A cross-institutional fraud detection method based on federated graph learning for privacy protection

Abstract

The application belongs to the technical field of financial security, and discloses a cross-institutional fraud detection method based on federated graph learning and oriented to privacy protection, which comprises the following steps: each participant constructs a transaction graph according to the transaction text description of the included account, and performs enhancement processing on the transaction graph to obtain an enhanced subgraph; each participant iteratively trains a graph Transformer using the enhanced subgraph until a local training end condition is reached, and uploads local model parameters of the graph Transformer to a central server based on a differential privacy mechanism; the central server aggregates the local model parameters of all participants to obtain global model parameters, and distributes the global model parameters to each participant; the global model parameters serve as initial model parameters for local training of each participant, and the federated training is repeated until a federated training end condition is reached. The application is more robust to cross-institutional, heterogeneous and sparse sample scenarios, and can reduce false negatives and false positives.

Description

Technical Field

[0001] This invention belongs to the field of financial security technology, specifically relating to a privacy-preserving cross-institutional fraud detection method based on federated graph learning. Background Technology

[0002] Traditional fraud detection methods primarily rely on rule-based systems and manual auditing, which depend on preset thresholds and expert knowledge. With the development of artificial intelligence, various machine learning methods such as support vector machines, random forests, and linear regression have emerged to improve fraud detection capabilities. However, these methods still have certain limitations, including insufficient adaptability to evolving fraud patterns, high false positive rates, and an inability to capture complex temporal dependencies within transaction sequences.

[0003] In recent years, deep learning technologies, represented by graph neural networks (GNNs), have demonstrated stronger fraud detection capabilities by enhancing feature learning and pattern recognition. For example, existing research has strengthened the message passing mechanism of GNNs by fusing edge and node features, thereby effectively capturing complex fraud patterns. Alternatively, by combining local and global memory networks through interactive aggregation and hierarchical information fusion, the problem of existing GNN methods neglecting global information has been alleviated, thus improving fraud detection performance. Although these methods have achieved good fraud detection results, their centralized paradigm has failed to solve the data silo problem in real-world scenarios. Recent research has used federated graph learning to address the data silo problem in the financial industry, achieving fraud detection in a distributed environment; however, these methods have not fully utilized the potential information contained in the graph structure.

[0004] Graph Neural Network (GNN) based methods, especially Graph Transformers, have opened up new avenues for efficient fraud detection. These methods naturally model the transaction relationships between accounts as a graph structure, where nodes represent financial accounts and edges represent transactions. Graph Transformers excel in graph representation learning, making them particularly suitable for financial fraud detection tasks. Theoretically, their expressive power can be defined by second-order invariant graph networks, and in experiments, they significantly outperform traditional message-passing GNNs. This is because real-world financial transaction data itself exhibits complex relational characteristics, which often cannot be fully represented by simple undirected graphs or one-sided structures.

[0005] However, applying existing graph transformer methods to financial fraud detection still faces two major challenges. First, mainstream graph transformers primarily focus on node-centric representations, failing to fully utilize multi-dimensional edge features. In fact, unlike other financial tasks, the key information for fraud detection lies mainly in edge features rather than node features, because the relational patterns within transaction edges contain signals that can identify fraudulent behavior. Second, due to strict privacy constraints, each financial institution operates as an independent data silo, making it difficult to directly apply existing graph transformers to effectively utilize cross-institutional information.

[0006] Furthermore, federated learning, as a distributed machine learning paradigm, enables privacy-preserving collaborative modeling and is therefore widely used in data-sensitive fields such as finance. For example, existing research has proposed a longitudinal federated learning method suitable for multivariate logistic regression, which can achieve credit scoring while protecting data privacy. A credit scoring model called FedKT has also been proposed, which significantly improves model effectiveness and robustness by integrating knowledge transfer techniques with the federated learning framework. However, unlike the tabular data used in credit scoring, graph-structured fraud detection data has complex relationships, making it difficult to capture cross-institutional patterns in federated learning; therefore, specialized federated graph learning methods are needed.

[0007] Federated graph learning can be divided into federated graph learning and subgraph federated learning based on the nature of the graph task. The former deals with independent graphs such as molecules, while the latter deals with partitioned subgraphs of the global graph, facing challenges such as missing links between subgraphs and insufficient information utilization. To address the problem of missing connections between subgraphs, existing research has integrated missing neighbor generation into the federated graph learning framework, restoring the structural integrity of subgraphs by filling in potential external connections. To fully utilize global graph information, a personalized federated graph Transformer has also been proposed, which effectively mines cross-subgraph information and enhances global learning capabilities by fusing learnable supernetworks and linear Transformers. However, these methods either mainly focus on low-hop neighborhoods, failing to capture long-range dependencies and high-order structural patterns between multi-hop neighbor nodes; or they overemphasize global information while neglecting the connectivity between subgraphs, making them unsuitable for application to graph structures representing transaction relationships between accounts. Summary of the Invention

[0008] The purpose of this invention is to provide a privacy-preserving cross-institutional fraud detection method based on federated graph learning, which is more robust to cross-institutional, heterogeneous, and sparse sample scenarios and can reduce false negatives and false positives.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A privacy-preserving cross-institutional fraud detection method based on federated graph learning includes the following steps:

[0011] Each participant constructs a transaction graph based on the transaction text descriptions of the included accounts, and then enhances the transaction graph to obtain an enhanced subgraph;

[0012] Each participant uses the enhanced subgraph to iteratively train the graph Transformer until the local training termination condition is met. Based on the differential privacy mechanism, the local model parameters of the graph Transformer are uploaded to the central server. The normalized attention score of the graph Transformer introduces edge attribute attention bias and shared bias, and the multi-head attention operator of the graph Transformer introduces an edge-based gating mechanism for execution.

[0013] The central server aggregates the local model parameters of all participants to obtain the global model parameters, and then distributes the global model parameters to each participant. The global model parameters serve as the initial model parameters for each participant's local training. The federated training is iterated repeatedly until the federated training termination condition is met.

