Lightweight Financial Fraud Detection Using a Symmetrical GAN-CNN Fusion Architecture

Abstract

With the rapid development of information technology and the deep integration of the Internet platform, the scale and form of financial transactions continue to grow and expand, significantly improving users’ payment experience and life efficiency. However, financial transactions bring us convenience but also expose many security risks, such as money laundering activities, forged checks, and other financial fraud that occurs frequently, seriously threatening the stability and security of the financial system. Due to the imbalance between the proportion of normal and abnormal transactions in the data, most of the existing deep learning-based methods still have obvious deficiencies in learning small numbers sample classes, context modeling, and computational complexity control. To address these deficiencies, this paper proposes a symmetrical structure-based GAN-CNN model for lightweight financial fraud detection. The symmetrical structure can improve the feature extraction and fusion ability and enhance the model’s recognition effect for complex fraud patterns. Synthetic fraud samples are generated based on a GAN to alleviate category imbalance. Multi-scale convolution and attention mechanisms are designed to extract local and global transaction features, and adaptive aggregation and context encoding modules are introduced to improve computational efficiency. We conducted numerous replicate experiments on two public datasets, YelpChi and Amazon. The results showed that on the Amazon dataset with a 50% training ratio, compared with the CNN-GAN model, the accuracy of our model was improved by 1.64%, and the number of parameters was reduced by approximately 88.4%. Compared with the hybrid CNN-LSTM–attention model under the same setting, the accuracy was improved by 0.70%, and the number of parameters was reduced by approximately 87.6%. The symmetry-based lightweight architecture proposed in this work is novel in terms of structural design, and the experimental results show that it is both efficient and accurate in detecting imbalanced transactions.

1. Introduction

With the rapid development of information technology and the widespread use of Internet platforms, the number and types of financial transactions have increased dramatically, providing great convenience for users’ lives. Financial transactions provide users with a good sense of transaction experience; however, at the same time, they also provide a breeding ground for financial fraud, such as by malicious actors through technical loopholes or illegal means, money laundering, counterfeiting checks, and other financial fraud, posing a serious threat to the security of financial institutions and the stability of the global financial system. According to data from the International Monetary Fund (IMF), globally, illegal money laundering activities account for about 2% to 5% of gross national product each year, and this proportion is on the rise year by year [1]. According to the Global Financial Crime Report 2024 published by Nasdaq, financial fraud brought nearly 500 billion dollars of losses to the world in 2023, and the frequent occurrence of financial fraud cases has become a major challenge for the financial industry. Therefore, strengthening financial fraud detection has become a prominent research focus.

One of the distinguishing features of financial fraud is the problem of category imbalance. Since the number of normal financial transactions is much larger than the number of abnormal fraudulent transactions, this poses a greater challenge for financial fraud detection. Financial fraud detection methods are categorized into two types: (1) traditional financial fraud detection methods, and (2) deep learning financial fraud detection methods. Traditional financial fraud detection methods mainly include random forest (RF) [2], decision tree (DT) [3], and logistic regression (LR) [4] methods. For example, Lin [2] analyzed a dataset of 150 bank transactions using a random forest technique to evaluate its effectiveness in identifying financial frauds, and the experimental results showed the high efficiency of this method in detecting fraudulent transactions. Devi and Kavitha [3] proposed a decision tree-based method for classifying credit card transactions as normal or suspicious and used multiple metrics for their evaluation, with high accuracy. Liu [4] demonstrated the use of logistic regression in telecom bank card fraud detection, with good results. Ali et al. [5] emphasized the efficiency of support vector machines (SVMs) in dealing with category-imbalanced datasets. Wu et al. [6] concluded that the simple Bayesian algorithm (NB) with its simple structure and high computational efficiency is suitable for the task of credit assessment. The NB algorithm is a simple and computationally efficient algorithm for credit assessment and it can be applied to the credit assessment task.

Despite the effectiveness of the above methods, there is a high risk of misclassification when dealing with complex patterns, and such models have limitations in dealing with complex data. For example, the selection of features needs to be based on specialized domain knowledge. In order to overcome the shortcomings of the above methods, scholars have proposed deep learning-based financial fraud detection methods. Thilagavathi et al. [7] proposed combining graph neural networks (GNNs) with anomaly detection to identify complex financial fraud such as account supervision, identity theft, etc. Li et al. [8] used graph convolutional networks (GCNs) for fraud detection in cryptocurrency trading networks. The evolutionary properties of the network were considered, combined with local structure and equilibrium theory to guide training. Experimental results show that the model is well validated. Nalayini et al. [9] proposed a CNN model based on a smart matrix algorithm for credit card fraud detection using large real-time datasets. Wu et al. [10] applied generative adversarial networks (GANs) in combination with deep learning for financial fraud detection to improve precision and recall through data augmentation and enhance financial transaction security. Fletcher et al. [11] proposed a financial fraud detection model based on a self-attention mechanism, which improves recognition accuracy by capturing remote dependencies and global information. Zhao et al. [12] proposed a credit card fraud detection method based on self-attention generative adversarial networks (SAGANs), which distinguishes the key features through a self-attention mechanism that generates data reflecting real fraudulent behavior, which significantly improves the detection accuracy and recall rate.

Although existing deep learning models in financial fraud detection have made some progress, there are still some problems. Unbalanced data categories can be a problem. Most deep learning relies on a large number of training samples, and when the training samples are insufficient, the training process is biased towards normal transaction data and ignores a few fraudulent behaviors, resulting in low recall. Furthermore, capture of contextual feature information may be inadequate. Existing methods have difficulty adequately capturing key information such as time series and money flows in dynamic transaction networks, which affects detection accuracy. Low computational efficiency can occur with a large number of parameters. Most of the existing deep learning models suffer from large numbers of parameters, high computational complexity, and high feature dimensionality.

To solve the above problems, this study proposes a new symmetrical structure-based GAN-CNN model for detecting financial fraud. In this model, based on a lightweight symmetrical design, fraud samples are generated to alleviate the category imbalance problem, and temporal features and graph structure features are extracted to enhance the model’s ability to model contextual information.

