{\text{ }}to{\text{ }}u} \right\}} \right|$$

where n denotes the node that can reach node u after k hops of the path.

Similar to the KatzIndex algorithm, the node centrality computation model of the Ada-Katz Centrality algorithm is also modeled based on the weights of the paths pointing to the node, but the difference lies in the fact that the process of calculating the path weights needs to be combined with the time and weight attributes, as shown in Fig. 2, the Ada-Katz centrality of node u is the sum of the weights of all paths pointing to it, and the paths include All paths that node a, b, c, d, e can reach node u.

Figure 2.

Ada-Katz centrality calculation model

By counting the sum of the weights of all the paths to node u, the node centrality in Ada-Katz algorithm is calculated as: (3)Cu=∑n∑ temporalweightedpathsz formntouΦ(z)$${C_u} = \mathop \sum \limits_n \mathop \sum \limits_{\begin{array}{*{20}{c}} {temporal\;weighted\;paths}\;z \\ {form\:n\:to\:u} \end{array}} \Phi (z)$$

where Φ(z) is the path z weight function and n is the start node of path z.

The computation of path weight Φ(z) in the Ada-Katz algorithm needs to be introduced with the following background knowledge. In a network structure data scenario that changes dynamically over time, edge flow data implies the transfer of information, e.g., social information transfer is represented in social networks, and money transfer is represented in financial transfer networks. In the process of information transfer, time passes and causes the node centrality to decay. At the same time, in the process of information transfer, a certain amount of information is lost after each transfer, i.e., the longer the path in the network, the more information is lost, and here we call it the information transfer decay property. To summarize, the path weight calculation rule in Ada-Katz algorithm needs to refer to the temporal time decay and information transfer decay characteristics.

Based on the above background knowledge, we model the paths as temporal sequential paths (Temporal Walks) with weight properties, which are called Temporal Weighted Walks in this paper, as shown in Figure 3.

Figure 3.

Temporal Weighted Walks

Temporal Weighted Walks is defined mathematically as: (4)z=(n0,n1,t1,w1),(n1,n2,t2,w2),⋯,(nj−1,nj,tj,wj);ti−1≤ti$$z = \left( {{n_0},{n_1},{t_1},{w_1}} \right),\left( {{n_1},{n_2},{t_2},{w_2}} \right), \cdots ,\left( {{n_{j - 1}},{n_j},{t_j},{w_j}} \right);{t_{i - 1}} \leq {t_i}$$

For the path Z in Eq. (4), the weights are specified as: (5)Φ(z)=w1β|z|∏i=1jϕ(ti,ti+1)$$\Phi (z) = {w_1}{\beta ^{|z|}}\prod\limits_{i = 1}^j \phi \left( {{t_i},{t_{i + 1}}} \right)$$

Where w₁ is the starting path weight, i.e., the amount of information carried by path z during information transfer. β(0 < β < 1) is the path length attenuation factor, from which the path weights are processed for information transfer attenuation. ϕ(ti,ti+1)$$\phi \left( {{t_i},{t_{i + 1}}} \right)$$ is the time decay function, from which the path weights are subjected to time decay processing. Combining Eq. (3) and Eq. (5), all the path weights pointing to node u are counted, and the Ada-Katz node centrality of node u is calculated as: (6)Cu=∑n∑ temporalweightedpathsz formntouw1β|z|∏i=1jϕ(ti,ti+1)$${C_u} = \mathop \sum \limits_n \mathop \sum \limits_{\begin{array}{*{20}{c}} {temporal\;weighted\;paths}\;z \\ {form\:n\:to\:u} \end{array}} {w_1}{\beta ^{|z|}}\mathop \prod \limits_{i = 1}^j \phi \left( {{t_i},{t_{i + 1}}} \right)$$

Where n is the start node of the temporally sequential weighted path z pointing to node u, β is the path length decay factor, w₁ is the weight of the start edge of the path, and ϕ(ti,ti+1)$$\phi \left( {{t_i},{t_{i + 1}}} \right)$$ is the time decay function, the degree of time decay can be controlled by using different types of time decay functions, and here we introduce two special time decay functions: (1)

ϕ(t1,t2)=1$$\phi \left( {{t_1},{t_2}} \right) = 1$$

This time decay function is constant. When this function is used and weight w₁ is not considered, the Ada-Katz node centrality is the same as the KatzIndex node centrality, so it can be seen that the KatzIndex algorithm is a special case of the Ada-Katz centrality measurement algorithm.

