A Cross-Institutional Financial Fraud Collaborative Detection Algorithm Based on FedGAT Federated Graph Attention Network

Abstract

Cross-institutional collaborative fraud detection is essential for combating increasingly sophisticated financial fraud, yet privacy regulations and data silos severely constrain knowledge sharing among institutions. This study aims to develop a privacy-preserving framework that enables effective collaborative fraud detection while protecting raw data, with particular emphasis on exploiting symmetry properties in federated architectures and graph topology analysis. We propose an Adaptive Federated Graph Attention Network (FedGAT), which employs spatio-temporal graph attention mechanisms to capture topological structures and dynamic fraud patterns within institutional transaction networks. The framework introduces a symmetric similarity matrix derived from graph topological features, where the symmetry property ($s_{ij}=s_{ji}$) ensures consistent and unbiased measurement of structural relationships between any pair of institutions. Based on this symmetric similarity metric, an adaptive weighted aggregation mechanism is designed for cross-institutional parameter fusion, enabling balanced knowledge transfer that respects the symmetric collaborative relationship among participating institutions. The symmetric information exchange protocol between local institutions and the central server further guarantees equitable contribution and benefit distribution throughout the federated learning process. The framework is evaluated on the Elliptic Bitcoin transaction dataset and the IEEE-CIS fraud detection dataset, with recall rate and false positive rate as primary performance metrics. Results show that FedGAT achieves a recall of 0.85 and a false-positive rate of 0.038 in single-institution detection, representing approximately 40% and 70% improvements over existing methods, respectively. In collaborative detection across five virtual institutions, the symmetry-aware adaptive aggregation mechanism enables all participants to achieve performance gains exceeding 15% while completely eliminating negative transfer effects observed in simple averaging approaches. This work contributes a novel symmetry-based federated learning framework that balances privacy protection with detection performance, advancing the literature on cross-institutional financial risk management.

1. Introduction

1.1. Background

Financial fraud poses a serious challenge to global financial systems. According to international statistics, economic losses from fraud exceed three trillion dollars annually, imposing profound impacts on financial institutions, consumers, and economic stability [1]. With the proliferation of digital payments and the increasing complexity of financial services, fraud schemes have become increasingly sophisticated. Fraudsters employ multi-account associations, cross-platform transfers, and false transaction networks to orchestrate collaborative fraud across multiple financial institutions. Consequently, traditional single-institution defense mechanisms are gradually becoming ineffective [2,3].

Financial fraud refers to intentional deception or misrepresentation conducted for unlawful financial gain, encompassing a wide range of illegal activities such as credit card fraud, identity theft, money laundering, insurance fraud, and investment scams. These fraudulent activities typically occur through several mechanisms: fraudsters may exploit stolen personal information to conduct unauthorized transactions, create synthetic identities by combining real and fabricated data, or establish complex transaction networks to obscure the origins of illicit funds. The financial systems implicated in fraud span multiple sectors, including commercial banks, online payment platforms, securities exchanges, insurance companies, and emerging cryptocurrency markets. Individuals become vulnerable to financial fraud due to various factors, such as insufficient awareness of security risks, weak authentication practices, excessive sharing of personal information on social media, and the increasing sophistication of social engineering attacks that manipulate victims into disclosing sensitive data or authorizing fraudulent transactions.

Despite substantial investments in fraud detection systems by financial institutions, existing detection methods remain largely constrained to local data analysis. These methods can only identify anomalies in individual transactions or accounts. However, carefully designed fraud schemes often do not manifest obvious anomalies within a single institution. Instead, they evade detection through coordinated activities spanning multiple financial platforms [4]. Data isolation between institutions creates “information silos,” making cross-institutional fraudulent networks difficult to detect [5]. Simultaneously, the implementation of the European Union’s General Data Protection Regulation (GDPR) and various national privacy regulations imposes stringent requirements on protecting personal financial data. These regulations directly prohibit direct sharing of raw data across institutions [6].

