A tax evasion identification penetration testing method

Abstract

The application discloses a tax evasion identification penetration testing method, comprising the following steps: simulating the behavior of a tax evasion company individual committing a crime in a transaction network to generate a graph adversarial sample that makes a tax evasion detection model fail to detect a single fake enterprise, performing local testing on the detection model, and calculating the false recognition rate; simulating the behavior of a tax evasion company gang committing a crime in a transaction network to generate a graph adversarial sample that reduces the overall detection performance of the detection model on multiple fake enterprises, performing global testing on the detection model, and calculating the performance reduction; and comprehensively evaluating the security of the model in the actual tax evasion identification scene by combining the false recognition rate of the local testing and the performance reduction of the global testing. The application realizes the evaluation of the security of the tax evasion identification method, diagnoses whether the method has a vulnerability, and thus guarantees the security of the application of the tax evasion identification method in a tax scene with high security requirements.

Description

Technical Field

[0001] This invention belongs to the fields of tax inspection and artificial intelligence information security, and specifically relates to a penetration testing method for identifying tax evasion. Background Technology

[0002] Taxation is a crucial source of national fiscal revenue, the material foundation for maintaining the normal operation of the country, and a guarantee for the stable and healthy development of the economy and society. However, in recent years, tax evasion in my country has become increasingly serious, causing significant losses in tax revenue. Among these crimes, the issuance of false value-added tax (VAT) invoices ranks first among tax-related offenses. On the one hand, with the implementation of the VAT reform, VAT has fully replaced business tax, covering a wide range of areas and gradually becoming the largest tax category in the country. On the other hand, because VAT invoices can be used to deduct VAT and corporate income tax, they essentially possess some of the functions of currency. Therefore, driven by profit, some people evade taxes and seek illicit gains by issuing false VAT invoices. The act of issuing false VAT invoices disrupts the order of the market economy, causes enormous social harm, and seriously damages national tax revenue. Furthermore, the methods of issuing false invoices are becoming increasingly diversified, professional, organized, and covert, which increases the difficulty of investigation and prosecution. Therefore, the efficient detection and identification of false VAT invoices has become an urgent challenge to be addressed.

[0003] Currently, tax audit case selection methods include: traditional methods such as manual selection, whistleblower selection, and computer-based selection; graph pattern matching-based selection methods; and machine learning-based selection methods. With rapid economic development and the implementation of the "Golden Tax Phase III" and "Internet + Taxation" initiatives, tax authorities have accumulated massive amounts of tax-related data. This vast amount of tax data has placed enormous pressure on tax audit work. Traditional case selection methods, relying solely on simple rule-based screening, are insufficient to uncover hidden and complex tax evasion patterns and can only serve as supplementary tools. Graph pattern matching-based selection methods require analyzing historical cases and combining expert experience to extract graph patterns related to tax evasion, resulting in a lag and inability to update in a timely manner. Therefore, tax officials and researchers are focusing more on machine learning-based technologies to detect tax evasion from large amounts of data. For detecting fraudulent VAT invoicing, the latest machine learning-based method constructs a transaction network based on the transaction relationships between taxpayers in VAT invoice records, and then uses a graph neural network model to extract the network embedding representations of taxpayer nodes in the transaction network for subsequent detection of fraudulent invoicing companies.

[0004] Adversarial examples refer to artificially constructed samples that, by adding specific perturbations to the original samples, cause classification models to make incorrect classification judgments on the newly constructed samples. The existence of adversarial examples stems from the difference between the model's decision boundary and the true decision boundary. Recent research has shown that even small adversarial perturbations on graphs can severely degrade the performance of even the most powerful graph neural network models. However, in high-security applications requiring tax evasion detection, graph neural network-based methods for detecting fraudulent VAT invoices suffer from security issues because they do not consider the impact of adversarial attacks (i.e., deliberately fraudulently issuing VAT invoices to initiate fraudulent and concealed transactions in the transaction network) on the model's detection results. Therefore, before practically applying graph neural network-based methods for detecting fraudulent VAT invoices in tax evasion scenarios, it is necessary to evaluate their robustness to adversarial attacks in tax scenarios to avoid vulnerabilities in the tax evasion detection model that could lead to missed detections of companies deliberately fraudulently issuing VAT invoices. How to efficiently generate adversarial examples in transaction networks to evaluate the robustness of tax evasion detection models is a key challenge.

[0005] To address the challenges, the following literature provides efficient methods for generating graph adversarial examples:

[0006] Reference 1. A gradient-based graph adversarial sample generation method for document classification by adding fake nodes (2019104963364).

[0007] Reference 2. A reinforcement learning-based graph adversarial sample generation method for document classification by adding fake nodes (2019104963241).