[0014] Each participant uses the graph Transformer trained in a federated manner to process the augmented subgraph to be detected, and obtains the node features of each node and the edge features between nodes in the augmented subgraph to be retrieved.

[0015] The node and edge features output from the last layer of the Transformer are input into the classification component to obtain the anomaly score of all edges. Transactions corresponding to edges with anomaly scores greater than the anomaly threshold are judged to have fraud risk, and a fraud warning is generated.

[0016] Several alternative methods are provided below, but they are not intended as additional limitations on the overall solution above. They are merely further additions or optimizations. Provided there are no technical or logical contradictions, each alternative method can be combined individually with respect to the overall solution above, or multiple alternative methods can be combined with each other.

[0017] Preferably, the enhancement process includes reverse message passing, port numbering, and master node identification.

[0018] Preferably, the edge attribute attention bias is the product of the transpose of the edge feature matrix of the transaction graph input by the current participant at the current layer and the local edge weight matrix, wherein the transaction graph input by the current participant at the first layer is the enhanced subgraph.

[0019] Preferably, the edge-based gating mechanism includes:

[0020] The transpose of the edge feature matrix of the transaction graph input by the current participant at the current layer is multiplied by the local gate control weight matrix, and then processed by the sigmoid function to output the edge gate control vector.

[0021] Preferably, the normalized attention score of the graph Transformer incorporates edge attribute attention bias and shared bias, including:

[0022] Summing the edge attribute attention bias and the shared bias yields the fused attention bias;

[0023] The unnormalized attention score is calculated by superimposing the fusion attention bias on the dot product of the query vector and the key vector.

[0024] The unnormalized attention score is processed using the softmax function to obtain the normalized attention score.

[0025] Preferably, the multi-head attention operator of the graph Transformer is implemented using an edge-based gating mechanism, including:

[0026] The normalized attention score, the node features of neighboring nodes, the value weight matrix, and the edge gating vector in the edge-based gating mechanism are multiplied by a dot product, and the result of the dot product is used as the node output feature of a single attention head.

[0027] Aggregate the output features of each individual attention head corresponding to all neighboring nodes, and obtain the updated node features of the central node through the node output projection matrix;

[0028] The unnormalized attention scores between nodes are used as the edge output features of a single attention head. The edge output features of all attention heads are concatenated, and the updated edge features between nodes are obtained through the edge output projection matrix.

[0029] Preferably, the step of uploading the local model parameters of the graph Transformer to the central server based on the differential privacy mechanism includes:

[0030] For the current federated training, calculate the model update amount between the local model parameters and the global model parameters issued by the central server;

[0031] The model update amount is pruned and differential privacy noise is added to obtain local privacy model parameters, which are then uploaded to the central server.

[0032] Preferably, the node features and edge features output from the last layer graph Transformer are input into the classification component to obtain anomaly scores for all edges, including:

[0033] The classification component includes a feedforward network and a sigmoid activation function. For any edge, the edge features and the corresponding two node features are concatenated, and the anomaly score of the edge is obtained through the feedforward network and the sigmoid activation function.

[0034] This invention addresses the shortcomings of existing graph Transformer methods, which primarily rely on node features and underutilize transaction edge features, as well as the data silos among financial institutions and the difficulty in directly sharing raw data. It proposes a privacy-preserving cross-institutional fraud detection method based on federated graph learning. This method introduces an edge-based gating mechanism and edge attribute attention bias during message passing using a hierarchical edge-attribute Transformer, enabling fine-grained modeling of key transaction relationships and abnormal transaction patterns. Through directed multigraph augmentation, it combines reverse message passing, port numbering, and self-identification to enhance the structural expressiveness of the transaction graph. Simultaneously, within a horizontal federated learning framework, it shares only bias parameters and incorporates differential privacy noise, achieving a balance between cross-institutional knowledge collaboration and privacy protection. Compared to existing technologies, this method offers the following advantages:

[0035] (1) It makes fuller use of transaction edge information and is better able to identify complex fraud relationships than existing methods that only look at node characteristics.

[0036] (2) It can perform collaborative testing by multiple institutions without exchanging original data, thus overcoming the limitation of data silos.

[0037] (3) By sharing only the bias parameters and using differential privacy protection, the detection effect, communication efficiency and privacy security are balanced.

[0038] (4) It is more robust to cross-institutional, heterogeneous and sparse sample scenarios, and can reduce false alarms and false alarms. Attached Figure Description

[0039] Figure 1 This is a flowchart of the privacy-preserving cross-agency fraud detection method based on federated graph learning according to the present invention.

[0040] Figure 2 This is a schematic diagram of the directed multigraph enhancement of the present invention;

[0041] Figure 3 This is a diagram illustrating the overall framework of the privacy-preserving cross-agency fraud detection method based on federated graph learning of the present invention.

[0042] Figure 4 This is a flowchart of the hierarchical edge attribute Transformer of the present invention;

[0043] Figure 5 The figure shows the ablation experimental results of the edge mechanism in the experiment of this invention;

[0044] Figure 6 This is a graph showing the training curve results of the shared bias strategy in the experiments of this invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0047] Fraud detection is crucial for maintaining the stability of the financial system. Given the networked and interconnected nature of financial fraud, graph neural network-based methods, especially graph Transformers, have demonstrated significant advantages in fraud detection. However, applying existing graph Transformer methods to fraud detection still faces two major challenges: firstly, insufficient utilization of edge features in financial transaction graphs; and secondly, difficulty in capturing cross-institutional connections in privacy-preserving scenarios and under the assumption of data silos.

[0048] To address the aforementioned issues, this embodiment proposes a privacy-preserving cross-institutional fraud detection method based on federated graph learning (a novel federated graph Transformer model, abbreviated as Fraud-FedGT). This method utilizes edge features in financial transaction graphs to achieve cross-institutional fraud detection in privacy-preserving scenarios. Specifically, to effectively utilize edge features, this embodiment proposes a hierarchical edge-attribute Transformer (HEAT), which enhances the model's recognition of key transaction features by employing edge-based gating mechanisms and edge-attribute attention biases. Furthermore, this embodiment proposes a directed multi-graph augmentation method to improve the representation quality of financial transaction graphs. To capture cross-institutional fraud patterns, this embodiment constructs a lateral federated learning framework with shared biases to explore complementary information between financial transaction graphs of different institutions. Finally, this embodiment also employs differential privacy technology to add noise to uploaded parameters to further protect user privacy.

