A method and device for identifying abnormal transaction behavior

Abstract

The application relates to an abnormal transaction behavior identification method and device, and relates to the technical field of information security, and solves the problems of high delay, large energy consumption and cross-institution data island of a traditional pure software scheme. The method comprises the following steps: converting transaction relationship data into a graph structure, wherein the nodes of the graph structure represent transaction parties, and the edges of the graph structure represent transactions; inputting an adjacency matrix of the graph structure into a spin-orbit torque magnetic random access memory, wherein the spin-orbit torque magnetic random access memory stores the weights of a graph neural network model; reading simulation convolution results generated by the spin-orbit torque magnetic random access memory from the adjacency matrix to obtain graph neural network features extracted from the graph structure by the graph neural network model; and inputting the graph neural network features into an abnormal transaction behavior evaluation model to obtain an evaluation result.

Description

Technical Field

[0001] This invention relates to the field of information security technology, and in particular to a method and apparatus for identifying abnormal transaction behavior. Background Technology

[0002] In recent years, with the acceleration of financial digitalization, abnormal transaction activities have become increasingly covert and complex. Traditional anti-abnormal transaction systems face problems such as slow processing speed, low identification accuracy, and difficulties in cross-institutional collaboration. Graph neural networks, due to their excellent graph structure modeling capabilities, have shown potential in abnormal transaction identification. However, traditional computing platforms cannot meet real-time requirements, and data security and trustworthiness are difficult to guarantee. Summary of the Invention

[0003] Based on the above analysis, embodiments of the present invention provide a method and apparatus for identifying abnormal transaction behavior, aiming to solve the problems of high latency, high energy consumption, and cross-institutional data silos in traditional pure software solutions.

[0004] In a first aspect, embodiments of the present invention provide a method for identifying abnormal transaction behavior, comprising: Transform transaction relationship data into a graph structure, where nodes in the graph structure represent transaction parties and edges represent transactions. The adjacency matrix of the graph structure is input into a spin orbital moment magnetic random access memory, wherein the spin orbital moment magnetic random access memory stores the weights of the graph neural network model. The simulated convolution result generated by the adjacency matrix through a spin-orbit moment magnetic random access memory is read to obtain the graph neural network features extracted from the graph structure by the graph neural network model; and The graph neural network features are input into the abnormal transaction behavior evaluation model to obtain the evaluation results.

[0005] Based on a further improvement of the above identification method, the weights of the graph neural network model are mapped through the differential resistance state of the two spin-orbit moment magnetic tunnel junctions in the spin-orbit moment magnetic random access memory.

[0006] Based on a further improvement of the above identification method, inputting the adjacency matrix of the graph structure into a spin-orbit moment magnetic random access memory includes: The adjacency matrix of the graph structure is converted into an analog voltage signal and input to the read bit line of the spin-orbit moment magnetic random access memory; The graph neural network features extracted from the graph structure by the graph neural network model are obtained by reading the simulated convolution result generated by the adjacency matrix through the spin orbit moment magnetic random access memory, including: The accumulated current output from the source line of the spin-orbit moment magnetic random access memory is read, and the output current is converted into a digital signal to obtain the graph neural network features.

[0007] Based on further improvements to the above identification method, the abnormal transaction behavior evaluation model includes a feature extraction module, a feature fusion module, and an output layer. The feature extraction module includes a convolutional neural network model and a recurrent neural network model. The convolutional neural network model is used to extract the spatial feature vector of the graph neural network features, and the recurrent neural network model is used to extract the temporal feature vector of the graph neural network features. The feature fusion module is used to fuse the outputs of the convolutional neural network model and the recurrent neural network model.

[0008] Based on a further improvement of the above recognition method, the feature fusion module is used to perform weighted fusion of the outputs of the convolutional neural network model and the recurrent neural network model, wherein the weight ratio of the output of the convolutional neural network model to the output of the recurrent neural network model is 0.7:0.3.

[0009] Based on the further improvement of the above recognition method, the convolutional neural network model uses spatial convolution kernels with dilation rates of 2 and 4 for convolution operations, respectively.

[0010] Based on a further improvement of the above identification method, the storage unit of the spin-orbit moment magnetic random access memory is a gating transistor and two spin-orbit moment magnetic tunnel junctions, and the two spin-orbit moment magnetic tunnel junctions share a spin-orbit moment write layer.