[0008] Reference 1 proposes a gradient-based graph adversarial example generation method for document classification by adding spurious nodes. First, it acquires the original graph data and a graph node classification model, constructs an attack target set, and then adds spurious nodes to the original graph data to obtain adversarial examples. Next, it selects an attack target (v*, y*) from the attack target set, inputs the current adversarial example into the classification model, calculates the gradient of the loss with respect to the input adjacency matrix, and connects the node corresponding to the element with the largest gradient value in the row corresponding to the spurious node to the added spurious node, obtaining a new adversarial example. This new adversarial example is then input into the classification model. If it makes the classification result of the attack target node v* y*, then the new adversarial example is the generated adversarial example; otherwise, the previous step is repeated until an adversarial example that makes the classification result y* is found. Reference 2 proposes a reinforcement learning-based graph adversarial example generation method by adding spurious nodes. Through reinforcement learning, the generation of adversarial examples by adding spurious nodes is viewed as a decision-making process involving interaction between the machine and the graph node classification model environment. The machine continuously adds edges and features to spurious nodes, collects model classification result information, and then generates adversarial examples that cause the classification model to misclassify the attack target node.

[0009] The methods described in the above literature mainly suffer from the following problems: Firstly, their graph adversarial sample generation methods generate graph adversarial samples that meet the attack requirements by adding fake nodes. These methods are suitable for document classification applications but not for VAT invoice fraud detection applications. In VAT invoice fraud detection, companies typically seek illegal profits by buying and selling invoices without any actual business transactions, or although there are genuine business transactions, the product names, quantities, unit prices, and transaction amounts do not match the invoice information. Fraudulent companies meticulously issue VAT invoices to create fictitious and concealed transactions in the transaction network to evade taxes. Registering fake companies to add fake nodes to the transaction network for VAT invoice fraud is impractical, and the cost of registering a fake company with the tax authorities is too high to implement. Secondly, their proposed graph adversarial sample generation methods can only generate adversarial samples that meet the attack requirements for each target node separately. This only causes the node classification model to misclassify individual nodes; the generated adversarial samples do not reduce the overall classification performance of the model. In tax evasion detection scenarios, the methods of issuing false invoices are becoming increasingly diversified, professional, organized, and covert. Tax evasion gangs operate through complex transaction chains, and they deliberately and systematically engage in legitimate transactions with compliant companies to reduce suspicion of tax evasion by their members issuing false invoices. This leads to the tax evasion detection model misclassifying these companies within the gang. Therefore, the graph adversarial examples generated by the methods described in the aforementioned literature cannot simulate the methods used by tax evasion gangs in VAT invoice false issuance detection scenarios, making it difficult to comprehensively evaluate the robustness and security of tax evasion detection models targeting VAT invoice false issuance. Summary of the Invention

[0010] To address the security shortcomings of existing tax evasion detection methods based on graph neural network models, this invention provides a penetration testing method for tax evasion detection to diagnose vulnerabilities in these methods. This ensures the security of applying these tax evasion detection methods in tax scenarios with high security requirements, which is of great significance in practical applications.

[0011] More specifically, this invention addresses the shortcomings of the graph adversarial sample generation methods proposed in the aforementioned literature when applied to tax evasion detection scenarios. It proposes a penetration testing method for tax evasion identification, which generates graph adversarial samples by adding or deleting perturbed transaction edges in the transaction network based on the model's gradient information (simulating fictitious and concealed transactions by fictitious enterprises in the transaction network). These samples cause the tax evasion detection model to fail to detect a single fictitious enterprise and reduce the overall detection performance of the model against multiple fictitious enterprises. This comprehensively tests the robustness of the tax evasion detection model, thereby fully evaluating the security of the model's application in real-world tax evasion identification scenarios.

[0012] This invention is achieved using the following technical solution:

[0013] A penetration testing method for identifying tax evasion includes the following steps:

[0014] 1) Simulate the behavior of individual tax evasion companies engaging in fraudulent and concealed transactions in the transaction network, generate graph adversarial samples that cause the tax evasion detection model to fail to detect individual fraudulent enterprises, conduct local testing on the detection model, and calculate the false recognition rate;

[0015] To evaluate the security of tax evasion detection algorithms based on graph neural network models, local testing was conducted. The purpose of local testing was to simulate the behavior of tax evasion companies engaging in transactions or deliberately concealing transactions by adding or removing some transaction edges between the target tax evasion node and other nodes in the transaction network. Then, the algorithm was tested to see if it could correctly identify these tax evasion companies.