Where c = ln 2/τ is the HALF-LIFE (half-life) factor, τ is the half-life time, and the half-life factor c is determined by the half-life time τ. In this paper, the time decay function is time decayed using an exponential function to capture the temporal change information in the dynamic network. 2)

Ada-Katz node centrality online update algorithm

In dynamic networks such as the Ethernet transaction network, the Ada-Katz centrality measurement algorithm tracks the centrality of nodes associated with edge streams as the edge stream data arrives and updates the node centrality online based on the node centrality computation model described above. When edge stream data arrives, two new parts of the path to the destination node u are added for edge stream data e(u,v,tuv,wuv)$$e\left( {u,v,{t_{uv}},{w_{uv}}} \right)$$, as shown in Fig. 4.

Figure 4.

Side flow data demonstration diagram

(1)

The first part of the path is the start node of this edge pointing to the destination node, i.e., the edge flow data just arrived itself as a path. At moment t = t_uv, the weight of this path is w_uvβ;

where C_u, C_v is the centrality of node u, v, t_u, t_v is the time before node u, v appears at moment t_uv, w_uv is the weight of edge flow data, β is the path length decay factor, and ϕ is the time decay function, respectively. Based on the above node centrality update formula, the

In summary, the Ada-Katz node centrality measurement algorithm can complete the online update operation of node centrality when the edge flow arrives, which achieves the purpose of real-time update.

The Ethernet anomalous account detection in this paper is based on the Ethernet account portrait technology in Chapter 3, takes the anonymous accounts in the transfer transaction network as the research object, combines the machine learning classification model, and solves the data imbalance problem common in the real scenarios in the field of anomaly detection through the data imbalance processing method, so as to improve the accuracy of anomalous account detection, and the overall framework is shown in Figure 5. Later, we will introduce the Ethernet account portrait technology as well as the machine learning classification model and data imbalance processing algorithm used in this chapter, and gradually analyze the advantages and disadvantages of the model through experiments, so as to gradually improve the accuracy of anomaly account detection.

Figure 5.

Ethereum abnormal account detection overall framework

The private chain company node registration verification work can be completed through the chameleon hash authentication algorithm, so that the regulator of the alliance blockchain module can accurately trace the risky company users through the information in the alliance blockchain module, obtain the identity information of the company users, and complete the accurate determination of the risky node identity.

In the company node registration phase of the private chain, the user node is denoted by u_id. The user node will firstly generate its own chameleon hash public and private key pairs by key generation algorithm, respectively (pk, sk), where the public key pk = g^x mod p, and generate a random number r. Using the generated meta- and public key pair information from the previous two steps m to perform the hash computation to get the Hash(y, m, r) = g^mpk^r to compute its chameleon hash value to get the chameleon hash value for the user’s identity, which will be used as the chameleon hash value as the result of chameleon hash encryption is stored in the enterprise table of the federated blockchain to generate the user identity. At the same time, the alliance blockchain regulatory node stores the trapdoor information of the node of the company, and if the user’s identity is validated, the storage is authorized by the alliance blockchain arbitration node and the node registration information is released through the Merkle tree. The flow of node registration stage is shown in Fig. 6.

Figure 6.

Flowchart of node registration phase

In the node identity tracing process, first input the private chain node data dataDu$$data_{D_u}$$ to be traced and initialize the identity sequence ID = ∅, obtain the data on the private chain, separate the private chain node information, parse to get the user node identity encryption result CT={CTuid}$$CT = \left\{ {C{T_{{u_{id}}}}} \right\}$$, determine whether the chameleon hash encryption result belongs to the hash table in the union blockchain, if not, output ID = 0; if it belongs to the hash table in the union blockchain, then calculate the private chain node of the chameleon hash public key pair PK value, retrieve the records on the chain, so as to obtain the identity of the company user corresponding to the chameleon hash value. And show the information u_id of the company user to the supervisory node to complete the node tracing, so as to realize the accurate tracing of the risky company, and the node tracing process is shown in Fig. 7.

Figure 7.

Node traceability phase flowchart

UCI Message is a social networking dataset collected from an online community of students at the University of California, Irvine. In the constructed dynamic graph, each node represents a user and each edge represents a message between two users.

Bitcoin-Alpha and Bitcoin-OTC, two publicly available datasets of the Bitcoin exchange network, are used to study the problem of node identification and classification in the Bitcoin network. In these two datasets, nodes are users on the platform, and when a user evaluates another user on the platform, edges appear between the corresponding two nodes.

Private MLM dataset: In this experiment, the data from the original data from June 2021 to July 2024 are selected for the experiment, and after data preprocessing, a financial transaction network containing 6739 account nodes and 72834 transaction edges is constructed.