The contributions of this study are reflected in the following aspects. To solve the category imbalance problem, GAN is used to generate data samples to balance the ratio of normal and fraudulent transactions in the dataset, thus enhancing the model’s ability to learn from a few fraudulent samples. To enhance the ability to capture contextual feature information, CNN is introduced to comprehensively consider the temporal features and graph structure features in the transaction data, to enhance the model’s ability to express the transaction relationship and to accurately recognize complex fraudulent behaviors. To reduce the complexity of model computation, the design of the lightweight symmetrical multi-scale attention mechanism and adaptive aggregation module effectively reduces the number of parameters and feature dimensions of the model, further improving the computational performance.

2. Related Work

2.1. Traditional Financial Transaction Fraud Detection Methods

Traditional financial transaction detection methods include methods such as plain Bayes, logistic regression, random forest, and decision tree. For example, Gupta et al. [13] investigated how to utilize the plain Bayes algorithm for credit card fraud detection, especially in the face of highly unbalanced data, and found that plain Bayes can effectively identify credit card fraud. Du et al. [14] proposed a transaction fraud detection method based on random forests, and the experimental results showed significant advantages. In order to cope with the data imbalance problem, Wan and Zhao [15] proposed an improved logistic regression model with a synthetic minority over-sampling technique (SMOTE) and cross-validation to optimize the model parameters, which significantly improved the model performance, especially in terms of accuracy and AUC value. Similarly, Aburbeian and Ashqar [16] proposed an enhanced version of a random forest classifier with SMOTE technique to address the data imbalance problem, and after optimizing the hyper-parameters, the method demonstrated stronger credit card fraud detection capabilities, especially in complex data environments.

In terms of algorithm comparison, Afriyie et al. [17] compared three algorithms, logistic regression, random forest, and decision tree, and the results showed that random forest performed the best in terms of accuracy and AUC value, making it the best detection algorithm by their recommendation. Li and Geng [18] used the BLS algorithm to deal with imbalanced credit data, and compared four algorithms’ performance, namely logistic regression, simple Bayes, support vector machine, and random forest; The results show that the Random Forest classifier is most suitable for classifying the credit dataset, and in the future, Random Forest can be used as a base classifier to develop an integrated model to further improve the prediction accuracy and performance. The assessment of the four algorithms showed that random forest performed best in terms of accuracy, AUC value, and other evaluation indexes. In addition, Zhu et al. [19] pointed out that logistic regression and plain Bayes are suitable for dealing with structurally simple classification tasks, support vector machines are adapted to high-dimensional small-sample problems, and random forests perform particularly well in terms of classification accuracy and model robustness. Kumar et al. [20] proposed a machine learning algorithm based on support vector machines (SVMs) for improving the accuracy of credit card fraud detection, effectively making up for the shortcomings of traditional methods in recognizing complex transaction fraud. Obeng et al. [21] analyzed the effect of multiple machine learning algorithms for fraud detection in financial transactions and found that the eXtreme Gradient Boosting (XGBoost) and random forest models performed well.

2.2. Deep Learning Based Fraud Methods for Financial Transactions

Deep learning methods play an important role in financial fraud detection. GNN and GCN enhance relationship modeling through graph structure and node features, GAN alleviates the data imbalance problem, and CNN extracts spatiotemporal features to improve detection accuracy; several studies have verified their effectiveness. For example, the ASA-GCN model proposed by Ahmad [22] enhances fraud detection performance through adaptive sampling and aggregation mechanisms. The GNN-CL model proposed by Cheng et al. [23] combines GNNs, CNNs, and LSTMs, which enhances the ability to handle complex transaction patterns. The FinFD-GCN model proposed by Yadegar and Rahmani [24] optimizes fraud detection in credit card transactions through graph convolutional networks. The spatiotemporal graph neural network proposed by Khosrav et al. [25] effectively improves the detection efficiency. The two-channel GAN model by Wei and Lee [26] significantly improves the accuracy and robustness of fraud detection in financial transactions.

The FraudGNN-RL model proposed by Zhang et al. [27] combines graph neural networks with reinforcement learning, which enhances the model’s ability to cope with dynamically changing fraud patterns. The credit card fraud detection model based on dilated convolutional neural networks (DCNNs) with a sampling technique designed by Karthika and Senthilselvi [28] improves the detection accuracy. The hybrid SMOTE-GAN technique proposed by Cheah et al. [29] effectively solves the data imbalance problem and improves the model performance. Strelcenia and Prakoonwit [30] investigated the application of generative adversarial networks (GANs) in credit card fraud detection. Li et al. [31] proposed the TA-Struc2Vec algorithm to detect credit card fraud by combining the topological features and transaction amount features of financial transaction networks to improve the efficiency of Internet financial fraud detection. Duan et al. [32] proposed a CaT-GNN model combining causal inference and time graph neural network to improve the robustness of credit card fraud detection.

2.3. Attention Mechanism Based Financial Transaction Fraud Detection Method

Financial fraud detection based on attention mechanisms can extract key features, recognize complex patterns, learn long-distance dependencies, apply dynamic fraud behaviors, and perform superiorly in different financial scenarios. For example, Zhao et al. [12] constructed a self-attentive generative adversarial network (GAN) combined with a graph attention technique to recognize fraudulent transactions, enhancing the model’s adaptive ability and data security. Wei and Lee [26] used a dual-channel graph attention network (GAT) combined with GBDT-DGAN to improve the accuracy of fraud detection and introduced blockchain technology to safeguard data privacy. Xie et al. [33] proposed a spatiotemporal gating network (STGN) to optimize credit card fraud detection through temporal attention and location-aware mechanisms, highlighting the spatiotemporal characteristics of transaction behaviors. Yu et al. [34] explored the application of the transformer model in credit card fraud detection and compared it with a variety of traditional models, demonstrating the transformer’s ability to capture complex transaction patterns. Ghosh et al. [35] combined a multilayer perceptron (MLP), feed-forward neural network (FNN), and attention LSTM in the Bitcoin network, which effectively enhanced fraud detection ability for Bitcoin transactions. Li et al. [36] designed a method based on sliding windows and attention LSTM for Internet financial credit assessment, which improved the accuracy of risk prediction.