[0016] 2) Simulate the behavior of tax evasion gangs in the transaction network by engaging in fraudulent and concealed transactions, generate graph adversarial samples that reduce the overall detection performance of the detection model against multiple fraudulent enterprises, conduct global testing on the detection model, and calculate the performance degradation.

[0017] To evaluate the security of tax evasion detection algorithms, a global test was conducted. The purpose of the global test was to add or remove some transaction edges between selected company nodes in the transaction network to simulate the intentional transactions and concealment of transactions between tax evasion criminal gangs, as well as the scenario where tax data is attacked and the transaction network is tampered with. Then, the detection performance of these tax evasion algorithms on the perturbed transaction network was tested.

[0018] 3) By combining the error recognition rate of local tests and the performance degradation of global tests, the security of the model in actual tax evasion detection scenarios can be comprehensively evaluated.

[0019] A further improvement of the present invention is that step 1) specifically includes the following implementation steps:

[0020] Step 1: Obtain the target tax evasion company's node

[0021] Obtain the tax evasion detection algorithm to be tested and the transaction network tax data it uses. Use the tax evasion detection algorithm to identify company nodes in the transaction network. Select companies in the test set that are actually labeled as tax evading enterprises and identified as tax evading enterprises by the algorithm as the target tax evading company nodes for local testing.

[0022] Step 2: Construct a set of potential transaction company nodes for the target tax evasion company node.

[0023] For each target tax evasion company node, construct its potential transaction company node set, that is, related company nodes in the transaction network whose business scope belongs to the same industry or the upstream and downstream industries of the industrial chain as the target tax evasion company;

[0024] Step 3: Generate disruptive transaction edges for the target tax evasion company node

[0025] Perturbation transaction edges are generated individually for each target tax evasion company node. This involves adding or deleting relevant transaction edges connected to the target tax evasion company node in the transaction network to simulate the behavior of the tax evasion company engaging in transactions or deliberately concealing transactions.

[0026] Step 4: Identify the target tax evasion company nodes and obtain the false identification rate.

[0027] Each target tax evasion company node is identified separately, and the proportion of incorrectly identified nodes to the total number of target tax evasion company nodes is the false identification rate.

[0028] A further improvement of the present invention is that, in Step 3 of step 1), the specific process is as follows:

[0029] 1-1) Obtain the graph neural network model of transaction network G = (A,X) and extract the embedded features of company network, where A is the adjacency matrix of the transaction network and X is the basic features of the companies selected by the tax experts from the rule base;

[0030] 1-2) Train a graph neural network model f on the transaction network G. θ θ represents the network parameters, and the loss function for calculating the target tax evasion company node u is: y u It is the real label of node u, the target tax evasion company;

[0031] 1-3) Calculate the loss function The gradient of the adjacency matrix A is obtained for any two nodes v in the transaction network G. i and v j gradient of the transaction edge between Where i,j∈{1,…,n} represent the index of the node, and n is the number of nodes;

[0032] 1-4) Select the edge with the largest gradient among the possible connected transaction edges of the target tax evasion company node u and its potential transaction company node set, and perturb it, that is, add or delete this perturbation transaction edge.

[0033] A further improvement of the present invention is based on whether the perturbation transaction edge originally existed.

[0034] A further improvement of the present invention is that, in Step 4 of step 1), the specific process is as follows:

[0035] 2-1) Retrain the graph neural network model on the transaction network after generating perturbation transaction edges for each target tax evasion company node to obtain node network embedding features;

[0036] 2-2) Follow-up steps of the tax evasion detection algorithm;

[0037] 2-3) Identify the target tax evasion company at this point and record whether the identification was correct;

[0038] 2-4) After repeating the first three steps to identify all the target tax evasion company nodes, calculate the false identification rate.

[0039] A further improvement of the present invention is that step 2) specifically includes the following implementation steps:

[0040] Step 1. Obtain the detection performance of the tax evasion detection algorithm.

[0041] Obtain the tax evasion detection algorithm to be tested and the transaction network tax data it uses; use the tax evasion detection algorithm to identify company nodes in the transaction network; and obtain the identification accuracy and recall rate of the tax evasion detection algorithm in the test set.

[0042] Step 2. Generate transaction edges to add or hide in the transaction network.

[0043] Select some company nodes in the transaction network, and add or remove some specific transaction edges between them;

[0044] Step 3. Obtain the performance degradation of the tax evasion detection algorithm after perturbation.

[0045] The graph neural network model is retrained on the transaction network after adding or hiding transaction edges to obtain the node network embedding features. This is then used for subsequent steps of the tax evasion detection algorithm. The accuracy and recall of the tax evasion detection algorithm in the test set are obtained. The difference between the accuracy and recall of the algorithm and the previous tax evasion detection algorithm in the same test set is the performance degradation.