[0011] Further improvements to the above identification method also include: If the evaluation result indicates abnormal transaction behavior, a zero-knowledge proof is generated and submitted to the blockchain.

[0012] Based on the further improvement of the above recognition method, the weights of the graph neural network model are mapped as follows: The first spin-orbit magnetic tunnel junction is in a parallel state, and the second spin-orbit magnetic tunnel junction is in an antiparallel state, with a mapping weight of 1; The first spin-orbit magnetic tunnel junction is in an antiparallel state, the second spin-orbit magnetic tunnel junction is in an antiparallel state, and the mapping weight is 0; The first spin-orbit magnetic tunnel junction is in an antiparallel state, and the second spin-orbit magnetic tunnel junction is in a parallel state. The mapping weight is -1.

[0013] Secondly, embodiments of the present invention provide a device for identifying abnormal transaction behavior, comprising: The first conversion unit is used to convert transaction relationship data into a graph structure, wherein the nodes of the graph structure represent transaction parties and the edges of the graph structure represent transactions. The storage unit is used to input the adjacency matrix of the graph structure into the spin orbital moment magnetic random access memory, wherein the spin orbital moment magnetic random access memory stores the weights of the graph neural network model; The reading unit is used to read the simulated convolution result generated by the adjacency matrix through the spin-orbit moment magnetic random access memory, and to obtain the graph neural network features extracted by the graph neural network model from the graph structure; and An evaluation unit is used to input the graph neural network features into an abnormal transaction behavior evaluation model to obtain evaluation results.

[0014] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: (1) Realize real-time computing with ultra-low power consumption and high energy efficiency: The present invention adopts the in-memory computing architecture of spin orbital moment magnetic random access memory, which transforms the time-consuming matrix multiplication and addition operation in graph convolution into parallel current accumulation operation in the analog domain, thus fundamentally avoiding the data transfer overhead in the traditional von Neumann architecture.

[0015] (2) Providing a high-precision, multi-dimensional risk assessment model: The present invention innovatively integrates the spatial relationship modeling of graph neural networks, the local spatial pattern extraction of convolutional neural networks, and the temporal dynamic capture capability of recurrent neural networks to construct a risk assessment model. Through a weighted fusion strategy, the risk assessment model can simultaneously characterize the features of abnormal transactions in three dimensions: network structure, local clustering patterns, and temporal evolution, significantly improving the accuracy and recall rate of abnormal transaction behavior detection and reducing the false positive rate.

[0016] (3) Supports large-scale, scalable hardware deployment: Based on a differential cell structure of one gating transistor and two spin-orbit magnetic tunnel junctions and a distributed array design of spin-orbit magnetic random access memory, the present invention achieves a unity of high-density storage and high-parallel computing at the hardware level. This design provides a linearly scalable hardware foundation for processing the ever-growing financial transaction graph data.

[0017] (4) Optimization of detection sensitivity and system stability: By setting a high-risk response threshold and supplementing it with verifiable zero-knowledge proofs, the system can automatically, instantly, and reliably trigger and respond to extremely high-risk abnormal transaction suspicions while controlling the false alarm rate and avoiding system overload. This design takes into account both the sensitivity of the monitoring system and the stability of its operation, ensuring its practicality and reliability in real-world high-concurrency financial environments.

[0018] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0019] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 A flowchart illustrating a method for identifying abnormal transaction behavior according to an embodiment of the present invention is shown.

[0020] Figure 2 An example of a directed graph according to an embodiment of the present invention is shown.

[0021] Figure 3 A schematic diagram of the cell structure of the spin orbital moment magnetic random access memory in an embodiment of the present invention is shown.

[0022] Figure 4 A conceptual structural diagram of a spin-orbit moment magnetic random access memory according to an embodiment of the present invention is shown.

[0023] Figure 5 A schematic diagram of the structure of an abnormal transaction behavior evaluation model according to an embodiment of the present invention is shown.

[0024] Figure 6 A schematic diagram of the structure of an identification device for abnormal transaction behavior according to an embodiment of the present invention is shown. Detailed Implementation

[0025] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0026] According to one aspect of the present invention, a method for identifying abnormal transaction behavior is provided. Figure 1 A flowchart illustrating a method for identifying abnormal transaction behavior according to an embodiment of the present invention is shown. Figure 1 As shown, the methods for identifying this abnormal transaction behavior include: Step S101: Convert the transaction relationship data into a graph structure.