There is an urgent need to establish a technical solution for collaborative fraud detection across financial institutions while strictly protecting user privacy. Existing defense frameworks must simultaneously avoid centralized data aggregation, fully exploit hidden fraud characteristics in institutional data, and address data heterogeneity resulting from different business scopes and customer bases across institutions [7]. The convergence of these requirements renders traditional single-institution detection methods inadequate. Novel approaches enabling multi-institutional collaborative learning under privacy protection constraints are imperative [8,9].

Recent studies have explored graph neural networks for fraud detection by modeling transaction relationships as graph structures to capture topological patterns associated with fraudulent behavior. Although these methods demonstrate improved detection capabilities compared to traditional feature-based approaches, they remain confined to individual institutions and cannot leverage knowledge from external sources. Federated learning has emerged as a promising paradigm for enabling collaborative model training without sharing raw data, yet existing federated approaches typically aggregate model parameters through simple averaging strategies that assume relatively homogeneous data distributions across participants. Such assumptions rarely hold in financial fraud detection scenarios, where institutions exhibit significant differences in business scopes, customer bases, and transaction patterns. These differences often cause conventional federated aggregation to produce negative transfer, degrading performance for certain participants rather than improving it. A unified framework that captures both transaction network topology and temporal dynamics while enabling privacy-preserving collaboration across heterogeneous institutions remains absent from the current literature.

Under this context, the key challenge lies in effectively leveraging transaction network information embedded in institutional data while balancing privacy protection and detection performance. Financial fraud is fundamentally a complex network phenomenon. Feature engineering alone is insufficient to capture its evasive and evolving nature. A truly effective cross-institutional collaborative defense system must address three dimensions: privacy, heterogeneity, and detection accuracy. This poses unprecedented challenges to existing technical frameworks.

The remainder of this paper is organized as follows. The Introduction section continues with a review of related work and a summary of the research contributions. The Methodology section formally defines the problem, introduces the spatio-temporal graph attention network, describes the federated learning and parameter aggregation strategy, and presents the overall FedGAT framework design. The Experiments and Analysis section details the experimental setup and datasets, compares the performance of single-institution and multi-institution collaborative fraud detection, and provides a discussion of the results. Finally, the Conclusion section summarizes the findings and outlines directions for future research.

1.2. Related Work

Financial fraud detection has long been a critical concern for financial system security. Early fraud detection relied primarily on rule-based systems and manual audits, which were inefficient and poorly suited to evolving fraud schemes. With advances in data science, machine learning methods have been widely applied to fraud detection. Research on anomaly detection in payment systems and credit card transactions employed traditional classification algorithms, including logistic regression, k-nearest neighbors, support vector machines, and decision trees. These methods distinguish normal from abnormal transactions through learning transaction features [10,11]. Meanwhile, clustering and statistical methods have been applied to the identification of anomalous transactions. For example, approaches combining K-Means clustering with Z-Score statistics detect abnormal transaction patterns [12]. While these traditional machine learning methods perform well in certain scenarios, their limited feature representation capability makes them unsuitable for capturing deep-level patterns in complex transaction data. Subsequently, deep learning methods introduced new possibilities for fraud detection. Researchers explored various deep learning architectures for detecting anomalous transactions. These include autoencoders based on long short-term memory networks, generative adversarial networks, and variational autoencoders. These methods demonstrate strong anomaly identification capabilities in unlabeled data scenarios [13,14]. However, most existing deep learning methods analyze only transaction features themselves, neglecting intrinsic associations and network structure information between transactions. Additionally, in practical applications, financial institutions face multiple relatively independent data islands. Collaborative detection capabilities between these institutions are severely limited, making it difficult to identify fraud networks spanning multiple platforms. Against this backdrop, comprehensive approaches simultaneously addressing network structure modeling, dynamic temporal feature extraction, and multi-institutional privacy protection remain an unsolved problem [15,16].