3. Method

3.1. Overview

This paper proposes a lightweight financial fraud detection model based on GAN-CNN, which adopts a symmetric branch architecture. As shown in Figure 1, the model includes key stages such as data augmentation, feature modeling, feature fusion, context modeling, and classification decision-making, forming a broader detection process. The model first uses a generative adversarial network (GAN) to synthesize diverse fraud samples, alleviating class imbalance and expanding the distribution of minority class samples to achieve data balancing, thereby providing stable training conditions for subsequent feature learning. A multi-scale attention mechanism simultaneously models local and global transaction features at different perceptual scales, with channel attention and spatial attention working together to enhance the representation accuracy of key patterns. The adaptive aggregation module screens high-correlation transaction relationships based on similarity calculations and random discard strategies, suppressing noise interference and enhancing the ability to identify key fraud patterns. The feature optimization fusion network (FRFN) combines deep separable convolutions and residual connections to integrate multi-level, multi-scale features, improving the sensitivity of complex fraud pattern recognition and model robustness. The context encoding module integrates time-frequency features and account out-degree features to characterize account behavior patterns in dynamic transaction networks, supplementing temporal and structural information to strengthen discriminative evidence. Boosting and bagging ensemble learning fuses outputs from various stages to achieve efficient and precise fraud detection for large-scale high-frequency transaction data.

3.2. Sample Generation

As shown in Figure 2, the GAN generates fraudulent transaction samples by learning real transaction data, balancing the data distribution, and preventing the model from over-biasing towards normal transactions [37]. The generated fraud data is mixed with real data and key nodes and edges are selected by a subgraph generator to enhance the diversity of fraud patterns, while the hidden graph structure information is used to optimize the model [38]. Different connection methods (e.g., dynamic connection) are used to capture complex fraudulent relationships in the transaction network and improve the generalization ability of the detection system. During the training process, the loss function is optimized to ensure the similarity in distribution between the generated data and the real fraudulent transaction data. The loss function is shown below:

$${\mathrm{LOSS}}_{\mathrm{GAN}}=\sqrt{{\displaystyle \sum }_{\mathrm{m}=1}^{\mathrm{M}}{\displaystyle \sum }_{\mathrm{n}=1}^{\mathrm{M}}{\Vert {\mathrm{x}}_{\mathrm{m}}-{\mathrm{x}}_{\mathrm{n}}\Vert }_2^2}$$

(1)

where ${\mathrm{x}}_{\mathrm{m}}$ and ${\mathrm{x}}_{\mathrm{n}}$ denote the feature vectors of the two transaction samples respectively, this loss function measures the Euclidean distance between the samples, ensuring that the generated data points maintain the structural properties of the fraudulent transactions in the feature space, thus improving the detection model’s ability to recognize fraudulent patterns. The method effectively mitigates the category imbalance problem and improves the accuracy and robustness of financial fraudulent transaction detection by using GAN to generate synthetic fraudulent transactions, optimizing the graph structure information, and dynamically constructing the transaction network.

3.3. Adaptive Aggregation Module

As shown in Figure 3, the adaptive aggregation module extracts key features in the transaction network through similarity computation, random discard, and weighted aggregation to improve the accuracy of financial fraud detection. Specifically, node similarity $\mathrm{m}\left(\mathrm{u},\mathrm{v}\right)$ is first computed to measure account transaction associations, and nodes and their neighboring features are processed by feature mapping $\mathrm{M}\left(\mathrm{v}\right)$. Before entering the aggregation phase, the model employs similarity-driven stochastic discarding to remove neurons with a certain probability α to reduce irrelevant transaction interference and prevent model overfitting. Subsequently, the adaptive aggregator weights and fuses the feature information to improve the recognition of abnormal transaction patterns (e.g., money laundering, ring trading, and high-frequency abnormal trading). Finally, the model is regularized using the Dropout mechanism combined with the SeLU activation function, calculated as follows:

$${\mathrm{h}}_{{\mathsf{\theta }}_{\mathrm{i}}}^0=\mathsf{\sigma }\left(\mathrm{dropout}\left({\mathrm{x}}_{\mathrm{i}},\mathsf{\alpha }\right)·{\mathsf{\omega }}_1\right)$$

(2)

where dropout $\left({\mathrm{x}}_{\mathrm{i}},\mathsf{\alpha }\right)$ randomly deletes neurons with probability $\mathsf{\alpha }$ to reduce the model’s dependence on specific transaction patterns and improve the generalization ability; ${\mathsf{\omega }}_1$ is used to linearly transform the features and optimize the feature expression; $\mathsf{\sigma }\left(·\right)$ is the activation function; and SeLU is used here to enhance the network’s nonlinear expression ability and improve the stability of training. Through this mechanism, the model can efficiently screen key transaction features, reduce noise interference, and improve the stability and accuracy of financial fraud detection.

3.4. Attention Mechanisms

As shown in Figure 4, the model introduces a lightweight multi-scale attention mechanism for modeling inter-account transaction associations and enhancing the characterization of financial fraud patterns. The input data is first subjected to layer normalization (LN) to stabilize the feature distribution and enhance the training effect, followed by the generation of query vectors (Q), key vectors (K), and value vectors (V). Q is processed in three parts: ReLU activation to enhance the nonlinear representation, linear transformation to adapt to diverse trading patterns, and attention computation to focus on key trading relationships and highlight abnormal behavioral features. Q and K are input into the lightweight attention module to generate attention weights to enhance the trading features that are highly correlated with the target account. Q and K are each subjected to a $3\times 3$ deep separable convolution (${\mathrm{DWConv}}_{3\times 3}$) and $1\times 1$ standard convolution (${\mathrm{Conv}}_{1\times 1}$), as shown in

Table 1. Parameter Comparison between Ordinary Convolution and Depthwise Separable Convolution.