[0049] like Figure 1 As shown, this embodiment provides a privacy-preserving cross-institutional fraud detection method based on federated graph learning, including the following steps:

[0050] Step 1: Each participant constructs a transaction graph based on the transaction text descriptions of the included accounts, and enhances the transaction graph to obtain an enhanced subgraph.

[0051] Financial transaction graphs are typically modeled as directed multigraphs. , where the node set Include Each element represents an account; edge set The transaction relationships between accounts are defined, and multiple edges can exist between any two nodes. Node feature matrix. Characterizing the features of each account node, For nodes The node features, and the edge feature matrix Characterizing the features of the transaction edge; among which, These represent the dimensions of the node features and edge features, respectively. Furthermore, for connecting nodes... and The edges are adopted This represents the corresponding feature vector.

[0052] Financial transaction graphs typically employ directed multigraph structures that contain rich information. However, traditional message-passing graph neural networks lack the ability to effectively utilize this information. To address these limitations, this embodiment proposes a directed multigraph augmentation method, transforming a message-passing graph neural network into a more efficient directed multigraph neural network. Directed multigraph augmentation enables the model to learn more expressive feature representations, thereby improving the effectiveness of fraud detection. Figure 2 As shown, the directed multigraph augmentation consists of three main parts: reverse messaging, port numbering, and master node identification. These components significantly enrich the information contained in the node and edge features, enabling the model to more effectively identify various complex fraud patterns, including scatter-cluster patterns, cyclic patterns, and stacking patterns.

[0053] Figure 2 Showing from the account to account A transaction and its sampling neighborhood were analyzed, and this neighborhood was enhanced with features through reverse messaging, port numbers, and master node identifiers. The sampled neighborhood, after directed multigraph enhancement, can clearly identify fraud patterns. Within the port numbers, ordered pairs... Indicates from node To the node The directed edges. In the master node identifier, the central node is marked as... All its neighboring nodes are marked as .

[0054] In this embodiment, reverse message passing is used to achieve bidirectional information exchange. Traditional graph neural networks only propagate messages along the edge direction, which prevents nodes without incoming edges from utilizing the graph topology for feature optimization. This allows such nodes to receive neighbor information and distinguish the direction of edges through the direction encoding in the edge features.

[0055] Port numbering: Distinguishing between connections from the same neighbor and different neighbors in a directed graph can be challenging. This embodiment uses port numbering to assign local identifiers to the neighbors of each node to detect complex fraud patterns. Each directed edge is assigned inbound and outbound port identifiers, and edges sharing the same source or destination node are assigned matching identifiers.

[0056] Master Node Identification: While reverse message passing and port numbers help detect more anomalous patterns, they are still insufficient to identify directed cycles. Therefore, this embodiment employs self-identification, facilitating cycle identification in the graph structure by assigning a unique binary tag to the central node, thereby enabling identification when messages propagate back to the starting node.

[0057] Step 2: Each participant uses the enhanced subgraph to iteratively train the graph Transformer until the local training termination condition is met. Based on the differential privacy mechanism, the local model parameters of the graph Transformer are uploaded to the central server. The normalized attention score of the graph Transformer introduces edge attribute attention bias and shared bias, and the multi-head attention operator of the graph Transformer introduces an edge-based gating mechanism for execution.

[0058] Graph Transformer can extract effective structural representations by analyzing the feature relationships and dependencies within the node feature matrix. Specifically, Graph Transformer consists of... The system consists of layers, each composed of a multi-head attention (MHA) module and a feedforward network (FFN). In the first... layer( The MHA module processes hidden representations through linear transformations. (in Generate a query matrix. Key matrix Sum matrix ,in Indicates the attention head index, This represents the attention head dimension.

[0059] Subsequently, the Graph Transformer calculates attention weights by scaling dot product operations to aggregate neighborhood information, and performs a nonlinear transformation on the attention output through a feedforward network. Furthermore, the model introduces residual connections and layer normalization mechanisms to stabilize the training process.

[0060] like Figure 3 As shown, the Fraud-FedGT proposed in this paper adopts a horizontal federated learning framework, in which the locally deployed HEAT model is generated by... The layer encoder module is composed of layers. Specifically, HEAT learns node features through a graph attention mechanism and learns edge features using an edge-based gating mechanism and an edge attribute attention bias mechanism. In addition, before inputting graph data into HEAT, it needs to be preprocessed by applying directed multigraph augmentation.

[0061] In Fraud-FedGT, the workflow begins with the server initializing model parameters (including global model parameters), while each participant prepares local transaction graph data (subgraphs). The raw transaction text data includes payer account ID, payee account ID, transaction amount, transaction time, transaction type, currency, bank or institution, and transaction tags. The process of constructing a transaction graph based on the raw data includes: cleaning invalid transaction records, such as those with missing account IDs, missing amounts, or abnormal timestamps; normalizing or logarithmically transforming continuous features such as transaction amounts; encoding discrete features such as transaction type, currency, and institution number; dividing the local subgraph by financial institution or data holder; converting transaction records into edge feature matrices; and converting account information or aggregated statistics into node feature matrices.

[0062] Then, directed multigraph augmentation is used to augment the subgraph to expand the sampling neighborhood. Subsequently, each participant performs HEAT-based local training on the augmented graph to capture cross-institutional fraud patterns, and optimizes the local model using binary cross-entropy loss. Next, the updated local parameters are uploaded to the server, and differential privacy is applied to prevent the leakage of sensitive information. Finally, the server aggregates the updates from all parties through weighted fusion. This iterative process of local training and global aggregation enables cross-institutional knowledge sharing until the global model converges to a stable state, at which point each participant obtains a local model incorporating global fraud detection knowledge.