To overcome limitations of relying purely on deep learning, the academic community has conducted extensive research on feature and model optimization. However, these studies commonly face bottlenecks across multiple dimensions. Su et al. proposed differential fraud risk detection methods by integrating raw financial statements, non-financial indicators, and industry-specific data, distinguishing between general and severe fraud. However, this research’s limitation lies in focusing solely on transaction data features, without modeling the relationships among accounts. Consequently, it cannot identify coordinated fraud involving multiple participants [17]. Bower et al. investigated the severe class imbalance problem in financial fraud data. They compared multiple data balancing techniques, including Resample, Synthetic Minority Oversampling Technique (SMOTE), Adaptive Synthetic Sampling (ADASYN), and RandomOverSampler, assessing their effects on classifiers such as k-nearest neighbors, support vector machines, and random forests. While this research highlighted the importance of data preprocessing for fraud detection, its scope was limited to single-institution tabular data processing. It neglected the network properties and cross-institutional collaborative characteristics of fraud [18]. Alam et al. extracted temporal, spatial, and relational features from transaction records through feature engineering. They applied machine learning algorithms, including decision trees, random forests, and CatBoost, for fraud detection. The features extracted in this study remain at the level of individual transactions or single accounts. They fail to leverage topological structure information of transaction networks, making it difficult to identify fraud networks based on multi-account collaboration [19]. Zeng and Dai constructed interpretable features based on earnings management analysis by integrating domain knowledge with machine learning for enterprise fraud detection. This research achieved better classification accuracy compared to methods using only raw financial statement features. Nevertheless, this study adopted a centralized single-institution detection paradigm without considering privacy regulation constraints on data sharing [20]. Blancaflor et al. explored machine learning applications in credit card fraud detection within the Philippine context, accounting for unique economic, technological, and social factors. This research’s value lies in emphasizing the geographical adaptability of fraud detection. Its primary limitation, however, is that it remains based on isolated data from a single financial institution, unable to achieve cross-institutional knowledge fusion [21]. Al-dahasi et al. conducted comprehensive data preprocessing and model optimization procedures, employing XGBoost and random forests for fraud detection while emphasizing the importance of addressing sample imbalance. However, their detection framework lacks capturing the dynamic evolution characteristics of transaction networks [22]. Li et al. proposed the XGB-GP framework, combining extreme gradient boosting with genetic programming to generate interpretable fraud detection models. Its core remains feature analysis based on financial indicators, remaining at the static feature level of individual enterprises [23]. Thakkar et al. conducted a bibliometric analysis of trends in artificial intelligence and machine learning research on fraud detection. They identified growing research community interest and key research directions, emphasizing the importance of strong legal frameworks, real-time analysis systems, and multi-stakeholder collaboration [24]. In summary, existing research has made significant progress in feature engineering, model optimization, and data preprocessing. However, most methods overlook topological structure features of transaction networks, relying solely on static feature analysis within single institutions. Simultaneously, existing frameworks cannot achieve true multi-institutional collaborative learning under privacy protection constraints. Therefore, developing an integrated framework simultaneously satisfying network structure modeling, dynamic temporal capture, multi-institutional privacy protection, and collaborative learning has become a necessary requirement for current research.

To systematically understand the current state of existing fraud detection methods and their relevance to this research, the following table summarizes major recent studies. The table synthesizes the characteristics of various studies in application scenarios, research content, and potential limitations. It aims to clearly elucidate the strengths and limitations of existing methods, thereby highlighting the innovative contributions of the Adaptive Federated Graph Attention Network (FedGAT) framework proposed in this research in multi-institutional collaboration, heterogeneous graph processing, and spatio-temporal feature fusion.

As summarized in

Table 1. Summary of representative studies on financial fraud detection.