Methods	Kernel Size	Input Channel	Output Channel	Layer	Parameters	Total (M)
Ordinary	3 × 3	1024	1024	${\mathrm{Conv}}_1$	1024 × 1024 × 3 × 3 = 9,437,184	≈28.3 M
				${\mathrm{Conv}}_2$	1024 × 1024 × 3 × 3 = 9,437,184
				${\mathrm{Conv}}_3$	1024 × 1024 × 3 × 3 = 9,437,184
	5 × 5	1024	1024	${\mathrm{Conv}}_1$	1024 × 1024 × 5 × 5 = 26,214,400	≈78.6 M
				${\mathrm{Conv}}_2$	1024 × 1024 × 5 × 5 = 26,214,400
				${\mathrm{Conv}}_3$	1024 × 1024 × 5 × 5 = 26,214,400
Depthwise Separable	3 × 3	1024	1024	${\mathrm{Conv}}_1$	1024 × 1024 × 3 × 3 = 9,437,184	≈9.4 M
				${\mathrm{Conv}}_2$
				${\mathrm{Conv}}_3$
	5 × 5	1024	1024	${\mathrm{Conv}}_1$	1024 × 1024 × 5 × 5 = 26,214,400	≈26.2 M
				${\mathrm{Conv}}_2$
				${\mathrm{Conv}}_3$

. Under the condition that the number of inputs, outputs, and channels is 1024, the depthwise separable convolution significantly reduces the number of parameters compared with the ordinary convolution. Specifically, for example, the 3 × 3 and 5 × 5 convolution kernels are reduced by 18.9 M and 52.4 M, respectively, reflecting higher computational efficiency and lightweight advantages. Subsequently, after combining the attention information, the dimensionality is adjusted by linear transformation (Linear) to fit the classification task. While maintaining the model’s high efficiency, the mechanism effectively integrates local and global transaction features, improving the detection capability of complex fraudulent behaviors such as abnormal fund flows, ring trading, and high-frequency trading, and making the identification results more accurate and stable.

3.5. Feature-Response Fusion (FRFN)

As shown in Figure 5, in financial fraud transaction behavior detection, the FRFN module is used to integrate different transaction features to improve detection accuracy. Specifically, first, the input matrix ${\mathrm{x}}_{\mathrm{H}\times \mathrm{W}\times \mathrm{C}}$ is layer normalized (LN) to stabilize the data distribution and prevent the gradient problem. Then, parallel convolution (Parallel ${\mathrm{Conv}}_{3\times 3}$) extracts trading patterns at different scales to enhance the model’s perception of long and short-term trading behaviors. The data is processed linearly and reshaped to match the input structure of the depthwise separable convolutional layer (${\mathrm{DWConv}}_{3\times 3}$).${\mathrm{DWConv}}_{3\times 3}$ extracts localized features to reduce computational complexity and maintain spatial features. Then, the data is flattened (Flatten), and the original features are fused with the processed features using residual concatenation, followed by batch normalization (BN) to stabilize the feature distribution. Further, another layer of ${\mathrm{DWConv}}_{3\times 3}$ enhances feature extraction to identify hidden fraud. Finally, the data is nonlinearly mapped by the SeLU activation function to improve the model’s robustness. The whole FRFN structure accurately detects abnormal transaction patterns through the steps of convolution, normalization, and nonlinear transformation to improve financial fraud detection performance.

3.6. Text Encoding

To capture the fraudster’s contextual behavior more accurately, we introduce a context encoding module, which combines time-frequency features without-degree features in order to generate context encodings that effectively characterize each node. As shown in Figure 6. The encoding size is controlled by hyperparameters and fused with hidden node embeddings in subsequent stages to finally output an optimized $Z_V$ representation of the node.

3.6.1. Time-Frequency Coding

To describe the trading behavior of a particular entity in a time series, its set of transactions is defined as follows:

$${\mathrm{E}}_{\mathrm{V}}^{\mathrm{out}}=\left\{\left(\mathrm{V},{\mathrm{u}}_1\right.\right.,{\mathrm{t}}_1),\left(\mathrm{V},{\mathrm{u}}_2,{\mathrm{t}}_2\right),\cdots (\mathrm{V},{\mathrm{u}}_{\mathrm{T}},{\mathrm{t}}_{\mathrm{T}}\left.)\right\}$$

(3)

where $\mathrm{V}$ denotes a fixed entity identifier (e.g., user or account) and ${\mathrm{u}}_{\mathrm{i}}$ represents the type of transaction or operation that occurred at time ${\mathrm{t}}_{\mathrm{i}}$ for that entity. All transactions are arranged in chronological order to construct a time series. To quantify the time intervals of these transactions, their average time intervals are calculated as follows:

$${\mathrm{t}}_{\mathrm{V}}={\displaystyle \frac{\left({\mathrm{t}}_2-{\mathrm{t}}_1\right)+\left({\mathrm{t}}_3-{\mathrm{t}}_2\right)+\cdots +\left({\mathrm{t}}_{\mathrm{T}}-{\mathrm{t}}_{\mathrm{T}-1}\right)}{\mathrm{T}-1}}$$

(4)

This formula calculates the time difference between neighboring transaction events and averages them to represent the time interval characteristics of the transaction. We perform a further feature transformation on the time intervals to obtain the time-frequency encoding (TFE), as follows:

$${\mathrm{T}}_{\mathrm{E}}\left({\mathrm{t}}_{{\mathrm{v}}_{\mathrm{i}}}\right)=\mathrm{SeLU}\left({\mathrm{W}}^{\mathrm{TE}}{\mathrm{t}}_{\mathrm{v}}\right)$$

(5)

where ${\mathrm{W}}^{\mathrm{TE}}$ is the weight matrix and $\mathrm{SeLU}$ is used as a nonlinear activation function to transform the temporal features to obtain a feature representation applicable to the model inputs.