The statistics of the above data set after preprocessing are shown in Table 1.

In order to verify the effectiveness of the proposed model Ada-Katz when used for anomalous account detection, this paper conducts comparison experiments with the following methods. 1)

Netwalk: this method is based on a random walk approach to generate contextual information and learns node embedding using an autoencoder model.

In this experiment, a comparison of the performance of different models on several publicly available datasets is presented, and this part of the analysis helps to understand the performance of various models and their applicability on different datasets.

The average AUC performance comparison for all tested time-slice networks is anomalous, as shown in Table 2.

Analyzing the experimental results in Table 2, the following conclusions can be obtained: 1)

The accuracy of detection results of Netwalk, AddGraph, TADDY and model Ada-Katz can reach more than 75% in each dataset, which indicates that the detection effect of the four models can learn the effective information of the graph to a certain extent, which is conducive to the detection of anomalous accounts in financial transactions. These methods all adopt the idea of deep learning and combine the characteristics of graph-structured data. They learn the dynamic evolution of the graph by considering the interactions in the previous time-slice network.

In this experiment, box plots are used to show the AUC results of different datasets with different training ratios. Box plot is a commonly used data visualization method, its upper and lower edges represent the upper and lower quartiles, the line in the middle represents the median, the “tentacles” at the top and bottom represent the maximum and minimum values in the dataset, the hollow circle represents the average value, and the solid circle represents the outliers. The horizontal axis in the box plot represents the training scale of the dataset, and the vertical axis represents the AUC value of the model at that training scale.

The training results of the model on UCI Message, Bitcoin-Alpha, and Bitcoin-OTC are shown in Figure 8, Figure 9, and Figure 10, respectively.

Figure 8.

AUC of different training ratio on the UCI-Message dataset

Figure 9.

AUC of different training ratio on the Bitcoin-Alpha dataset

Figure 10.

AUC of different training ratio on the Bitcoin-OTC dataset

The experimental results show that on the dataset UCI-Message, the AUC value of the model does not have a significant upward or downward trend as the training ratio increases. On the datasets Bitcoin-Alpha and Bitcoin-OTC, the AUC value of the model shows a zigzag upward trend with the increase of the training ratio.

Additionally, it has been observed that there are some differences in the model’s performance on different datasets. For example, on the Bitcoin-OTC dataset, the AUC value of the model is generally high and can reach more than 94%. On the UCI-Message dataset, on the other hand, the AUC value of the model is lower and the accuracy is above 77%. This indicates that the model in this paper has different adaptability to different datasets and can be used for financial anomaly detection on multiple datasets. However, the model also achieves an AUC value of over 77% in the worst case, and in the best case, the AUC value can reach 97%.

Through affiliation analysis, correlation analysis and discriminative power analysis, 17 qualitative indicators and 9 quantitative indicators, totaling 26 indicators, were finally selected to construct a blockchain technology-enabled supply chain finance credit risk evaluation index system, as shown in Table 3.

In this paper, 20 construction enterprises and suppliers in the engineering and construction industry served by Z platform are selected, and the operation data and financial data of these 20 enterprises from 2021 to 2023 are collected, and the data results of 26 qualitative and quantitative indicators of these three years of the 20 companies are collated and summarized in a weighted average according to the credit risk evaluation index system constructed in Table 3, and the resulting matrix of the component scoring coefficients is as Table 4 Shown.

It can be found that F1 represents X22: due diligence, X24: enterprise information integrity, X26: encryption security, X23: traceability ability, X19: distributed ledger, X20: information sharing degree, X25: transaction consensus mechanism, X21: smart payment, X18: peer-to-peer transaction, reflecting the investment of enterprises in blockchain technology; F2 represents X2: operating profit margin, X10: net sales margin, X1: return on equity, X3: total profit, and X7: interest coverage ratio, reflecting the profitability and quality of earnings and the ability to repay debts. F3 represents X5: quick ratio, X13: R&D investment intensity, and X15: inventory turnover ratio, reflecting the company’s solvency, asset flow, and investment in R&D expenditures. F4 represents X9: the growth rate of total assets and X8: the growth rate of net profit, which reflects the growth ability of the enterprise; F5 represents the proportion of X12: technical personnel and X11: the ratio of new technologies, reflecting the proportion of scientific and technological personnel and the research and development of new technologies; F6 represents X17: return on invested capital, X4: accounts receivable turnover ratio, X16: accounts receivable account period, X6: asset-liability ratio, reflecting the profitability of the company’s investment, the operating capacity of accounts receivable and the company’s capital structure; F7 represents X14: Shareholder Equity Ratio, which reflects the capital structure of the core companies in the supply chain.