[0046] A further improvement of the present invention is that, in Step 2), the specific process is as follows:

[0047] 3-1) Obtain the transaction network G = (A,X) and extract the graph neural network embedding features of the company network, and train the graph neural network model f on the transaction network G. θ ;

[0048] 3-2) Determine the maximum number of perturbation edges Δ for each transaction edge, which is set according to the specific transaction network and the security requirements of the tax evasion detection model;

[0049] 3-3) Solve the following objective function using gradient descent to obtain the edge perturbation indicator vector s, where V is the taxpayer company node in the test set;

[0050]

[0051] st,1 T s≤Δ, s∈[0,1] n

[0052] 3-4) Processing the edge disturbance indicator vector s: First, filter the disturbance transaction edges represented by the edge disturbance indicator vector that conform to reality. If the business scope of the two company nodes represented by the edge disturbance indicator vector belongs to the same industry or to the upstream and downstream industries of the industrial chain, then the disturbance transaction edge is considered to conform to reality; otherwise, let its corresponding edge disturbance indicator variable s i =0; then, if the side disturbance indicator variable s i If the value is greater than 0.5, set it to 1; otherwise, set it to 0. Finally, check if the edge perturbation indicator vector s satisfies 1. T If i ≤ Δ, then randomly add some edge perturbation indicator variables s. i Set it to 0 until this condition is met;

[0053] 3-5) Based on the edge disturbance indicator variable s with a value of 1 i The perturbed transaction network is obtained by adding or deleting transaction edges in the transaction network.

[0054] A further improvement of the present invention is that the maximum number of perturbation edges Δ is set to 5% of the total number of transaction edges in the transaction network.

[0055] Compared with the prior art, the present invention has at least the following beneficial technical effects:

[0056] 1. This invention proposes to generate graph adversarial samples by adding or deleting perturbed transaction edges in the transaction network, which is more practical and feasible in the scenario of fraudulent issuance of value-added tax invoices.

[0057] 2. This invention proposes to simulate the behavior of individual tax evasion companies and tax evasion gangs engaging in fraudulent and concealed transactions in a transaction network. Graph adversarial samples are generated to cause the tax evasion detection model to fail to detect a single fraudulent enterprise and to reduce the overall detection performance of the detection model for multiple fraudulent enterprises. This allows for a comprehensive test of the robustness of the tax evasion detection model, which can more comprehensively evaluate the security of the model in real-world tax evasion identification scenarios.

[0058] 3. This invention provides a penetration testing method for identifying tax evasion. On the one hand, it diagnoses the vulnerabilities of these methods, thereby ensuring the security of applying these tax evasion identification methods in tax scenarios with high security requirements, which is of great significance in practical applications. On the other hand, its research can also help design more robust and secure tax evasion detection models. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the overall process of a penetration testing method for identifying tax evasion according to the present invention;

[0060] Figure 2 This is a schematic diagram of the process for partial testing of the present invention;

[0061] Figure 3 A schematic diagram illustrating the process of generating perturbation transaction edges for the target tax evasion company node during local testing;

[0062] Figure 4 This is a schematic diagram of the global testing process of the present invention;

[0063] Figure 5 A flowchart illustrating the process of adding or hiding transaction edges in the transaction network during global testing. Detailed Implementation

[0064] To more clearly illustrate the technical solution of the present invention, a penetration testing method for identifying tax evasion according to the present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not limit it in any way.

[0065] In this embodiment, the tax evasion detection algorithm to be tested for security is FBNE-PU (Gao Y, Shi B, Dong B, et al. Tax Evasion Detection with FBNE-PU Algorithm Based on PnCGCN and PU Learning[J]. IEEE Transactions on Knowledge and Data Engineering,2021,PP(99):1-1.). It uses the graph neural network model PnCGCN to extract the network embedding features of company nodes in the transaction network, then integrates the basic features of the companies selected by experts and the learned network embedding features as the company features, and combines PULearning to obtain the pseudo-labels of all company nodes in the transaction network. Finally, it uses the MLP model to detect tax evasion. In addition, in this embodiment, the transaction network tax data used by FBNE-PU for tax evasion detection is TAX20H, which is a tax dataset for tax evasion detection obtained by preprocessing the 2020 Shanghai value-added tax invoice issuance records. The TAX20H transaction network contains 43,082 company nodes, with 54,332 transaction edges between them. Each company node has 300-dimensional node features, but only 1,724 company nodes have labels. Each node has only one label, and there are two categories for the labels: positive samples represent tax evasion companies, and negative samples represent legitimate companies. Of the 1,724 labeled company nodes in TAX20H, 931 are tax evasion companies and 793 are legitimate companies.