3.6.2. Out-Degree Coding

To extract the out-degree information, the out-degree encoding is computed for the out-degree feature associated with each index ${\mathsf{\theta }}_{\mathrm{i}}$, as follows:

$${\mathrm{d}}_{{\mathrm{e}}_{{\mathsf{\theta }}_{\mathrm{i}}}}=\mathrm{SeLU}\left({\mathrm{W}}^{\mathrm{Ode}}{\mathrm{d}}_{{\mathsf{\theta }}_{\mathrm{i}}}^{\mathrm{out}}\right)$$

(6)

where $\mathrm{SeLU}$ (scaled exponential linear unit) is the nonlinear activation function, and ${\mathrm{W}}^{\mathrm{Ode}}$ is the weight matrix, which is responsible for linearly transforming the input out-degree feature ${\mathrm{d}}_{{\mathsf{\theta }}_{\mathrm{i}}}^{\mathrm{out}}$ and nonlinearly mapping it through $\mathrm{SeLU}$ to optimize the out-degree feature ${\mathrm{d}}_{{\mathrm{e}}_{{\mathsf{\theta }}_{\mathrm{i}}}}$.

3.6.3. Time-Frequency and Out-Degree Coding Fusion

In financial transaction fraud detection, the model extracts transaction behavior features by combining time-frequency features and out-degree coding to identify anomalous transaction patterns. First, the model fuses the time-frequency information ${\mathrm{T}}_{\mathrm{E}}\left({\mathrm{t}}_{{\mathrm{V}}_{\mathrm{i}}}\right)$ and out-degree feature ${\mathrm{Ode}}_{{\mathrm{V}}_{\mathrm{i}}}$ to compute the contextual coding ${\mathrm{C}}_{{\mathrm{E}}_{\mathrm{v}}}$ as follows:

$${\mathrm{C}}_{{\mathrm{E}}_{\mathrm{v}}}=\left(1-\mathsf{\beta }\right){\mathrm{T}}_{\mathrm{E}}\left({\mathrm{t}}_{{\mathrm{v}}_{\mathrm{i}}}\right)+{\mathsf{\beta }\mathrm{Ode}}_{{\mathrm{v}}_{\mathrm{i}}}$$

(7)

Subsequently, the model splices the context-encoded ${\mathrm{C}}_{{\mathrm{E}}_{\mathrm{v}}}$ with the multilevel features $\widehat{{\mathrm{h}}_0}$ and performs a nonlinear mapping via the weight transformation ${\mathrm{W}}_{\mathrm{Z}}$ and the $SeLU$ activation function in order to obtain the optimized transaction features ${\mathrm{Z}}_{\mathrm{V}}$ as follows:

$${\mathrm{Z}}_{\mathrm{V}}=\mathrm{SeLU}\left({\mathrm{W}}_{\mathrm{Z}}\left(\widehat{{\mathrm{h}}_0}\&{\mathrm{C}}_{{\mathrm{E}}_{\mathrm{v}}}\right)\right)$$

(8)

The initial feature representation $\widehat{{\mathrm{h}}_0}$ is calculated by the feature refined fusion network (FRFN), integrating transaction data, historical behavioral characteristics, and account association information at multiple levels, as follows:

$$\widehat{{\mathrm{h}}_0}=\mathrm{FRFN}\left({\mathrm{h}}_{\mathrm{v}}^1\&{\mathrm{h}}_{\mathrm{v}}^2\&\cdots \&{\mathrm{h}}_{\mathrm{v}}^{\mathrm{k}}\&{\mathrm{h}}_{\mathrm{r}}\&{\mathrm{X}}_{\mathrm{v}}\right)$$

(9)

Through this computational process, the model is able to learn the behavioral patterns of trading accounts to effectively detect anomalous transactions, improving the accuracy of fraud detection and the security of the financial system.

4. Experiment

4.1. Datasets

The YelpChi and Amazon datasets are standard datasets commonly used in the field of financial fraud detection and are widely used for the training and evaluation of graph neural network models. As shown in

Table 2. Datasets.

Dataset	YelpChi	Amazon
Nodes	45,954	11,944
edges	3,846,979	4,398,392
(fraud%)	14.5%	10%
Relation	R-U-R, R-S-R, R-T-R	U-P-U, U-S-U, U-V-U

, the YelpChi dataset [39] contained 45,954 nodes and 3,846,979 edges, mainly covering hotel and restaurant review data, in which the proportion of fraudulent reviews was 14.53%. The dataset defines three types of relationships: R-U-R denotes connections between reviews posted by the same user, R-S-R denotes relationships between reviews with the same rating, and R-T-R denotes connections between reviews posted in the same period. The Amazon dataset [1] contained 11,944 nodes and 4,398,392 edges covering reviews of products in the category of music equipment. The percentage of fraudulent reviews was 6.87%. The relationship types include U-P-U for connections between users reviewing the same product, U-S-U for connections between users giving the same ratings in the same period, and U-V-U for relationships between users calculated based on the similarity of review texts.

4.2. Experimental Setup and Metrics

4.2.1. Experimental Setup and Realization

All experiments were conducted under the same training parameters, which are shown in

Table 3. Setting of experimental environment and other parameters.

Project	Content
CPU	Intel Core i7-4700, 2.70 GHz × 12
Memory	32 GB
Operating system	CentOS 7.8 64-bit
Hard disk	1TB
GPU	Nvidia Titan-X × 2
Python	3.7.2
PyTorch	1.4.0
CUDA	10.0
Learning rate	10–3
Momentum	0.73
Weight decay	5 × 10–4
Batch	16
Saturation	1.7
Subdivisions	64

. Referring to existing models [40,41,42,43,44,45,46,47,48,49,50,51,52,53,54], we adopted different training ratio settings for the YelpChi and Amazon datasets. Specifically, the experiments were conducted at 20% and 50% training ratios for the YelpChi dataset, while the training ratio for the Amazon dataset was set at 50%. The experiments were implemented based on the PyTorch framework, and parameters such as optimizer, learning rate, and convolutional layer structure were kept the same. To reduce the effect of randomness, we repeated each set of experiments independently 10 times; the mean and standard deviation are reported. During training, the learning rate was set to 10⁻³, the weight decay coefficient to 5 × 10⁻⁴, and the momentum parameter to 0.73 to control training stability and prevent overfitting. Node embeddings were extracted across multiple layers based on the FRFN module structure, utilizing deep separable convolutions and residual connections to enhance representational capacity. The entire model was trained for 185 epochs, with each epoch using a batch size of 16 for iterative optimization. To improve the model’s generalization performance, we introduced the dropout mechanism in the multi-scale attention mechanism and adaptive aggregation module, with a probability of 0.5. For the time-frequency and out-degree features in the context encoding module, the SeLU activation function was used for nonlinear mapping to enhance the representation of dynamic transaction behavior. Additionally, we set differentiated hyperparameters for tuning on the YelpChi and Amazon datasets to adapt to different graph structures and fraud rates. YelpChi used a saturation parameter of 1.7, while Amazon used a subclassification score of 64, ensuring the model’s stability and effectiveness across multiple scenarios.