According to the collected data and expert scoring data obtained through the questionnaire survey, the indicators of Guangdong A Investment Holding Company Limited are shown in Table 5.

The calculated probability of compliance is 97.74%, indicating that with the empowerment of Z Platform blockchain technology, Guangdong A Investment Holding Company Limited has a 97.74% probability of compliance and a low credit risk.

The following calculates the probability of a company’s compliance in a traditional supply chain finance model without the empowerment of blockchain technology. In this model, only 15 quantitative evaluation indicators need to be considered, while 11 indicators related to blockchain technology, such as new technology multiplier, R&D investment intensity, peer-to-peer transaction, distributed ledger, etc. need not be considered. Still taking the data of 20 enterprises served by Z platform from 2021 to 2023 as samples, the probability of compliance of enterprises is calculated by principal component analysis and binary logistic regression analysis.

Through the calculation of IBM SPSS 28.0 software, the results of the principal component analysis method are shown in Table 6, in which the cumulative contribution rate of the five principal component variables with initial eigenvalues greater than 1 is 73.62%, which is greater than 60%, and meets the requirements of Logistic regression analysis.

The next step is to carry out regression analysis, according to the results of regression analysis, the principal component variable with significance less than 0.05 is F1, the constant is -2.317, the significance is less than 0.001, so the probability of keeping the contract of Guangdong A Investment Holding Co. under the traditional model is 58.29%, there is a higher credit risk, and the bank will reject its loan request.

By calculating the probability of compliance of Guangdong A Investment Holding Company Limited under different modes, we find that with the empowerment of blockchain technology, the probability of compliance of the enterprise is 97.74%, which is much higher than the probability of compliance of the enterprise under the traditional supply chain finance mode, which is 58.29%. With such a result, the bank will make a different decision when assessing the risk of the loan application of Guangdong A Investment Holding Co. If the probability of compliance is 97.74%, the bank is more likely to approve the loan application and the enterprise will have the funds to continue the construction project. If the enterprise’s probability of compliance is 58.29%, then the bank will reject its loan application, and as a capital-driven engineering and construction industry, the enterprise will find it difficult to continue its project, thus affecting the enterprise’s project progress and negatively impacting the enterprise or even the entire supply chain. The following analyses the reasons why the likelihood of the enterprise complying is higher than that of the traditional supply chain finance model due to the empowerment of blockchain technology.

First: in the traditional supply chain finance model, the enterprise’s return on net assets, operating profit margin, total profit, accounts receivable turnover, quick ratio, interest coverage multiple, net profit growth rate, total asset growth rate, net sales interest rate, R&D investment intensity, inventory turnover, return on invested capital are weaker than the average of the 20 sample enterprises, which will have a negative impact on the bank’s risk assessment of its loan applications, thus preventing the enterprise from obtaining financing. This will have a negative impact on the bank’s risk assessment of its loan application, thus preventing the enterprise from obtaining financing; And through the empowerment of Z platform blockchain technology, Guangdong A Investment Holding Co., Ltd. scored the highest value of 10 on both indicators of smart payment, and transaction consensus mechanism, and all other blockchain-related indicators scored 7 or more, indicating that blockchain technology can improve the transparency and traceability of enterprise information in the supply chain and reduce the risk of fraud. Banks can learn about the transaction records, inventory, logistics and other information of each enterprise in the supply chain through blockchain technology, so as to more accurately assess the credit risk of the enterprise. Through blockchain technology, transaction records are recorded on a distributed ledger and cannot be tampered with, which prevents tampering of transaction records in the supply chain, reduces the risk of fraud, improves the probability of compliance by enterprises and the trust of banks in the enterprises, helps banks to better assess the credit risk and financing ability of enterprises, and improves the chances of enterprises in the supply chain to obtain bank loans.

Second: In the principal component analysis of the credit risk evaluation index system of supply chain finance, a total of seven principal component variables are extracted, of which F1 denotes X22: prudent investigation, X24: enterprise information integrity, X26: encryption security degree, X23: traceability, X19: distributed ledger, X20: information sharing degree, X25: transaction consensus mechanism, X21: smart Payment, X18: Peer-to-Peer Transaction. We find that the principal component variable F1 contains all the qualitative indicators related to blockchain technology, thus indicating that blockchain technology is an indicator that has a very important impact on the evaluation of credit risk in supply chain finance, which can improve the transparency, credibility, and efficiency of supply chain finance transactions, reduce the credit risk of suppliers, and thus promote the development of the supply chain finance market.