Authors	Application Scenario	Research Content	Potential Limitations
Fang et al. [25]	Loan fraud detection	Graph-enhanced fraud detection (GraphFA), introducing obfuscation detection sampler to identify hidden fraudulent edges, modeling customer information as graph structure	Limited to single-institution environment without involving multi-institutional collaborative learning; lacks temporal dynamic pattern modeling, capturing only static network structure
Cui et al. [26]	Multi-institutional financial transactions	Combining GNN with reinforcement learning (FraudGNN-RL), fusing spatio-temporal semantic features, introducing federated learning for multi-institutional privacy protection	Although incorporating federated learning, lacks adaptive handling of heterogeneous graph structure differences across institutions; may suffer from negative transfer
Menon et al. [27]	Credit card fraud detection	Integrating ensemble learning with deep learning (Fraud Shield), employing voting, stacking, CNN, and LSTM techniques	Based only on the feature space of individual transactions without leveraging relationship networks among accounts; unable to identify coordinated fraud across multiple accounts; lacks privacy protection mechanisms
Karst et al. [28]	Multi-institutional data sharing	Synthetic data sharing platform (FinDEx), achieving data sharing under privacy protection through synthetic data technologies such as generative adversarial networks	Although addressing privacy concerns, generated data may lose original network structure characteristics; does not directly perform fraud detection, requiring combination with other methods; validity across heterogeneous financial institutions unverified
Innan et al. [29]	Financial fraud detection	Quantum graph neural network (QGNN), enhancing QGNN performance using variational quantum circuits	Although introducing quantum computing to enhance computational efficiency, still limited to single-institution scenarios; lacks temporal feature modeling; heterogeneous graph processing capability unexplored
Miao [30]	Financial fraud detection survey	Summarizing deep learning applications in financial fraud detection, emphasizing advantages of deep learning over traditional methods	As a survey, lacks a concrete integrated framework or novel methods; does not address systematic solutions for multi-institutional collaboration or privacy protection; articulates current status without providing solutions
Faccia et al. [31]	Financial statement fraud detection	Applying natural language processing for sentiment analysis, identifying anomalies through text polarity and subjectivity	Analyzes only text features without capturing topological information of transaction networks; serves as a supplementary method rather than an independent detection means; inapplicable to cross-institutional transaction network scenarios
Tang and Liu [32]	Multi-industry financial fraud detection	Distributed knowledge distillation framework (based on Transformer), transferring knowledge from multiple teacher networks to student network	Although involving distributed processing, knowledge transfer targets industry differences rather than institutional heterogeneity; lacks privacy protection mechanisms; difficult to extend to multi-institutional collaboration
Li et al. (Bixuan) [33]	Financial statement fraud identification	Using machine learning to analyze key financial indicators, addressing data imbalance through SMOTE	Limited to financial indicator features, completely overlooking transaction network structure; unable to discover cross-institutional fraud patterns; single-institution and static analysis lacking temporal dynamics
Su et al. [17]	Enterprise fraud risk detection	Integrating financial and non-financial data, utilizing AdaBoost to distinguish different fraud types	Uses only aggregated data features without establishing relationship graphs; adopts centralized paradigm without privacy protection; limited capability in identifying coordinated fraud

, existing fraud detection methods remain limited in several respects. Most approaches focus on static transaction features and neglect the topological structure of transaction networks. Methods incorporating graph neural networks typically operate within single-institution settings and do not support cross-institutional collaboration. Federated learning-based approaches have been proposed to address privacy concerns, yet these methods fail to account for institutional heterogeneity, potentially leading to negative transfer during model aggregation. The proposed FedGAT framework addresses these limitations through an integrated design. The spatio-temporal graph attention mechanism enables simultaneous modeling of transaction network topology and dynamic temporal patterns, moving beyond static feature analysis. The federated learning architecture supports collaborative fraud detection across multiple institutions while preserving the privacy of raw data, thereby overcoming the constraints inherent to single-institution detection paradigms. The adaptive weighted aggregation mechanism, which leverages graph topological similarity, mitigates negative transfer arising from heterogeneous data distributions across institutions. This study contributes to the literature by providing a unified framework that reconciles privacy-preserving collaborative learning with effective fraud detection in heterogeneous multi-institutional environments.

1.3. Research Contributions

• Design of Spatio-Temporal Graph Attention Network: To address the gap that existing methods neglect transaction network topology and temporal dynamics, this research proposes a graph attention network architecture integrating spatial and temporal dimensions. The spatial layer employs multi-head attention mechanisms to learn associations between accounts. The temporal layer models dynamic features in transaction sequences through GRU and introduces temporal decay mechanisms to assign higher weights to recent transactions. This dual-dimensional fusion design provides a more complete fraud-feature representation compared to single-dimensional modeling. The model simultaneously captures both topological and temporal anomalies.
• Adaptive Weight Federated Aggregation Mechanism: To overcome negative transfer caused by institutional heterogeneity in conventional federated averaging, this research innovatively introduces graph topology similarity metrics into the federated learning framework. By computing similarity matrices across institutions, it dynamically adjusts the contribution weights of each institution during parameter aggregation. This mechanism effectively addresses differences among heterogeneous institutions. Heterogeneous institutions benefit from collaborative learning.
• Cross-Institutional Collaborative Framework with Privacy Protection: To bridge the gap between privacy constraints and cross-institutional knowledge sharing, this research achieves multi-institutional fraud collaborative detection while protecting data privacy. Through the federated learning paradigm, institutional data always remains local. Only model parameters are uploaded for aggregation. The framework combines spatio-temporal attention networks and adaptive aggregation. It provides a feasible technical solution for collaborative risk control among financial institutions.