[0063] To enable the local model to effectively utilize information from the financial transaction graph, this embodiment improves the original graph Transformer by proposing a hierarchical edge attribute Transformer. For example... Figure 4 As shown, the hierarchical edge-attribute Transformer not only employs a graph attention mechanism to process node-level information, but also fully integrates edge features with edge-attribute attention biases through an edge-based gating mechanism. Furthermore, shared biases from the lateral federated learning framework are transferred to HEAT to help the local model learn supplementary information from other mechanisms.

[0064] Graph attention mechanisms are a fundamental component that enables Transformer models to selectively focus on key graph regions. In Fraud-FedGT, this is achieved through nodes... The features and corresponding edge features are used to calculate the attention score for adjacent nodes. The neighborhood masking strategy restricts the attention calculation to the range of direct connections, thereby facilitating the extraction of local transaction patterns. In this embodiment, a graph attention mechanism is applied under the federated learning framework. The attention bias is superimposed on the dot product of the query vector and the key vector to calculate the unnormalized attention score. The calculation process is shown in formula (1). Then, the unnormalized attention score is processed by the softmax function to obtain the normalized attention score. The calculation process is shown in formula (2).

[0065] (1)

[0066] (2)

[0067] in, Indicates the first Participants under individual attention No. Layer nodes The unnormalized attention score between them; Indicates the first Participants under individual attention No. Layer nodes The query vector; Indicates the first Participants under individual attention No. Layer nodes The key vector; It represents the Hadamardi (or Hadama) stack; Indicates the first Participants under individual attention No. Layer nodes Fusion attention bias; It is the first Participants under individual attention No. Layer nodes Normalized attention score, The dimension of the key feature. This refers to the softmax function.

[0068] Subsequently, multiple attention heads simultaneously calculate the scaled dot product. The calculation process for nodes is shown in Equation (3). The normalized attention score, the node features of neighboring nodes, the value weight matrix, and the edge gating vector in the edge-based gating mechanism are multiplied by a dot product. The result of the dot product is used as the node output feature of a single attention head. The output features of the single attention heads corresponding to all neighboring nodes are aggregated, and the updated node features of the center node are obtained through the node output projection matrix. The calculation process for edges is shown in Equation (4). The unnormalized attention score between nodes is used as the edge output feature of a single attention head. The edge output features of all attention heads are concatenated, and the updated edge features between nodes are obtained through the edge output projection matrix.

[0069] (3)

[0070] (4)

[0071] in, Indicates a multi-head attention operator; This represents the connection of all attention heads; Indicates the central node Add the elements of all adjacent nodes; and Representing the participating parties No. Layer input nodes Node representation (i.e., participants) No. Layer output nodes (node ​​representation). Represents the value weight matrix; It is a connection node The side-gating vector; Indicates the participating parties No. Layer input nodes Edge characteristics between (i.e., the participants) No. Layer output nodes (Edge features between) Representing the participating parties No. The node output projection matrix and edge output projection matrix of the layer.

[0072] Edge-based gating and shared bias. To enable the local model to fully integrate edge features and supplementary information from other institutions, this paper adopts an edge-based gating mechanism, edge attribute attention bias, and shared bias. The edge-based gating mechanism selectively propagates valuable information to identify important features of financial transactions. Its calculation process is shown in Equation (5).

[0073] (5)

[0074] in, It is the sigmoid function; This represents the local control weight matrix.

[0075] Edge-attribute attention bias adjusts the attention score based on edge features, thereby enhancing the focus on abnormal transactions and facilitating the differentiation between legitimate and fraudulent activities. Shared bias improves detection accuracy by leveraging complementary information across multiple financial institutions to promote the learning of complex fraud patterns. Participants under individual attention No. Layer fusion attention bias As shown in formula (6).

[0076] (6)

[0077] in, Indicates the participating parties No. Layer edge attribute attention bias; Indicates the first Participants under individual attention No. The local edge weight matrix of the layer; Indicates the first The first thing to pay attention to. Shared bias of layers.

[0078] By employing an edge mechanism and shared biases, the hierarchical edge-attribute Transformer effectively utilizes local edge features and acquires comprehensive information through federated learning. Furthermore, unlike traditional attention methods that evaluate neighbor importance and apply weights to the entire neighbor communication process, the edge mechanism effectively filters the target neighbor interactions at the dimensional level, thereby extracting key feature dimensions. This combined node-level and dimensional-level weighting enhances the representational power of Fraud-FedGT, contributing to better detection of complex financial transaction patterns.

[0079] To protect model parameters, this embodiment introduces a differential privacy mechanism during the upload process. For the current federated training, the model update amount between the local model parameters and the global model parameters issued by the central server is calculated. The model update amount is then pruned and differential privacy noise is added to obtain the local privacy model parameters, which are then uploaded to the central server. Formally, the model parameters are denoted as... Differential privacy operator is denoted as The cut boundary is denoted as The process is shown in formula (7).

[0080] (7)

[0081] in, Represents the truncation function; This indicates a mean of 0 and a scale of 0. Laplace noise.

[0082] To facilitate understanding, this embodiment uses the shared bias in model parameters as an example. First, we introduce the federated training rounds. , will the The first round of central server maintenance The first thing to pay attention to. The shared bias of the layer is denoted as . No. Each participating party receives the message from the server. And participate in the fusion attention bias calculation according to formula (8) during local training:

[0083] (8)

[0084] After local training is completed, the first Each participant receives the updated local shared bias. Its update size relative to the shared bias issued by the server is:

[0085] (9)

[0086] To reduce the risk of privacy leaks during parameter upload, the shared bias update amount is pruned and differential privacy noise is added:

[0087] (10)

[0088] After receiving the shared bias updates uploaded by each participant, the server performs weighted aggregation based on the local sample size to obtain the next round of shared bias:

[0089] (11)

[0090] in, Indicates the first Round Participants under individual attention No. Layer fusion attention bias, For the first Round Participants under individual attention No. Shared bias of layers, For the first Round Participants under individual attention No. Shared bias update amount of the layer For the first Round The first thing to pay attention to. The noise value added to the shared bias update of the layer. For the clipping function, It is Gaussian noise. Indicates the number of participants. Indicates the first The number of local samples for each participant. This represents the total number of local samples. Indicates the clipping threshold. Represents the identity matrix. Represents the differential privacy noise coefficient. For the first Round The first thing to pay attention to. Shared bias at the layer. Through the above update method, the shared bias can absorb complementary fraud pattern information from various institutions round by round without sharing the original transaction graph data.