[0066] The overall process of the method of this invention is as follows: Figure 1 As shown, the method mainly involves three steps: local testing, global testing, and comprehensive security assessment. More specifically, the present invention provides a penetration testing method for identifying tax evasion, comprising the following steps:

[0067] Step 1. Simulate individual tax evasion by a single company, conduct local testing, and calculate the false recognition rate.

[0068] Figure 2 The overall process of local testing of the present invention is shown, and its specific steps are as follows:

[0069] S201. Obtain the node of the target tax evasion company.

[0070] Obtain the transaction network tax data TAX20H and the tax evasion identification algorithm FBNE-PU to be tested. Use the tax evasion identification algorithm FBNE-PU to identify company nodes in the transaction network TAX20H. Select companies in the test set that are actually labeled as tax evading enterprises and are correctly identified as tax evading enterprises by the tax evasion algorithm as target tax evading company nodes for local testing.

[0071] In this embodiment, the 1724 labeled company nodes in TAX20H are divided into training set, validation set and test set in a ratio of 1:1:8. There are 745 company nodes in the test set that are actually labeled as tax evasion companies. Among these 745 tax evasion company nodes, FBNE-PU correctly identified 673 tax evasion companies. Therefore, these 673 company nodes are used as the target tax evasion company nodes for local testing.

[0072] S202. Construct a set of potential transaction company nodes for the target tax evasion company node.

[0073] In tax scenarios, companies typically transact with related companies within the same or upstream / downstream industries of their business scope. Therefore, the first step is to construct an industry transaction association rule base based on prior tax knowledge of the industries of the companies involved in the transactions. Then, a set of potential transaction company nodes is constructed for each target tax evasion company node: based on the target tax evasion company's industry, its upstream and downstream industries within the prior industry transaction association rule base are identified. The set of related company nodes belonging to these industries in the transaction network constitutes its potential transaction company node set. In this embodiment, assuming a target tax evasion company belongs to the cement industry, according to the prior industry transaction association rule base, its upstream industries are limestone raw materials, clay raw materials, and other mineral mining industries, and its downstream industries are infrastructure, construction, and real estate. Therefore, the set of company nodes belonging to these industries in TAX20H constitutes the potential transaction company node set for this target tax evasion company.

[0074] S203. Generate a disruptive transaction edge for the target tax evasion company node.

[0075] The process of generating perturbation transaction edges for each target tax evasion company node u is as follows: Figure 3 As shown:

[0076] S301. Obtain the transaction network and graph neural network model

[0077] The graph neural network model PnCGCN is used to extract the embedded features of the company network, and the transaction network G = (A, X) is obtained from the tax data TAX20H, where A is the adjacency matrix of the transaction network and X is the basic features of the companies selected by tax experts from the rule base. In this embodiment, the TAX20H transaction network has 43,082 company nodes and 54,332 transaction edges between them. Each company node has corresponding node features with a dimension of 300. Therefore, A ∈ {0,1} 43082×43082 , X∈R 43082×300 .

[0078] S302. Train the graph neural network model and calculate the loss function of the node u of the target tax-evading company.

[0079] In this embodiment, following the same partitioning method as in S201, the 1724 labeled company nodes in TAX20H are divided into training, validation, and test sets in a 1:1:8 ratio. Then, the graph neural network model PnCGCN is trained on the transaction network G, and the loss function for the target tax evasion company node u is defined as follows:

[0080]

[0081] where f with parameter θ θ This represents the trained graph neural network model PnCGCN. It is a loss function, such as cross-entropy. This represents the predicted label and the actual label y of the target node u. u The larger the value of the loss function, the worse the prediction result of the trained graph neural network model PnCGCN for the target node u.

[0082] S303. Calculate the gradient of the transaction edges in the transaction network.

[0083] Calculate the target loss function The partial derivative of the adjacency matrix A of the transaction network is used to obtain the edge gradient matrix g, thus allowing any two nodes v in the transaction network to be considered as edge gradient matrices. i and v j The gradient information of the edges between them is shown below, which gives the gradient of each transaction edge.

[0084]

[0085] S304. Select Disruption Transaction Edge

[0086] The perturbation of the trading edge should be based on the loss function. The selection is based on the goal of maximizing the gradient. Therefore, gradient g is selected from the transaction edges that may connect the target tax evasion company node u to the set of potential transaction company nodes constructed in step S202. uj The largest edge is perturbed. Depending on whether the perturbed edge originally existed, it is either deleted or added to the transaction network (by modifying the adjacency matrix A of the transaction network G), resulting in a transaction network after generating perturbed transaction edges for each target tax evasion company node u.