2. Methodology

2.1. Problem Statement

The core of financial fraud detection lies in leveraging network structure information in transaction data to identify anomalous accounts. In cross-institutional scenarios, transaction data from a single financial institution i can be represented as a directed weighted graph:

$$G_i=(V_i,E_i,X_i)$$

(1)

Here, $V_i=\{u_1^{\left(i\right)},u_2^{\left(i\right)},\dots ,u_{n_i}^{\left(i\right)}\}$ denotes the set of account nodes in institution i. $n_i$ is the total number of accounts. $E_i\subseteq V_i\times V_i$ denotes the set of transaction edges. $X_i\in {\mathbb{R}}^{n_i\times d}$ denotes the node feature matrix. d is the feature dimension. The properties of transaction edge $(u_j,u_k)\in E_i$ can be described as

$$e_{jk}^{\left(i\right)}=(u_j,u_k,w_{jk}^{\left(i\right)},t_{jk}^{\left(i\right)},a_{jk}^{\left(i\right)})$$

(2)

Here, $w_{jk}^{\left(i\right)}\in {\mathbb{R}}^{+}$ denotes transaction amount. $t_{jk}^{\left(i\right)}\in {\mathbb{R}}^{+}$ denotes transaction timestamp. $a_{jk}^{\left(i\right)}\in \{0,1\}$ indicates whether the transaction is fraudulent (label, available only in training set). In financial systems, typically M mutually independent financial institutions require collaborative fraud detection. Define the graph collection across institutions as

$$\mathcal{G}=\{G_1,G_2,\dots ,G_M\}$$

(3)

Each $G_i$ may have significantly different scale and structure (node count $n_i$, edge count $|E_i|$, feature dimension $d_i$, etc.). Privacy regulations and competitive constraints require that raw transaction data cannot be directly shared or aggregated across institutions. This constraint is expressed as

$$D_i\cap D_j=Ø,\forall i\ne j,i,j\in \{1,2,\dots ,M\}$$

(4)

Here, $D_i$ denotes the raw transaction dataset of institution i. This means institutions cannot perform centralized data fusion. Due to differences in service scope, customer base, and business characteristics, the topological structures of transaction graphs exhibit significant heterogeneity. Define graph structure similarity between institutions i and j as

$$\sigma (G_i,G_j)=f(\mathrm{GraphFeatures}\left(G_i\right),\mathrm{GraphFeatures}\left(G_j\right))$$

(5)

Here, $\mathrm{GraphFeatures}(·)$ extracts topological features of graphs (such as node count, edge count, average degree, clustering coefficient, etc.). $f(·)$ is the similarity computation function (such as cosine similarity). Given labeled data, the objective of single-institution fraud detection is to minimize classification loss. For institution i, define the empirical loss function of its local detection model ${\theta }_i$ as

$${\mathcal{L}}_i\left({\theta }_i\right)={\displaystyle \frac{1}{|V_i^{\mathrm{train}}|}}\sum _{u\in V_i^{\mathrm{train}}}\ell (f_{{\theta }_i}(u,G_i),a_u^{\left(i\right)})$$

(6)

Here, $V_i^{\mathrm{train}}$ denotes the training set accounts of institution i. $f_{{\theta }_i}(·)$ denotes the detection model with parameter ${\theta }_i$ predicting account u on graph $G_i$. $\ell (·)$ is the binary cross-entropy loss function. $a_u^{\left(i\right)}$ is the true label. Detection based solely on single-institution data faces inherent performance limits. Cross-institutional collaborative learning has the potential to improve overall detection performance by fusing fraud characteristics from multiple institutions. Under privacy protection constraints, the core problem of achieving this goal is described as follows

Problem 1.