[0091] This embodiment aims to predict the anomaly score for each edge (transaction). The final classification component for local participants employs a simple feedforward network. Combined with the Sigmoid activation function The calculation process is shown in formula (12).

[0092] (12)

[0093] in, Indicates the participating parties At the node and Anomaly score in the prediction of transactions between them; Indicates vector concatenation. Participants The nodes output by the Transformer in the last layer of the graph Node characteristics, Participants The nodes output by the Transformer in the last layer of the graph Node characteristics, Participants The nodes output by the Transformer in the last layer of the graph and Edge features between them.

[0094] Furthermore, Fraud-FedGT is optimized by minimizing the binary cross-entropy loss between the predicted value and the true label. (Local loss function) As shown in formula (13).

[0095] (13)

[0096] in, Indicates participants internal nodes The true labels of transactions between them. During training, the local optimizer uses Adam, with a learning rate of 0.001 and a weight decay factor of 1×10⁻⁶. -5 The batch size was set to 1024; in each round of federated training, each participant performed 5 local epochs, and the server-side federated communication rounds were set to 100 rounds. To alleviate the class imbalance problem, a weighted binary cross-entropy loss was introduced for fraudulent sample classes during training; to prevent overfitting, Dropout was added to the attention layer and feedforward network of the HEAT encoder, with a dropout rate set to 0.3, and an early stopping strategy was adopted, stopping training when the validation set F1 score did not improve for 10 consecutive rounds. In addition, norm pruning was performed on the model gradient, with a pruning threshold set to 1.0 to improve training stability.

[0097] Step 3: The central server aggregates the local model parameters of all participants to obtain the global model parameters, and distributes the global model parameters to each participant based on the differential privacy mechanism. The global model parameters serve as the initial model parameters for each participant's local training, and the federated training is iterated repeatedly until the federated training termination condition is met.

[0098] Financial fraud detection not only requires full utilization of edge features but also necessitates the construction of a privacy-preserving cross-institutional detection framework. Therefore, this embodiment employs a differential privacy-based horizontal federated learning framework to achieve cross-institutional fraud detection in a privacy-preserving scenario. Specifically, in the Fraud-FedGT model, the central server coordinates the training process, which is... Each local participant jointly participates in the training. Each participant stores the global graph. a subgraph The subgraphs generated by different participants may partially overlap or be completely independent. The central server does not store any graph data, and privacy regulations restrict each participant from accessing or obtaining information from other participants. The model parameter aggregation process in federated learning is shown in Equation (14).

[0099] (14)

[0100] in, Participants In the The local model parameters of the wheel; It is the first Global model parameters of the wheel; It is the first The data size of each local model parameter; It represents the total data size of all local model parameters.

[0101] After federated training, each participant uses the graph Transformer from the federated training to process the augmented subgraph to be detected, obtaining the node features of each node in the augmented subgraph to be retrieved and the edge features between the nodes; the node features and edge features output by the last layer graph Transformer are input into the classification component to obtain the anomaly scores of all edges. Transactions corresponding to edges with anomaly scores greater than the anomaly threshold are judged to have fraud risk, generating fraud warnings or outputting fraud assessment reports for credit agencies to refer to.

[0102] This invention proposes Fraud-FedGT, a model designed to address cross-institutional fraud detection in financial transaction graphs while maintaining privacy. By combining an improved graph Transformer architecture with a lateral federated learning framework, Fraud-FedGT enables collaborative fraud detection among multiple institutions while preserving data confidentiality. The model fully leverages edge features through three complementary mechanisms: edge-based gating, edge attribute attention bias, and shared bias, effectively capturing complex fraud patterns. Furthermore, directed multigraph augmentation enhances the model's representational capabilities by distinguishing multiple edges between transaction directions and entities. Differential privacy ensures robust data protection during parameter exchange.

[0103] To visually demonstrate the advantages of the method of the present invention, the following experiments were conducted in this embodiment.

[0104] (1) Dataset description:

[0105] This experiment uses three publicly available Anti-Money Laundering (AML) datasets of three different sizes from the Kaggle platform, each containing two variants: High Illegal Ratio (HI) and Low Illegal Ratio (LI). Table 1 shows the number of nodes, edges, and illegal ratio for each dataset.

[0106] Table 1 Key Information of the AML Dataset

[0107]

[0108] (2) Baseline model:

[0109] This experiment compares the proposed Fraud-FedGT model with three baseline models in the field of financial fraud detection. The first category consists of traditional machine learning methods: the XGBoost (eXtreme Gradient Boosting) model and the LightGBM (Light Gradient Boosting Machine) model, both based on gradient boosting decision trees. This experiment also uses a Graph Feature Preprocessor (GFP) to improve the XGBoost and LightGBM models; the resulting GFP-XGBoost and GFP-LightGBM models are also used as baseline models. The second category includes widely used graph neural network models in this field: GIN (Graph Isomorphism Network) and PNA (Principal Neighbourhood Aggregation). Considering the importance of edge features in financial fraud detection tasks, this experiment also uses GIN and PNA (i.e., GIN+EU (GIN with Edge Update, a graph isomorphic network with edge update function) and PNA+EU (PNA with Edge Update, a main neighborhood aggregation network with edge update function)) with edge update function as baseline models. The third category is federated learning models in distributed environments. This study uses three classic federated learning baselines: FedAvg (Federated Averaging), FED-PUB (Federated Personalized sUBgraph learning framework), and PerFedGT (Personalized Federated Graph Transformer model).