[0027] In this embodiment, the nodes of the graph structure represent the transacting parties, and the edges of the graph structure represent the transactions. For example, there are the following five transaction records:

[0028] Based on the above five transaction data points, we can construct as follows: Figure 2 The graph shown is a directed graph. A directed graph is converted into an adjacency matrix, where rows represent outgoing edges and columns represent incoming edges. If an edge has a weight, that weight is added: the matrix elements for A→B are 100, and the matrix elements for B→C are 50. Therefore... Figure 2 The adjacency matrix of the directed graph shown is as follows:

[0029] Step S102: Store the adjacency matrix of the graph structure into a spin-orbit moment magnetic random access memory.

[0030] Figure 3 A schematic diagram of the cell structure of the spin-orbit moment magnetic random access memory in an embodiment of the present invention is shown. Figure 3 As shown, the unit structure of this embodiment of the invention employs a single gate transistor ( Figure 3 (not shown) and two spin-orbit magnetic tunnel junctions ( Figure 3 The two spin-orbit magnetic tunnel junctions (MTJs) are oriented in opposite directions. The MTJ is a multi-layered stacked structure consisting of a heavy metal layer, a free layer, a barrier layer, a reference layer, an artificial ferromagnetic layer, and a top electrode. The heavy metal layer is made of tungsten (…). Figure 3 (represented by the character "W"), free layer ( Figure 3 The material represented by the character "FL" in the text is Co. 40 Fe 40 B 20 The barrier layer is made of MgO, and the reference layer ( Figure 3 The material represented by the character "RL" in the text is Co. 40 Fe 40 B 20 The artificial ferromagnetic layer is made of symmetrical artificial antiferromagnetic material (SAF), and the top electrode ( Figure 3 The material (represented by the character "top") is either titanium or ruthenium. Figure 3 In the middle, there is a multilayer stacked structure consisting of three columns of layers: heavy metal layer / free layer / barrier layer / reference layer / artificial ferromagnetic layer / top electrode. The middle column is a spin-orbit magnetic tunnel junction, and the left and right columns are electrodes. Figure 3(Represented by the character "E" in Chinese). The writing principle of a spin-orbit-moment (MTJ) magnetic tunnel junction is based on the spin-orbit-moment effect, that is, when the writing current flows through the heavy metal layer, a vertical spin current is generated, thereby driving the magnetic moment of the free layer to flip, and the direction of the magnetic moment flip is controlled by the polarity of the applied current. The readout principle of a MTJ is based on the tunneling magnetoresistance effect, that is, the tunneling magnetoresistance ratio between the high and low resistance states of the MTJ can reach about 150%, and its resistance state can be determined by sensing the read current.

[0031] Figure 4 A conceptual structural diagram of a spin-orbit moment magnetic random access memory according to an embodiment of the present invention is shown. Figure 4 As shown, the spin-orbit-moment magnetic random access memory (SOM) in this embodiment of the invention includes a spin-orbit-moment magnetic tunnel junction array based on a gating transistor and two spin-orbit-moment magnetic tunnel junction unit structures, a read / write separated word line and bit line control architecture, and peripheral read circuitry. The read word line of the SOM is labeled RWL, the read bit line is labeled RBL, the write word line is labeled WWL, the write bit line is labeled WBL, and the source line is labeled SL. When the write enable signal is 1, a write pulse (e.g., pulse width less than 2 nanoseconds) signal can be applied to the SOM to store the weights of the graph neural network model. This write current flows through the shared spin-orbit write layer (i.e., the heavy metal layer and the top electrode) of the two spin-orbit-moment magnetic tunnel junction unit structures. The spin Hall effect generated by the current produces spin accumulation in the spin-orbit write layer, which then acts on the adjacent free layer magnetic moment through spin-orbit torque, inducing it to undergo magnetization reversal. In this embodiment, the spin-orbit moment magnetic random access memory can reach a size of 1024*1024 cell structure, supporting 10 9 Address linear expansion. When the read enable signal is 1, a read pulse signal (e.g., pulse width less than 2 nanoseconds) can be applied to the spin-orbit magnetic random access memory to input the adjacency matrix of the graph structure. Due to the conductance characterization weight of the spin-orbit magnetic tunnel junction, according to Kirchhoff's laws, the read current generated by this read voltage is the graph neural network feature extracted from the graph structure.