[0087] S204. Identify the target tax evasion company node and obtain the false identification rate.

[0088] In each transaction network The graph neural network model PnCGCN is trained to obtain node network embedding features, which are then used in the subsequent steps of the tax evasion detection algorithm FBNE-PU. The algorithm then identifies the target tax evasion company node u and records whether the identification is correct. The proportion of incorrectly identified nodes to the total number of target tax evasion company nodes is the misclassification rate. In this embodiment, after local testing, the tax evasion detection algorithm FBNE-PU incorrectly identified 14 tax evasion company nodes out of 673 originally correctly identified target tax evasion company nodes; therefore, its misclassification rate is [missing information].

[0089] Step 2. Simulate a tax evasion gang's activities, conduct a global test, and calculate the performance degradation.

[0090] Figure 4 The overall process of global testing of this invention is demonstrated, which specifically includes the following steps:

[0091] S401. Obtain the detection performance of the tax evasion detection algorithm.

[0092] The tax evasion detection algorithm FBNE-PU to be tested is obtained, and the transaction network G = (A, X) is obtained from the tax data of the transaction network TAX20H. The tax evasion detection algorithm FBNE-PU is used to identify the company nodes in the transaction network G, and the accuracy (Accuracy1) and recall (Recall1) of the tax evasion detection algorithm in the test set are obtained. In this embodiment, the 1724 labeled company nodes in TAX20H are divided into training set, validation set and test set in a ratio of 1:1:8. In the test set, there are 745 company nodes with the real label of tax evasion enterprises and 634 company nodes with the real label of normal enterprises. The accuracy (Accuracy1) and recall (Recall1) of FBNE-PU in the test set are 90.36% and 96.24%, respectively.

[0093] S402. Generate transaction edges to add or hide in the transaction network.

[0094] The process involves carefully selecting certain company nodes within the transaction network and adding or removing specific transaction edges between them, as detailed below. Figure 5 As shown:

[0095] S501. Training a Graph Neural Network Model

[0096] The graph neural network model PnCGCN, which extracts the network embedding features of the company, is obtained, and the transaction network G = (A,X) is obtained from the transaction network tax data TAX20H. The graph neural network model f is then trained on the transaction network G. θ .

[0097] S502. Determine the maximum number of perturbation edges Δ on the transaction edge.

[0098] The maximum number of disturbed edges Δ in a transaction network is determined. This number can be set according to the specific transaction network and the security requirements of the tax evasion detection model. For example, the maximum number of disturbed edges Δ can be set to 5% of the total number of transaction edges in the transaction network. In this embodiment, there are 54,332 transaction edges between company nodes in the TAX20H transaction network. Therefore, the maximum number of disturbed edges Δ = 54,332 × 5% ≈ 2716.

[0099] S503. Solve the objective function using gradient descent to obtain the edge perturbation indicator vector s.

[0100] First, define the following objective function:

[0101]

[0102] st,1 T s≤Δ, s∈[0,1] n

[0103] Where V is the set of company nodes in the transaction network test set, Δ is the maximum number of perturbed edges in the transaction determined in step S502, and the edge perturbation indicator vector s∈[0,1] n It has n = N(N-1) / 2 flags (where N is the number of company nodes in the transaction network) indicating whether any two company nodes v in the transaction network G are disturbed. i and v j The edge e between ij The indicator variable, where S ij =1 indicates that it will be at node v i and v j Add or remove transaction edges between ij .

[0104] Then, the objective function is solved using the following gradient descent method to obtain the edge perturbation indicator vector s:

[0105]

[0106] Where t represents the number of iterations, η t This represents the learning rate at the t-th iteration. This indicates that the objective function f(s) is in s (t-1) The gradient at point s is obtained after T iterations using gradient descent. (T) ∈[0,1] n In this embodiment, T = 50, η t =0.01.

[0107] S504. Based on the actual prior industry transaction association rules, process the edge perturbation indicator vector s.

[0108] First, we screen the perturbation transaction edges represented by the edge perturbation indicator vector that conform to reality. If the business scope of the two company nodes represented by the edge perturbation indicator vector belongs to the same industry or to the upstream and downstream industries of the industrial chain (this can be determined by constructing an industry transaction association rule base based on the prior tax knowledge built in step S202), then the perturbation transaction edge is considered to conform to reality. Otherwise, let its corresponding edge perturbation indicator variable s i =0; then, if the side disturbance indicator variable s i If the value is greater than 0.5, set it to 1; otherwise, set it to 0. Finally, check if the edge perturbation indicator vector s satisfies 1. T If s ≤ Δ, then randomly add some edge perturbation indicator variables s. i Set it to 0 until this condition is met.