Under the privacy constraint in Equation (4), how can collaborative learning be performed using the transaction graph collection $\mathcal{G}$ across institutions such that each institution obtains a global fraud detection model with superior performance to single-institution models while keeping original data local? Formally, define the objective as

$$\underset{{\theta }_1,{\theta }_2,\dots ,{\theta }_M}{min}\sum _{i=1}^M{\mathcal{L}}_i\left({\theta }_i^{*}\right),s.t.D_i\cap D_j=Ø,\forall i\ne j$$

(7)

Here ${\theta }_i^{*}$ denotes the optimal model parameters for institution i under privacy constraints. The ${\theta }_i^{*}$ obtained through collaborative mechanisms should outperform independent training.

2.2. Spatio-Temporal Graph Attention Network

• Limitations of Traditional Methods: Traditional feature engineering-based machine learning methods focus only on static properties of individual transactions. They neglect relationship networks among accounts. Pure graph neural network methods can capture topological structures but lack modeling of transaction temporal dynamics [34]. This single-dimensional processing causes detection models to be unable to simultaneously identify structural and temporal anomalies. Carefully designed multi-step fraud activities easily evade detection [35].
• Advantages of Spatio-Temporal Fusion: The spatio-temporal graph attention network shown in Figure 1 overcomes these limitations through parallel design of spatial and temporal layers. The spatial layer leverages multi-head graph attention mechanisms to learn the association strength between accounts and the fraudulent network topology. The temporal layer captures dynamic features and periodic patterns in transaction sequences through GRU and temporal decay mechanisms. Features from both layers are fused to output fraud predictions. The model identifies complex fraud from both topological and temporal dimensions, significantly improving detection precision and robustness compared to single-dimensional methods.

To effectively identify fraudulent accounts within a single financial institution, this section proposes a Spatio-Temporal Graph Attention Network (ST-GAT). Graph attention networks are adopted instead of graph convolutional networks because financial transaction networks are typically sparse with heterogeneous node degrees. GCN’s fixed aggregation weights treat all neighbors equally and cannot distinguish the varying importance of different transaction relationships. In contrast, GAT’s learnable attention coefficients enable the model to adaptively focus on the most informative neighboring transactions, which is particularly suitable for identifying subtle fraud patterns hidden among legitimate activities. The core idea of this network design is to simultaneously capture topological relationships and dynamic evolution features of transaction networks through parallel design of spatial and temporal layers. For transaction graph $G_i$ of institution i, graph attention mechanisms are first employed to learn association strength among accounts. Attention weights are computed by comparing feature similarity of node pairs. This process targets multiple independent attention heads. For the h-th attention head, the weight of edge $(u_j,u_k)$ is defined as

$${\alpha }_{jk}^{(i,h)}={\displaystyle \frac{exp\left(\mathrm{LeakyReLU}\left({\mathbf{a}}^{\left(h\right)T}\left[{\mathbf{W}}^{\left(h\right)}{\mathbf{x}}_j^{\left(i\right)}\parallel {\mathbf{W}}^{\left(h\right)}{\mathbf{x}}_k^{\left(i\right)}\right]\right)\right)}{{\sum }_{m\in {\mathcal{N}}_j^{\left(i\right)}\cup \left\{j\right\}}exp\left(\mathrm{LeakyReLU}\left({\mathbf{a}}^{\left(h\right)T}\left[{\mathbf{W}}^{\left(h\right)}{\mathbf{x}}_j^{\left(i\right)}\parallel {\mathbf{W}}^{\left(h\right)}{\mathbf{x}}_m^{\left(i\right)}\right]\right)\right)}}$$

(8)