[0110] (3) Evaluation and implementation details:

[0111] Given the significant class imbalance inherent in the dataset used, traditional accuracy metrics are insufficient for performance evaluation. Therefore, this experiment uses the F1 score as the primary evaluation metric to align with the evaluation practices commonly adopted by banking and regulatory institutions in practice. Test results are presented based on model parameters that achieve optimal performance on the validation dataset. The experiment utilizes an Intel Core i7-14700HX processor with a base clock frequency of 2.1GHz, a maximum boost frequency of 5.5GHz, a thermal design power (TDP) of 55W, and 33MB of L3 cache.

[0112] (4) Experimental results:

[0113] (4.1) Performance comparison:

[0114] This experiment compares the proposed Fraud-FedGT with three baseline methods on six anti-money laundering datasets, using the F1 score as the evaluation metric. Table 2 summarizes the experimental results on each dataset. All reported performance results are the average of five runs using different random seeds.

[0115] Table 2. Comparison of F1 scores (%) of different methods on six AML datasets.

[0116]

[0117] Based on the experimental results in Table 2, the following conclusions can be drawn:

[0118] Compared to the first type of machine learning methods, the Fraud-FedGT proposed in this invention outperforms the baseline model on all experimental datasets. In particular, Fraud-FedGT exhibits significant and comprehensive advantages over XGBoost and LightGBM, which may be due to the difficulty of traditional machine learning models in effectively handling complex graph structure data. Notably, even when compared with two recent machine learning methods enhanced by graph feature preprocessing (GFP-XGBoost and GFP-LightGBM), Fraud-FedGT maintains an overwhelming performance advantage.

[0119] Compared to the second type of mainstream GNN models, the Fraud-FedGT proposed in this invention also demonstrates superior performance. In particular, compared to the original PNA, PNA+EU with edge update functionality shows significant performance improvements on all experimental datasets, highlighting the crucial role of effectively utilizing edge features in achieving optimal financial fraud detection efficiency.

[0120] Compared to the third type of federated learning model, the Fraud-FedGT proposed in this invention also demonstrates a significant performance advantage. It is worth noting that while PerFedGT achieves a slight lead on the AML-Small-HI dataset, Fraud-FedGT shows a more significant lead on the other five datasets. Overall, compared to FED-PUB and PerFedGT, which employ personalized subgraph federated learning strategies, Fraud-FedGT maintains its competitive advantage due to its superior ability to capture cross-institutional fraud patterns. The performance improvements shown by Fraud-FedGT, FED-PUB, and PerFedGT compared to FedAvg, which does not use graph structures, highlight the importance of effectively utilizing graph structures for improving the performance of fraud detection models.

[0121] (4.2) Ablation studies:

[0122] This section presents an ablation study of the proposed Fraud-FedGT to evaluate the contribution of each component to the overall model performance. The following variants were designed and implemented for analysis, and the experimental data are recorded in Table 3.

[0123] Fraud-FedGT(w / og): This variant removes the edge-based gating mechanism from Fraud-FedGT, thus not utilizing the edge-based gating mechanism to selectively propagate valuable information.

[0124] Fraud-FedGT(w / oa): This variant removes the edge attribute attention bias from Fraud-FedGT and does not utilize the edge attribute attention bias to distinguish between normal and fraudulent transactions.

[0125] Fraud-FedGT(w / os): This variant removes the shared bias from Fraud-FedGT, thus not utilizing the shared bias to allow the local model to explore complementary information between financial transaction graphs of different institutions.

[0126] Fraud-FedGT(w / oo): This variant removes the directed multigraph augmentation method from Fraud-FedGT and instead of using graph-level data augmentation techniques, it directly utilizes the original graph structure data.

[0127] Table 3 Ablation studies of F1 scores (%) for different variants on six AML datasets

[0128]

[0129] Based on the data shown in Table 3, the following conclusions can be drawn:

[0130] The F1 score comparison reveals a significant performance gap between Fraud-FedGT and Fraud-FedGT(w / og). Removing the edge-based gating mechanism leads to a substantial drop in model performance, highlighting its crucial role in effectively identifying key transaction features and propagating valuable information within graph structures.

[0131] Fraud-FedGT (w / oa) achieves a higher F1 score than Fraud-FedGT (w / og), but still falls short of the full Fraud-FedGT model. This indicates that although the role of edge attribute attention bias is not as critical as the edge-based gating mechanism, it still makes a significant contribution to the overall model performance. The complementary effect between the two components is crucial for maximizing edge feature utilization and achieving superior fraud detection performance.

[0132] Compared to the full Fraud-FedGT, Fraud-FedGT(w / os) exhibits a lower F1 score, thus validating the effectiveness of shared biases in a lateral federated learning framework. This finding suggests that shared biases help explore complementary information derived from heterogeneous financial transaction graphs among participating institutions, ultimately improving fraud detection capabilities.

[0133] Although a performance gap remains compared to Fraud-FedGT, Fraud-FedGT(w / oo) achieves results closest to Fraud-FedGT among all variants. This not only demonstrates the effectiveness of directed multigraph augmentation but also indirectly highlights the significant impact of the other modules proposed in this paper.

[0134] (4.3) Component Analysis:

[0135] This section evaluates the effectiveness of the four core components by systematically replacing each core module in Fraud-FedGT. To balance the comprehensiveness of the experiment with computational efficiency, component analysis is performed on three representative datasets: Small-LI and Small-HI are used to examine the component's performance under different levels of data heterogeneity, while Medium-HI is used to verify performance on large-scale datasets.

[0136] (4.3.1) The impact of edge-based gating mechanisms:

[0137] To verify the effectiveness of edge-based gating mechanisms, two comparative variants were designed in this experiment. Fraud-FedGT-RandomGate retains the gating structure but replaces the learnable gating values ​​with fixed random values ​​sampled from a uniform distribution U(0,1) to test the effectiveness of specific gating values. Fraud-FedGT-StaticGate also retains the gating structure but uses fixed gating values ​​derived directly from normalized edge features without the need for learning, to test the effectiveness of rule-based gating mechanisms.