[0109] S505. Add or delete transaction edges in the transaction network based on the edge perturbation indicator vector s.

[0110] Based on the edge perturbation indicator variable s with a value of 1 i In the transaction network, node v i and v j Add or remove the corresponding transaction edge e between them ij The perturbed transaction network is obtained by modifying the adjacency matrix A of the transaction network G.

[0111] S403. Obtain the performance degradation of the tax evasion detection algorithm after perturbation.

[0112] Transaction network after adding or hiding transaction edges The graph neural network model PnCGCN is trained to obtain node network embedding features, and then the subsequent steps of the tax evasion detection algorithm FBNE-PU are performed. The accuracy (Accuracy2) and recall (Recall2) of the FBNE-PU algorithm are obtained on the test set. |Accuracy2-Accuracy1| and |Recall2-Recall1| represent the performance degradation. In this embodiment, after global testing, FBNE-PU performs well on the transaction network. In the test set, the accuracy (Accuracy2) was 86.74% and the recall (Recall2) was 94.50%. Therefore, the performance degradation was |Accuracy2-Accuracy1| = 3.62% and |Recall2-Recall1| = 1.74%.

[0113] Step 3. Conduct a security assessment by considering both the false recognition rate and performance degradation.

[0114] The security of a tax evasion detection algorithm is evaluated by combining the misclassification rate in local tests with the performance degradation |Accuracy2 - Accuracy1| and |Recall2 - Recall1| in global tests. If misclassification < 3%, and |Accuracy2 - Accuracy1| < 5% and |Recall2 - Recall1| < 5%, the algorithm is considered secure and can be practically applied in tax evasion detection scenarios; otherwise, the algorithm is considered insecure. In this embodiment, in the local tests of the tax evasion detection algorithm FBNE-PU, the misclassification rate was 2% < 3%; in the global tests of FBNE-PU, the performance degradation |Accuracy2 - Accuracy1| = 3.62% < 5% and |Recall2 - Recall1| = 1.74% < 5%. Therefore, after comprehensive evaluation, the tax evasion detection algorithm FBNE-PU is deemed secure.

[0115] Those skilled in the art will readily understand that the above description is merely an embodiment of the method of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A penetration testing method for identifying tax evasion, characterized in that, Includes the following steps: 1) Simulate the behavior of individual tax evasion companies engaging in fraudulent and concealed transactions in the transaction network, generate graph adversarial samples that cause the tax evasion detection model to fail to detect individual fraudulent enterprises, conduct local testing on the detection model, and calculate the false recognition rate; To evaluate the security of tax evasion detection algorithms based on graph neural network models, local testing was conducted. The purpose of local testing was to simulate the behavior of tax evasion companies engaging in transactions or deliberately concealing transactions by adding or removing some transaction edges between the target tax evasion node and other nodes in the transaction network. Then, the algorithm was tested to see if it could correctly identify these tax evasion companies. 2) Simulate the behavior of tax evasion gangs in the transaction network by engaging in fraudulent and concealed transactions, generate graph adversarial samples that reduce the overall detection performance of the detection model against multiple fraudulent enterprises, conduct global testing on the detection model, and calculate the performance degradation. To evaluate the security of tax evasion detection algorithms, a global test was conducted. The purpose of the global test was to add or remove some transaction edges between selected company nodes in the transaction network to simulate the intentional transactions and concealment of transactions between tax evasion criminal gangs, as well as the scenario where tax data is attacked and the transaction network is tampered with. Then, the detection performance of these tax evasion algorithms on the perturbed transaction network was tested. 3) By combining the error recognition rate of local tests and the performance degradation of global tests, the security of the model in actual tax evasion detection scenarios can be comprehensively evaluated.

2. The penetration testing method for identifying tax evasion according to claim 1, characterized in that, Step 1) specifically includes the following implementation steps: Step 1: Obtain the target tax evasion company's node Obtain the tax evasion detection algorithm to be tested and the transaction network tax data it uses. Use the tax evasion detection algorithm to identify company nodes in the transaction network. Select companies in the test set that are actually labeled as tax evading enterprises and identified as tax evading enterprises by the algorithm as the target tax evading company nodes for local testing. Step 2: Construct a set of potential transaction company nodes for the target tax evasion company node. For each target tax evasion company node, construct its potential transaction company node set, that is, related company nodes in the transaction network whose business scope belongs to the same industry or the upstream and downstream industries of the industrial chain as the target tax evasion company; Step 3: Generate disruptive transaction edges for the target tax evasion company node Perturbation transaction edges are generated individually for each target tax evasion company node. This involves adding or deleting relevant transaction edges connected to the target tax evasion company node in the transaction network to simulate the behavior of the tax evasion company engaging in transactions or deliberately concealing transactions. Step 4: Identify the target tax evasion company nodes and obtain the false identification rate. Each target tax evasion company node is identified separately, and the proportion of incorrectly identified nodes to the total number of target tax evasion company nodes is the false identification rate.