[0032] like Figure 4As shown, the weights of the pre-trained model are written to the spin-orbit moment magnetic random access memory (MORAM) by gating WWL, WBL, and SL. Each set of MTJ resistance values ​​represents a weight value. Using the adjacency matrix as the model input, the system gated the RWL column corresponding to the convolution calculation weights, and input a set of voltage pulses (i.e., the adjacency matrix) in parallel to each row of the MORAM via RBL. Then, the accumulated current is read from SL. Due to Kirchhoff's laws, this constitutes a convolution operation. Subsequently, the ADC reads the convolution result, and the system saves the result for subsequent calculations. The magnitude of the voltage pulse represents the magnitude of the value in the adjacency matrix.

[0033] In this embodiment, the spin-orbit moment magnetic random access memory stores the weights of the graph neural network model. Graph Neural Network (GNN) is an algorithmic framework based on deep learning for processing graph-structured data. It performs tasks such as classification, prediction, and generation by extracting features from nodes, edges, and the overall graph.

[0034] In this embodiment, the weights of the graph neural network model are mapped through the differential resistive states of the two spin-orbit-moment magnetic tunnel junctions in the spin-orbit-moment magnetic random access memory. A single weight of the graph neural network model can be obtained from... Figure 4 The differential resistive state mapping of the two MTJs in the shared spin orbital moment write layer is shown in the figure.

[0035] For example, the first spin-orbit magnetic tunneling junction is in a parallel state, and the second spin-orbit magnetic tunneling junction is in an antiparallel state, mapped to a weight of "1"; the first spin-orbit magnetic tunneling junction is in an antiparallel state, and the second spin-orbit magnetic tunneling junction is in an antiparallel state, mapped to a weight of "0"; the first spin-orbit magnetic tunneling junction is in an antiparallel state, and the second spin-orbit magnetic tunneling junction is in a parallel state, mapped to a weight of "-1". Alternatively, the first spin-orbit magnetic tunneling junction can be specified as being in a parallel state, and the second spin-orbit magnetic tunneling junction as being in an antiparallel state, mapped to a weight of "-1"; the first spin-orbit magnetic tunneling junction is in a parallel state, and the second spin-orbit magnetic tunneling junction is in a parallel state, mapped to a weight of "0"; the first spin-orbit magnetic tunneling junction is in an antiparallel state, and the second spin-orbit magnetic tunneling junction is in a parallel state, mapped to a weight of "1".

[0036] Step S103: Read the simulated convolution result generated by the adjacency matrix through the spin orbit moment magnetic random access memory to obtain the graph neural network features extracted by the graph neural network model from the graph structure.

[0037] In this embodiment, graph convolution operations of the graph convolutional neural network model can be directly executed on the spin-orbit-moment magnetic random access memory, achieving integration of computation and storage. The adjacency matrix of the graph structure is converted into an analog voltage signal and applied to the read lines of the spin-orbit-moment magnetic random access memory. According to Ohm's law and Kirchhoff's laws, the total output current I_out of each column of source lines is the sum of the product of the input voltage and the conductance of the MTJ cell at the crossover point:

[0038] This operation physically completes the graph convolution operation in one step. The peripheral circuitry of the spin-orbit moment magnetic random access memory (including a sensitive amplifier (SA) and an analog-to-digital converter (ADC)) can read and quantize the analog current output by the column bit line to obtain the graph neural network features.

[0039] Step S104: Input the graph neural network features into the abnormal transaction behavior evaluation model to obtain the evaluation results.

[0040] In this embodiment, the abnormal transaction behavior assessment model can perform a deeper, multi-dimensional analysis of graph neural network features to generate the final comprehensive probability of abnormal transaction behavior.

[0041] In this embodiment, the abnormal trading behavior evaluation model needs to be pre-trained. During pre-training, the abnormal trading behavior evaluation model first performs forward propagation to calculate the output and loss, then backpropagates the loss gradient back to each layer of the abnormal trading behavior evaluation model and the fusion weights, and finally uses an optimizer (such as Adam) to update all parameters. Training the abnormal trading behavior evaluation model on large-scale data first to obtain initial weights accelerates convergence and improves generalization ability.

[0042] In some embodiments, the abnormal transaction behavior assessment model includes a feature extraction module, a feature fusion module, and an output layer. The feature extraction module includes a convolutional neural network model and a recurrent neural network model. The convolutional neural network model is used to extract spatial feature vectors of graph neural network features, and the recurrent neural network model is used to extract temporal feature vectors of graph neural network features. The feature fusion module is used to fuse the outputs of the convolutional neural network model and the recurrent neural network model.