4.2.2. Metrics

This paper uses overall accuracy (OA), confusion matrix (CM), number of model parameters, gigaflops per second (GMACs), precision, recall, and F1 as evaluation metrics. Overall accuracy measures the model’s overall classification performance; the confusion matrix visually displays the prediction results for each category and the degree of confusion; GMACs indicates the number of multiply-accumulate operations the model performs per second during inference, reflecting computational complexity; the number of parameters measures the model’s size and structural complexity; precision evaluates the proportion of samples correctly classified as positive by the model; recall indicates the proportion of all true positive samples that are successfully identified; F1 is the harmonic mean of precision and recall, comprehensively reflecting the trade-off between accuracy and coverage. The relevant formulas for these metrics are as follows:

$$\mathrm{OA}={\displaystyle \frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}}$$

(10)

$$\mathrm{Precision}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}}$$

(11)

$$\mathrm{Recall}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}}$$

(12)

$$\mathrm{F}1={\displaystyle \frac{2\mathrm{PrecisionRecall}}{\mathrm{Precision}+\mathrm{Recall}}}={\displaystyle \frac{2\mathrm{TP}}{2\mathrm{TP}+\mathrm{FP}+\mathrm{FN}}}$$

(13)

where true positives (TPs) are the number of fraudulent transactions correctly identified by the model, false positives (FPs) are the number of legitimate transactions incorrectly identified as fraudulent, false negatives (FNs) are the number of fraudulent transactions not identified by the model, and true negatives (TNs) are the number of legitimate transactions correctly identified by the model as non-fraudulent.

4.3. Experimental Results

4.3.1. Experimental Results on YelpChi and Amazon Datasets

• Comparison of classification performance with different training ratios

As shown in

Table 4. Classification accuracy (%) of each model on YelpChi vs. Amazon datasets with different training ratios.

Method	YelpChi		Amazon
	20%	50%	50%
XGBoost [40]	85.47 ± 0.28	87.92 ± 0.29	88.79 ± 0.41
CatBoost [40]	92.11 ± 0.21	95.09 ± 0.22	93.49 ± 0.29
GCN [41]	87.42 ± 0.27	91.04 ± 0.21	94.81 ± 0.22
SGC [42]	85.95 ± 0.22	90.11 ± 0.25	96.72 ± 0.12
GAT [41]	91.28 ± 0.15	94.77 ± 0.12	93.52 ± 0.34
AR-GCN [43]	91.82 ± 0.15	95.02 ± 0.19	96.42 ± 0.39
FAGCN [44]	91.44 ± 0.14	94.82 ± 0.17	92.87 ± 0.28
FdGars [45]	92.15 ± 0.19	94.92 ± 0.14	96.44 ± 0.52
GEM [46]	91.79 ± 0.19	92.95 ± 0.24	93.96 ± 0.34
GraphSAGE [47]	88.94 ± 0.21	91.45 ± 0.27	91.46 ± 0.27
GraphSAINT [48]	92.48 ± 0.15	95.77 ± 0.21	96.77 ± 0.31
GraphConsis [49]	91.22 ± 0.22	94.84 ± 0.09	94.84 ± 0.09
CARE-GNN [50]	91.47 ± 0.28	94.52 ± 0.15	94.63 ± 0.16
HHLN-GNN [1]	94.82 ± 0.14	95.97 ± 0.21	94.92 ± 0.18
GCN+SMOTE [51]	93.67 ± 0.15	96.86 ± 0.12	94.79 ± 0.15
CNN+SMOTE [52]	93.58 ± 0.17	96.78 ± 0.13	94.68 ± 0.22
CNN-GAN [53]	94.52 ± 0.22	94.52±0.22	94.73±0.16
Hybrid CNN-LSTM–attention [54]	94.75 ± 0.13	96.89 ± 0.15	95.67 ± 0.23
Proposed	94.97 ± 0.22	97.15 ± 0.25	96.37 ± 0.23

, the proposed Symmetrical GAN-CNN model achieved an overall accuracy of 94.97% and 97.15% on the YelpChi dataset and 96.37% on the Amazon dataset under a 20% and 50% training ratio, respectively. Compared with GraphConsis, CARE-GNN, and HHLN-GNN models, the accuracy at the 20% training ratio improved by 3.75%, 3.67%, and 0.15% on the YelpChi dataset, respectively; at the 50% training ratio, it improved by 2.31%, 2.63%, and 1.18% on YelpChi, and on Amazon, it improved by 1.53%, 1.74%, and 1.45%. In addition, the accuracy of the model on the Amazon dataset was also superior to that of GCN+SMOTE, CNN+SMOTE, CNN-GAN, and hybrid CNN-LSTM–attention models, with improvements of 1.58%, 1.69%, 1.64%, and 0.70%, respectively. This result shows that symmetrical GAN-CNN maintains excellent performance under both data-scarce and sufficient conditions and remains stable, especially when the training samples are small. This is mainly due to the introduction of the GAN sample generation module, which effectively mitigates the category imbalance, avoids model bias towards majority class transactions, and improves the learning ability for fraudulent transactions (minority class samples).

2.