3. The penetration testing method for identifying tax evasion according to claim 2, characterized in that, In Step 3 of Step 1), the specific process is as follows: 1-1) Obtain the graph neural network model of transaction network G = (A,X) and extract the embedded features of company network, where A is the adjacency matrix of the transaction network and X is the basic features of the companies selected by the tax experts from the rule base; 1-2) Train a graph neural network model f on the transaction network G. θ θ represents the network parameters, and the loss function for calculating the target tax evasion company node u is: y u It is the real label of node u, the target tax evasion company; 1-3) Calculate the loss function The gradient of the adjacency matrix A is obtained for any two nodes v in the transaction network G. i and v j gradient of the transaction edge between Where i,j∈{1,…,n} represent the index of the node, and n is the number of nodes; 1-4) Select the edge with the largest gradient among the possible connected transaction edges of the target tax evasion company node u and its potential transaction company node set, and perturb it, that is, add or delete this perturbation transaction edge.

4. The penetration testing method for identifying tax evasion according to claim 3, characterized in that, Based on whether the perturbation transaction edge originally existed.

5. The penetration testing method for identifying tax evasion according to claim 2, characterized in that, In Step 4 of Step 1), the specific process is as follows: 2-1) Retrain the graph neural network model on the transaction network after generating perturbation transaction edges for each target tax evasion company node to obtain node network embedding features; 2-2) Follow-up steps of the tax evasion detection algorithm; 2-3) Identify the target tax evasion company at this point and record whether the identification was correct; 2-4) After repeating the first three steps to identify all the target tax evasion company nodes, calculate the false identification rate.

6. The penetration testing method for identifying tax evasion according to claim 2, characterized in that, Step 2) specifically includes the following implementation steps: Step 1. Obtain the detection performance of the tax evasion detection algorithm. Obtain the tax evasion detection algorithm to be tested and the transaction network tax data it uses; use the tax evasion detection algorithm to identify company nodes in the transaction network; and obtain the identification accuracy and recall rate of the tax evasion detection algorithm in the test set. Step 2. Generate transaction edges to add or hide in the transaction network. Select some company nodes in the transaction network, and add or remove some specific transaction edges between them; Step 3. Obtain the performance degradation of the tax evasion detection algorithm after perturbation. The graph neural network model is retrained on the transaction network after adding or hiding transaction edges to obtain the node network embedding features. This is then used for subsequent steps of the tax evasion detection algorithm. The accuracy and recall of the tax evasion detection algorithm in the test set are obtained. The difference between the accuracy and recall of the algorithm and the previous tax evasion detection algorithm in the same test set is the performance degradation.

7. The penetration testing method for identifying tax evasion according to claim 6, characterized in that, Step 2) of Step 2, the specific process is as follows: 3-1) Obtain the transaction network G = (A,X) and extract the graph neural network embedding features of the company network, and train the graph neural network model f on the transaction network G. θ ; 3-2) Determine the maximum number of perturbation edges Δ for each transaction edge, which is set according to the specific transaction network and the security requirements of the tax evasion detection model; 3-3) Solve the following objective function using gradient descent to obtain the edge perturbation indicator vector s, where V is the taxpayer company node in the test set; s.t.,1 R s≤Δ,s∈[0,1] n 3-4) Processing the edge disturbance indicator vector s: First, filter the edge disturbance transactions represented by the edge disturbance indicator vector that conform to reality. If the business scope of the two end company nodes of the edge disturbance transaction represented by the edge disturbance indicator vector belongs to the same industry or to the upstream and downstream industries of the industrial chain, then the edge disturbance transaction is considered to conform to reality; otherwise, set its corresponding edge disturbance indicator variable si = 0. Then, if the edge disturbance indicator variable si > 0.5, set it to 1; otherwise, set it to 0. Finally, check whether the edge disturbance indicator vector s satisfies 1. T If s≤Δ, then some edge perturbation indicator variables si are randomly set to 0 until this condition is met. 3-5) Based on the edge perturbation indicator variable si with a value of 1, the perturbed transaction network is obtained by adding or deleting transaction edges in the transaction network.

8. The penetration testing method for identifying tax evasion according to claim 7, characterized in that, The maximum number of disturbed edges Δ is set to 5% of the total number of transaction edges in the transaction network.

Images