[0043] Figure 5 A schematic diagram of the structure of an abnormal transaction behavior assessment model according to an embodiment of the present invention is shown. Figure 5As shown, this abnormal trading behavior assessment model employs a fusion of convolutional neural network (CNN) and recurrent neural network (RNN) operations. Its architecture comprises two branches: the CNN branch uses dilated convolutional kernels with dilation rates of d=2 and d=4, aiming to expand the receptive field to effectively capture multi-hop relationships in trading data, and its output is a vector feature representing the local topological structure; the RNN branch utilizes gated recurrent units (GRUs) to track and model the temporal evolution of trading behavior features. Finally, a fusion layer weights and sums the output features of the two branches, and activates them via a sigmoid function to output the final risk probability.

[0044] In some embodiments, the weight ratio of the output of the convolutional neural network model to the output of the recurrent neural network model is 0.7:0.3.

[0045] In some embodiments, Figure 1 The method also includes: if the evaluation result indicates abnormal transaction behavior, then generating a zero-knowledge proof and submitting it to the blockchain.

[0046] In this embodiment, a risk response threshold can be set (e.g., 0.92). When the probability of an abnormal transaction exceeds the risk response threshold, it is determined to be a highly suspicious event. To protect business privacy and achieve verifiability, a zero-knowledge proof is generated and submitted to the blockchain. The zero-knowledge proof verifies that "the calculation process is correct and the risk score does indeed exceed the threshold without disclosing specific transaction details and internal model parameters." Subsequently, this proof, along with event digest hash, timestamp, and other metadata, is submitted to the blockchain network as a transaction for notarization and alert distribution.

[0047] Figure 6 A schematic diagram of the structure of an abnormal transaction behavior identification device according to an embodiment of the present invention is shown. Figure 6 As shown, the abnormal transaction behavior identification device of this embodiment includes: a first conversion unit 601, used to convert transaction relationship data into a graph structure, wherein the nodes of the graph structure represent transaction parties and the edges of the graph structure represent transactions; a storage unit 602, used to input the adjacency matrix of the graph structure into a spin orbital moment magnetic random access memory, wherein the spin orbital moment magnetic random access memory stores the weights of a graph neural network model; a reading unit 603, used to read the simulated convolution result generated by the adjacency matrix through the spin orbital moment magnetic random access memory to obtain the graph neural network features extracted by the graph neural network model from the graph structure; and an evaluation unit 604, used to input the graph neural network features into an abnormal transaction behavior evaluation model to obtain an evaluation result.

[0048] It should be understood that Figure 6The devices described in the document contain various modules and references. Figure 1 The steps described in the method correspond to each other. Therefore, the operations and features described above for the method also apply to... Figure 6 The devices and modules contained herein will not be described in detail here.

[0049] Compared with the prior art, the embodiments of the present invention can achieve at least one of the following beneficial effects: (1) Realize real-time computing with ultra-low power consumption and high energy efficiency: The present invention adopts the in-memory computing architecture of spin orbital moment magnetic random access memory, which transforms the time-consuming matrix multiplication and addition operation in graph convolution into parallel current accumulation operation in the analog domain, thus fundamentally avoiding the data transfer overhead in the traditional von Neumann architecture.

[0050] (2) Providing a high-precision, multi-dimensional risk assessment model: The present invention innovatively integrates the spatial relationship modeling of graph neural networks, the local spatial pattern extraction of convolutional neural networks, and the temporal dynamic capture capability of recurrent neural networks to construct a risk assessment model. Through a weighted fusion strategy, the risk assessment model can simultaneously characterize the features of abnormal transactions in three dimensions: network structure, local clustering patterns, and temporal evolution, significantly improving the accuracy and recall rate of abnormal transaction behavior detection and reducing the false positive rate.

[0051] (3) Supports large-scale, scalable hardware deployment: Based on a differential cell structure of one gating transistor and two spin-orbit magnetic tunnel junctions and a distributed array design of spin-orbit magnetic random access memory, the present invention achieves a unity of high-density storage and high-parallel computing at the hardware level. This design provides a linearly scalable hardware foundation for processing the ever-growing financial transaction graph data.