Improvement in the effect of the attention mechanism on accuracy

As shown in Table 4, the accuracy of the benchmark attention model GAT on the YelpChi dataset at a 20% training ratio was 91.28%, and on the Amazon dataset it was 93.52% at a a 50% training ratio, which was a significant improvement in accuracy compared with the XGBoost model without the introduction of the attention mechanism, indicating that the attention mechanism helped to enhance the model’s ability to focus on key features. The Symmetrical GAN-CNN model proposed in this paper further improved the detection performance based on the fusion of channel attention and spatial attention mechanisms, multi-scale feature extraction, and a depth-separable convolutional structure. The accuracy on the YelpChi and Amazon datasets reached 94.97% and 96.37%, respectively, improvements of 3.69% and 2.85% over the GAT model, 3.75% and 1.53% over the GraphConsis model, 0.15% and 1.45% over the HHLN-GNN model, and up to 9.50% and 7.58% compared with the XGBoost model. At the 50% training ratio, symmetrical GAN-CNN achieved 97.15% and 96.37% accuracy on the YelpChi and Amazon datasets, outperforming all kinds of comparative models, with improvements of 2.38% and 2.85% compared with the GAT model, 2.31% and 1.53% compared with the GraphConsis model, 1.18% and 1.53% compared with the HHLN-GNN model, In addition, compared with the GCN model and the model method that introduced SMOTE technology, the performance of model in this paper improved by 1.30% and 1.39%, respectively, on the YelpChi dataset; compared with the CNN-GAN and hybrid CNN-LSTM–attention models, it also improved by 0.45% and 0.22%, respectively. These results show that the proposed lightweight multi-scale attention mechanism has significant advantages in improving detection accuracy, in which channel attention strengthens the focus on key features such as fund flow direction and frequency, and spatial attention and DWConv structure synergistically model the local and global transaction relationship so that the model has a stronger discriminative ability in complex fraud identification.

3.

Enhanced Recognition of Anomalous Structures

As shown in Table 4, the overall accuracy of the symmetrical GAN-CNN improved on both the YelpChi and Amazon datasets compared with GraphSAINT, FdGars, HHLN-GNN, GCN+SMOTE, CNN+SMOTE, CNN-GAN, and hybrid CNN-LSTM–attention models, with the maximum improvement of 1.45% on the Amazon dataset. The model extracts transaction features at different scales by introducing depth-separable convolutions, where the convolution is used to capture local anomalous behaviors, and the convolution focuses on modeling the relationship between long-distance money flows, thus enhancing the ability to express complex transaction structures. Meanwhile, the context coding module, which combines time-frequency and out-degree information, effectively improves the model’s ability to model dynamic trading behavior. In the feature fusion stage, the model introduces the batch normalization (BN) operation to standardize the feature distribution, improve the training stability, and reduce the feature dimension and computational cost while improving the accuracy of identifying abnormal fund flows, ring trades, and high-frequency frauds.

4.3.2. Model Compactness, Computational Efficiency, and Minority Class Recognition Capability Analysis

This paper selected GCN, SGC, GAT, FAGCN, FdGars, GraphSAINT, HHILN-GNN, GCN+SMOTE, CNN+SMOTE, CNN-GAN, and hybrid CNN-LSTM–attention as comparison models, and comprehensively evaluated their performance in terms of classification accuracy, parameter count, GMACs, precision, recall, and F1. As shown in

Table 5. Model size evaluation.

Method	Acc (%)	Parameters (M)	GMACs (G)	Precision	Recall	F1
XGBoost [40]	85.47	1.2	0.01	—	—	—
CatBoost [40]	92.11	1.5	0.02	—	—	—
GCN [41]	87.42	2.8	0.2	—	0.5211	0.5533
SGC [42]	85.95	2	0.25	—	—	—
GAT [41]	91.28	2.5	0.5	—	0.5386	0.4649
AR-GCN [43]	91.82	4.2	0.55	—	—	—
FAGCN [44]	91.44	4.5	0.5	—	—	—
FdGars [45]	92.15	4.8	0.55	—	—	—
GEM [46]	91.79	5	0.7	—	—	—
GraphSAGE [47]	88.94	2.2	0.4	—	0.5266	0.5471
GraphSAINT [48]	92.48	2.8	0.45	—	—	—
GraphConsis [49]	91.22	4	0.48	—	0.6208	0.6070
CARE-GNN [50]	91.47	5.1	0.75	—	0.7038	0.6138
HHLN-GNN [1]	94.82	5.5	0.8	0.968	0.723	0.812
GCN+SMOTE [51]	93.67	36.4	1.63	0.945	0.705	0.782
CNN+SMOTE [52]	93.58	49.23	2.79	0.937	0.683	0.776
CNN-GAN [53]	94.52	42.37	2.56	0.952	0.710	0.791
Hybrid CNN-LSTM–attention [54]	94.75	39.65	1.87	0.961	0.711	0.807
Proposed	94.97	4.92	0.62	0.971	0.767	0.823

, the detailed information of different models and the model proposed in this paper is listed under the YelpChi dataset with a training ratio of 20%. From this, we can obtain the following observations:

• Parameter scale and model compactness analysis: While most models achieve a classification accuracy of approximately 90%, this is typically accompanied by a large model parameter scale. For example, HHLN-GNN achieves an accuracy of 94.82% on the YelpChi dataset, with a parameter count of 5.5 M, while the symmetrical GAN-CNN model proposed in this paper maintains a slightly higher accuracy of 94.97% with a parameter count of only 4.92 M. Additionally, the classification accuracy of Symmetrical GAN-CNN is 94.97%, which is higher than that of CNN-GAN (94.52%) and hybrid CNN-LSTM–attention (94.75%). Furthermore, the GCN+SMOTE and CNN+SMOTE models have parameter counts of 36.4 M and 49.23 M, respectively, which are significantly higher than the model proposed in this paper. Overall, symmetrical GAN-CNN demonstrates superior performance in parameter count control, showcasing stronger model compression capabilities and deployment adaptability.
• Comparison of computational overhead under the GMACs metric: Compared with HHLN-GNN, the proposed model reduces computational overhead by 0.18 G, while it is slightly higher than GraphSAINT (+0.17 G). In addition, compared with GCN+SMOTE, CNN+SMOTE, CNN-GAN, and hybrid CNN-LSTM–attention, the symmetrical GAN-CNN reduces GMACs by 62%, 78%, 76%, and 67%, respectively, demonstrating its powerful lightweight advantages and inference efficiency. This performance is particularly critical in financial fraud detection tasks, effectively supporting real-time processing and providing a rapid response to large-scale, high-frequency transaction data, offering greater application value.
• Comprehensive evaluation of minority class recognition capabilities based on Precision, recall, and F1: To more comprehensively evaluate the model’s performance in minority class identification tasks, this paper introduces three additional metrics for supplementary analysis—precision, recall, and F1. On the YelpChi dataset, the symmetrical GAN-CNN model outperformed mainstream comparison models in recall, with improvements of 14.6%, 6.3%, 5.7%, and 5.6% compared with GraphConsis, CARE-GNN, CNN-GAN, and hybrid CNN-LSTM–attention models, respectively, significantly enhancing its ability to identify genuine fraudulent transactions. Additionally, compared with GCN+SMOTE and CNN+SMOTE models that employ oversampling strategies, the symmetrical GAN-CNN model achieved improvements of 6.2% and 8.4% in recall and 4.1% and 4.7% in F1, further validating its detection advantages in imbalanced data scenarios. Overall, symmetrical GAN-CNN outperformed the CNN-GAN and hybrid CNN-LSTM–attention models in F1 metrics, demonstrating stronger classification robustness and minority class sensitivity, making it suitable for practical applications such as identifying high-risk transactions.