[0138] like Figure 5As shown in (a), Fraud-FedGT-RandomGate exhibits significant performance degradation across all datasets, particularly on Small-LI. This substantial decline indicates that while random filtering provides some regularization, it fails to capture task-relevant patterns and may inadvertently suppress crucial fraud signals in low-heterogeneity scenarios. Fraud-FedGT-StaticGate demonstrates more competitive performance but remains consistently lower than Fraud-FedGT. This persistent gap suggests that fixed rules based on domain knowledge are ill-suited to the diverse and evolving nature of fraudulent activities. Experiments conducted at different scales and levels of heterogeneity demonstrate the superior performance of edge-based gating.

[0139] (4.3.2) The effect of edge attribute attention bias:

[0140] To evaluate the proposed edge attribute attention bias, two contrasting variants were designed in this experiment. Fraud-FedGT-FixedBias uses pre-defined static weights extracted directly from normalized edge features without parameter learning to verify whether domain knowledge is sufficient to support the model. Fraud-FedGT-EdgeConcat, on the other hand, aggregates edge features and node representations before attention computation to verify the effectiveness of the early fusion stage.

[0141] Figure 5 (b) shows the performance on datasets with different levels of heterogeneity. Fraud-FedGT consistently maintains the leading performance, achieving 0.7431, 0.7492, and 0.4518 on Small-HI, Medium-HI, and Small-LI, respectively. The performance comparison reveals a key point: static weights fail to capture context-dependent fraud patterns, particularly evident in the significant gap on Medium-HI (0.7492 vs. 0.6047). While early concatenation is more competitive, it obscures the edge-node interaction structure crucial for effective attention. These results validate that learning-based attention-level edge fusion can be dynamically tuned, essential for robust fraud detection in diverse data scenarios.

[0142] (4.3.3) The effect of shared bias:

[0143] Figure 6The performance evolution of three federated learning methods during training is demonstrated. Among all datasets, Fraud-FedGT-LocalOnly, employing an independent training mode without parameter sharing, exhibits the slowest convergence speed and lowest final performance across all datasets. Its training curve plateaus prematurely, indicating its difficulty in capturing fraud patterns across institutions, and also demonstrating the fundamental limitations of independent learning in collaborative fraud detection scenarios.

[0144] Figure 6 The training curves of three federated learning strategies—LocalOnly (no parameter sharing), Fraud-FedGT (the model proposed in this invention), and FullSharing (full parameter sharing)—were compared on the Small-HI, Medium-HI, and Small-LI datasets. For example, Small-HI-LocalOnly refers to the training curve of Fraud-FedGT-LocalOnly on the Small-HI dataset.

[0145] In contrast, the Fraud-FedGT-FullSharing method, which fully shares all model parameters, exhibits fast convergence and achieves the highest performance. However, by exposing the complete model architecture and parameter distribution, this method introduces significant privacy risks, making it difficult to apply in real-world deployment scenarios where institutional data sensitivity is critical.

[0146] The Fraud-FedGT proposed in this invention achieves a balance between two extreme approaches. Its training curve exhibits a continuously rising trend, with competitive convergence speed, achieving near-FullSharing performance levels by transmitting only bias parameters. This superior efficiency demonstrates that the bias vector can effectively encode complementary fraud information across institutions. The small performance gap validates that the strategic parameter selection can both protect privacy and achieve efficient federated learning, providing an effective solution for cross-institutional fraud detection.

[0147] (4.3.4) The effect of directed multigraph enhancement:

[0148] This experiment validated the effectiveness of directed multigraph augmentation by applying each augmentation technique individually to obtain different variants. Table 4 shows the ablation results of separating the three directed multigraph components. Fraud-FedGT-OnlyPN achieved better results than Fraud-FedGT-OnlyRMP and Fraud-FedGT-OnlyEgoID, demonstrating the strong characterization ability of port numbering. Furthermore, the Fraud-FedGT method, which fully integrates the three components, consistently performed best, indicating that integrating all components produces a synergistic effect, which is crucial for capturing the inherent complex multidimensional patterns of fraudulent transaction networks.

[0149] Table 4 Ablation studies with directed multigraph enhancement

[0150]

[0151] (4.4) Friedman test:

[0152] This experiment systematically evaluated the statistical significance of performance differences among the comparative methods using nonparametric statistical tests. The specific procedure is as follows: by analyzing relative rankings, the Friedman test was applied to comprehensively evaluate the performance of the 12 methods on six AML datasets. The null hypothesis (H0) states that there are no significant differences in performance among all methods across different datasets, while the alternative hypothesis (H1) states that at least one method exhibits significantly different performance from the others. Following the existing statistical testing framework, this experiment first calculated the individual ranking of each method on each dataset, and then derived its average ranking to comprehensively evaluate the relative performance of all methods. Table 5 presents the results of the Friedman test.

[0153] Table 5. Friedman test results between the baseline model and Fraud-FedGT

[0154]

[0155] The Friedman test results showed that there were statistically significant differences among the 12 methods across six AML datasets. With 11 degrees of freedom, p < 0.01, and Kendall correlation coefficient W > 0.7, this method overturns the null hypothesis that all methods have equal performance. In terms of average ranking, the proposed Fraud-FedGT ranks first with an average ranking of 1.17, significantly outperforming the baseline model. This method demonstrates superior generalization ability and robustness on datasets of varying sizes and heterogeneous levels, consistently maintaining its leading position.

[0156] This invention combines federated learning with a graph Transformer architecture, achieving efficient fraud detection while ensuring cross-institutional data privacy through differential privacy parameter exchange. Secondly, by employing an edge-based gating mechanism and edge attribute attention bias, the proposed Fraud-FedGT overcomes the limitations of data silos and effectively captures complex cross-institutional fraud patterns. Directed multigraph augmentation further enhances the detection capability for complex patterns such as circular transactions. This invention adopts an edge-centric design philosophy, prioritizing transaction relationships over account attributes, which aligns with the essential characteristics of financial transaction graphs, thereby extracting effective fraud detection signals to assist the model in identification.