[0052] (4) Optimization of detection sensitivity and system stability: By setting a high-risk response threshold and supplementing it with verifiable zero-knowledge proofs, the system can automatically, instantly, and reliably trigger and respond to extremely high-risk abnormal transaction suspicions while controlling the false alarm rate and avoiding system overload. This design takes into account both the sensitivity of the monitoring system and the stability of its operation, ensuring its practicality and reliability in real-world high-concurrency financial environments.

[0053] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying abnormal transaction behavior, characterized in that, include: Transform transaction relationship data into a graph structure, where nodes in the graph structure represent transaction parties and edges represent transactions. The adjacency matrix of the graph structure is input into a spin orbital moment magnetic random access memory, wherein the spin orbital moment magnetic random access memory stores the weights of the graph neural network model. The graph neural network features extracted from the graph structure by the graph neural network model are obtained by reading the simulated convolution result generated by the adjacency matrix through the spin orbit moment magnetic random access memory. The graph neural network features are then input into the abnormal transaction behavior evaluation model to obtain the evaluation results.

2. The identification method according to claim 1, characterized in that, The weights of the graph neural network model are mapped through the differential resistance state of the two spin-orbit magnetic tunnel junctions in the spin-orbit magnetic random access memory.

3. The identification method according to claim 1, characterized in that, Inputting the adjacency matrix of the graph structure into the spin orbit moment magnetic random access memory includes: The adjacency matrix of the graph structure is converted into an analog voltage signal and input to the read bit line of the spin-orbit moment magnetic random access memory; The graph neural network features extracted from the graph structure by the graph neural network model are obtained by reading the simulated convolution result generated by the adjacency matrix through the spin orbit moment magnetic random access memory, including: The accumulated current output from the source line of the spin-orbit moment magnetic random access memory is read, and the output current is converted into a digital signal to obtain the graph neural network features.

4. The identification method according to claim 1, characterized in that, The abnormal transaction behavior assessment model includes a feature extraction module, a feature fusion module, and an output layer. The feature extraction module includes a convolutional neural network model and a recurrent neural network model. The convolutional neural network model is used to extract the spatial feature vector of the graph neural network features, and the recurrent neural network model is used to extract the temporal feature vector of the graph neural network features. The feature fusion module is used to fuse the outputs of the convolutional neural network model and the recurrent neural network model.

5. The identification method according to claim 4, characterized in that, The feature fusion module is used to perform weighted fusion of the outputs of the convolutional neural network model and the recurrent neural network model, wherein the weight ratio of the output of the convolutional neural network model to the output of the recurrent neural network model is 0.7:0.

3.

6. The identification method according to claim 5, characterized in that, The convolutional neural network models use spatial convolutional kernels with dilation rates of 2 and 4 for convolution operations, respectively.

7. The identification method according to claim 1, characterized in that, The storage cell of the spin-orbit-moment magnetic random access memory consists of a gating transistor and two spin-orbit-moment magnetic tunnel junctions, with the two spin-orbit-moment magnetic tunnel junctions sharing a spin-orbit write layer.

8. The identification method according to claim 1, characterized in that, Also includes: If the evaluation result indicates abnormal transaction behavior, a zero-knowledge proof is generated and submitted to the blockchain.

9. The identification method according to claim 2, characterized in that, The weights of the graph neural network model are mapped as follows: The first spin-orbit magnetic tunnel junction is in a parallel state, and the second spin-orbit magnetic tunnel junction is in an antiparallel state, with a mapping weight of 1; The first spin-orbit magnetic tunnel junction is in an antiparallel state, the second spin-orbit magnetic tunnel junction is in an antiparallel state, and the mapping weight is 0; The first spin-orbit magnetic tunnel junction is in an antiparallel state, and the second spin-orbit magnetic tunnel junction is in a parallel state. The mapping weight is -1.

10. A device for identifying abnormal transaction behavior, characterized in that, include: The first conversion unit is used to convert transaction relationship data into a graph structure, wherein the nodes of the graph structure represent transaction parties and the edges of the graph structure represent transactions. The storage unit is used to input the adjacency matrix of the graph structure into the spin orbital moment magnetic random access memory, wherein the spin orbital moment magnetic random access memory stores the weights of the graph neural network model; The reading unit is used to read the simulated convolution result generated by the adjacency matrix through the spin-orbit moment magnetic random access memory, and to obtain the graph neural network features extracted by the graph neural network model from the graph structure; and An evaluation unit is used to input the graph neural network features into an abnormal transaction behavior evaluation model to obtain evaluation results.

Images