4.4. Ablation Experiments

To systematically evaluate the impact of each key module in the proposed model on the classification performance, this study included progressive ablation experiments using the YelpChi dataset with a training ratio of 20% as an example. Starting from the base model, the experiment sequentially introduces the category imbalance processing module, the attention mechanism, and the context encoding module, and progressively analyzes the magnitude of the enhancement of the classification accuracy by each module to quantify its practical effects on sample enhancement, feature extraction and behavior modeling.

• Impact of category imbalance processing on experimental results

As shown in

Table 6. Effect of category imbalance treatment on experimental results on YelpChi dataset.

Module Configuration	Accuracy (%)
Baseline (total removal)	90.31
+Category imbalance processing (GAN sample generation)	94.97

, in the base model, all the enhancement modules were removed and only the most basic convolutional structure with node input features were retained; the classification accuracy was 90.31%. A class imbalance processing module based on a generative adversarial network (GAN) was introduced to generate synthetic fraudulent transaction samples and extend the size of the few class samples in the training set. The module learns the distribution characteristics of real fraud samples through adversarial training, thus improving its ability to discern the minority class. Experimental results show that the introduction of this module improved the accuracy rate from 90.31% to 94.97%, with an improvement of 4.66 percentage points, which verifies its significant role in mitigating the skewness of the category distribution and improving the robustness of the model.

2.

Impact of Attention Mechanism on Experimental Results

Based on the introduction of category imbalance processing, the lightweight multi-scale attention mechanism module was further integrated. This module is modeled through the synergy of channel attention and spatial attention so that the model can dynamically pay attention to the nodes and edges in the transaction graph that have strong correlation with fraudulent behaviors, which enhances the ability to perceive both local and global anomalous patterns in the structure of the transaction graph. The experimental results are shown in

Table 7. Effect of Attention Mechanisms on Experimental Results on the YelpChi Dataset.

Module Configuration	Accuracy (%)
+Category imbalance processing (GAN sample generation)	91.25
+Category Imbalance Processing + Attention Mechanism	94.97

; 91.25% to 94.97% is an increase of 3.72 percentage points, indicates that the mechanism has a strong effect in improving feature discrimination and capturing potential risk paths.

3.

Impact of Context Encoding Module on Experimental Results

Based on the previous two modules, the context encoding module is further introduced to fully explore the trading behavior characteristics of accounts in the time dimension. This module integrates contextual information such as transaction frequency, active period, and out-degree of nodes, constructs temporal embedding, and participates in the generation process of the final node representation, to enhance the model’s ability to model dynamic behavioral patterns. The experimental results are shown in

Table 8. Effect of context encoding module on experimental results on YelpChi dataset.

Module Configuration	Accuracy (%)
+Category Imbalance Handling + Attention Mechanism	92.31
+Category Imbalance Handling + Attention Mechanism + Context Encoding	94.97

; the accuracy rate increases from 92.31% to 94.97%, an increase of 2.66 percentage points, indicating that the introduction of contextual information can provide more comprehensive behavioral descriptions for the model and further enhance its comprehensive discriminative performance.

4.

The Impact of Symmetry on Experimental Results

Experimental results from the YelpChi dataset demonstrate that the symmetrical structures significantly enhance the model’s performance. As shown in

Table 9. The impact of symmetry on experimental results from the YelpChi dataset.

Module Configuration	Accuracy (%)
With symmetry module	94.97
Without symmetry module	92.16
Use weight sharing	94.97
Remove weight sharing	93.78

, incorporating the symmetry module into the GAN-CNN detection framework improved the model accuracy from 92.16% to 94.97%, an increase of 2.81 percentage points, indicating that symmetric structures play a crucial role in enhancing feature extraction efficiency and pattern recognition capabilities. Additionally, combining the symmetry module with a weight sharing strategy further optimizes the network structure, reducing parameter redundancy while maintaining stable model performance. Without the symmetry and weight-sharing modules, the accuracy rates were 92.16% and 93.78%, respectively, both lower than the 94.97% achieved with the complete structure. This validates the significant value of symmetry design in constructing lightweight and efficient financial fraud detection models.

5. Conclusions

In this paper, a lightweight financial fraud detection model based on a symmetrical GAN-CNN structure is proposed to effectively deal with the problems of category imbalance, insufficient contextual feature extraction, and high computational complexity. The model includes a GAN to generate fraud samples to alleviate sample imbalance, a CNN to extract temporal and graph structure features, and an integrated lightweight multi-scale attention mechanism and adaptive aggregation module to improve feature expression capability and computational efficiency. The experiments were conducted on YelpChi and Amazon datasets, and the results showed that the model outperformed mainstream methods such as FdGars, GraphSAINT, and HHLN-GNN in terms of accuracy, parameter size, and GMACs, with good detection performance and application potential.

Although this achieved some success, it may have certain shortcomings in terms of accuracy when it comes to data compression and extreme class imbalance. In the future, we will combine multimodal data and sample generation to solve the above problems and increase the transparency of the research model.