[0157] In the context of increasingly sophisticated fraud methods, the Fraud-FedGT proposed in this invention provides important reference and support for financial institutions in their anti-fraud practices. By enabling cross-institutional collaboration in privacy-preserving scenarios, financial service providers can jointly combat fraud without violating data protection regulations. The horizontal federated learning framework is particularly beneficial to smaller institutions lacking sufficient fraud samples, allowing them to leverage collective intelligence while maintaining data sovereignty. The edge-centric design facilitates seamless integration with existing transaction monitoring systems. Furthermore, the differential privacy mechanism provides an adjustable privacy budget, enabling institutions to strike a balance between detection accuracy and compliance requirements. The superior performance of this method on various AML datasets demonstrates its potential in reducing false positive rates and operating costs, addressing key issues in cross-institutional fraud pattern detection while respecting the characteristics of data silos in a highly competitive financial market.

[0158] In another embodiment, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a privacy-preserving cross-agency fraud detection method based on federated graph learning.

[0159] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory.

[0160] In another embodiment, the present invention also provides a computer device including a processor and a memory storing a plurality of computer instructions, which, when executed by the processor, implement steps of a privacy-preserving cross-agency fraud detection method based on federated graph learning.

[0161] The memory and processor are electrically connected directly or indirectly to enable data transmission or interaction. For example, these components can be electrically connected to each other via one or more communication buses or signal lines. The memory stores a computer program that can run on the processor, and the processor implements the method of the present invention by running the computer program stored in the memory.

[0162] The memory can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), and Electrically Erasable Programmable Read-Only Memory (EEPROM). The memory stores the program, which the processor executes upon receiving execution instructions.

[0163] A processor can be an integrated circuit chip with data processing capabilities. The processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention.

[0164] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

[0165] The above embodiments merely illustrate several implementation methods of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the appended claims.

Claims

1. A privacy-preserving cross-institutional fraud detection method based on federated graph learning, characterized in that, Includes the following steps: Each participant constructs a transaction graph based on the transaction text descriptions of the included accounts, and then enhances the transaction graph to obtain an enhanced subgraph; Each participant uses the enhanced subgraph to iteratively train the graph Transformer until the local training termination condition is met. Based on the differential privacy mechanism, the local model parameters of the graph Transformer are uploaded to the central server. The normalized attention score of the graph Transformer introduces edge attribute attention bias and shared bias, and the multi-head attention operator of the graph Transformer introduces an edge-based gating mechanism for execution. The central server aggregates the local model parameters of all participants to obtain the global model parameters, and then distributes the global model parameters to each participant. The global model parameters serve as the initial model parameters for each participant's local training. The federated training is iterated repeatedly until the federated training termination condition is met. Each participant uses the graph Transformer trained in a federated manner to process the augmented subgraph to be detected, and obtains the node features of each node and the edge features between nodes in the augmented subgraph to be retrieved. The node and edge features output from the last layer of the Transformer are input into the classification component to obtain the anomaly score of all edges. Transactions corresponding to edges with anomaly scores greater than the anomaly threshold are judged to have fraud risk, and a fraud warning is generated.

2. The privacy-preserving cross-institutional fraud detection method based on federated graph learning according to claim 1, characterized in that, The enhanced processing includes reverse message passing, port numbering, and master node identification.

3. The privacy-preserving cross-institutional fraud detection method based on federated graph learning according to claim 1, characterized in that, The edge attribute attention bias is the product of the transpose of the edge feature matrix of the transaction graph input by the current participant at the current layer and the local edge weight matrix, where the transaction graph input by the current participant at the first layer is the enhanced subgraph.

4. The privacy-preserving cross-institutional fraud detection method based on federated graph learning according to claim 1, characterized in that, The edge-based gating mechanism includes: The transpose of the edge feature matrix of the transaction graph input by the current participant at the current layer is multiplied by the local gate control weight matrix, and then processed by the sigmoid function to output the edge gate control vector.

5. The privacy-preserving cross-institutional fraud detection method based on federated graph learning according to claim 1, characterized in that, The normalized attention score of the graph Transformer introduces edge attribute attention bias and shared bias, including: Summing the edge attribute attention bias and the shared bias yields the fused attention bias; The unnormalized attention score is calculated by superimposing the fusion attention bias on the dot product of the query vector and the key vector. The unnormalized attention score is processed using the softmax function to obtain the normalized attention score.

6. The privacy-preserving cross-institutional fraud detection method based on federated graph learning according to claim 1, characterized in that, The multi-head attention operator of the graph Transformer is implemented using an edge-based gating mechanism, including: The normalized attention score, the node features of neighboring nodes, the value weight matrix, and the edge gating vector in the edge-based gating mechanism are multiplied by a dot product, and the result of the dot product is used as the node output feature of a single attention head. Aggregate the output features of each individual attention head corresponding to all neighboring nodes, and obtain the updated node features of the central node through the node output projection matrix; The unnormalized attention scores between nodes are used as the edge output features of a single attention head. The edge output features of all attention heads are concatenated, and the updated edge features between nodes are obtained through the edge output projection matrix.

7. The privacy-preserving cross-institutional fraud detection method based on federated graph learning according to claim 1, characterized in that, The process of uploading local model parameters of the graph Transformer to the central server based on differential privacy mechanism includes: For the current federated training, calculate the model update amount between the local model parameters and the global model parameters issued by the central server; The model update amount is pruned and differential privacy noise is added to obtain local privacy model parameters, which are then uploaded to the central server.

8. The privacy-preserving cross-institutional fraud detection method based on federated graph learning according to claim 1, characterized in that, The node and edge features output from the last layer of the Transformer are input into the classification component to obtain anomaly scores for all edges, including: The classification component includes a feedforward network and a sigmoid activation function. For any edge, the edge features and the corresponding two node features are concatenated, and the anomaly score of the edge is obtained through the feedforward network and the sigmoid activation function.

Images