Here ${\mathbf{a}}^{\left(h\right)}\in {\mathbb{R}}^{2d^{\prime }}$ is the parameter vector of the h-th attention head. ${\mathbf{W}}^{\left(h\right)}\in {\mathbb{R}}^{d^{\prime }\times d}$ is the linear projection matrix. $\left|\right|$ denotes vector concatenation operation. ${\mathcal{N}}_j^{\left(i\right)}$ is the neighborhood set of node j. Based on the above attention weights, the spatial layer fuses neighborhood information through multi-head attention aggregation. Outputs from multiple attention heads are concatenated and pass through nonlinear activation functions to produce the spatial feature representation of nodes:

$${\mathbf{h}}_j^{\left(s\right)}=\sigma \left(\left.∥\_{h=1}^H\sum _{k\in {\mathcal{N}}_j^{\left(i\right)}\cup \left\{j\right\}}{\alpha }_{jk}^{(i,h)}{\mathbf{W}}^{\left(h\right)}{\mathbf{x}}_k^{\left(i\right)}\right){\mathbf{W}}^{\left(out\right)}\right)$$

(9)

Here H is the number of attention heads. $\sigma (·)$ is the activation function (such as ReLU). ${\mathbf{W}}^{\left(out\right)}\in {\mathbb{R}}^{H{d}^{\prime }\times d_s}$ is the output projection matrix. Meanwhile, the temporal layer captures dynamic features in transaction sequences. Since fraud patterns exhibit clear temporal properties, temporal information when transactions occur is crucial. Transaction edge weights are adjusted through temporal decay based on timestamps. Recent transactions have a larger influence:

$$w_{jk}^{\left(decay\right)}=w_{jk}^{\left(i\right)}·exp\left(-\lambda ·{\displaystyle \frac{t_{max}^{\left(i\right)}-t_{jk}^{\left(i\right)}}{t_{max}^{\left(i\right)}}}\right)$$

(10)

Here, $\lambda >0$ is the temporal decay coefficient. $t_{max}^{\left(i\right)}$ is the endpoint of the observation time window. $t_{jk}^{\left(i\right)}$ is the transaction occurrence time. For node j’s transaction history, a series of transaction records is organized in chronological order. Feature sequences are constructed. Let node j’s transaction sequence be ${\mathcal{T}}_j=\{(k_1,t_1^{\left(i\right)}),(k_2,t_2^{\left(i\right)}),\dots ,(k_L,t_L^{\left(i\right)})\}$, where $(k_m,t_m^{\left(i\right)})$ denotes the counterparty and time of the m-th transaction. Temporal feature encoding is as follows:

$${\mathbf{C}}_j^{\left(i\right)}=\left[{\mathbf{c}}_1^{\left(i\right)};{\mathbf{c}}_2^{\left(i\right)};\dots ;{\mathbf{c}}_L^{\left(i\right)}\right]\in {\mathbb{R}}^{L\times d_c}$$

(11)

Here, ${\mathbf{c}}_m^{\left(i\right)}=\varphi (w_{j{k}_m}^{\left(decay\right)},{\mathbf{x}}_{k_m}^{\left(i\right)},{\mathbf{x}}_j^{\left(i\right)})$ is the representation fusing transaction amount, counterparty features, and initiating account features. $\varphi (·)$ is the feature fusion function (such as concatenation followed by a linear layer). $d_c$ is the encoding dimension. Transaction sequences are modeled through gated recurrent units (GRU) to extract temporal dependencies. GRU performs recursive computation on sequence ${\mathbf{C}}_j^{\left(i\right)}$. The final output is the temporal representation of nodes:

$${\mathbf{h}}_j^{\left(t\right)}=\mathrm{GRU}\left({\mathbf{C}}_j^{\left(i\right)}\right)={\mathbf{h}}_L^{\left(i\right)}\in {\mathbb{R}}^{d_t}$$

(12)

The GRU recursion is expressed as follows:

$${\mathbf{h}}_m^{\left(i\right)}=(1-{\mathbf{z}}_m^{\left(i\right)})\odot {\mathbf{h}}_{m-1}^{\left(i\right)}+{\mathbf{z}}_m^{\left(i\right)}\odot {\tilde{\mathbf{h}}}_m^{\left(i\right)},m=1,\dots ,L$$

(13)

The update gate, reset gate, and candidate hidden state are defined by standard GRU formulas. ⊙ denotes element-wise multiplication. Features from spatial and temporal directions are fused for fraud prediction. Feature fusion is realized through concatenation and nonlinear transformation. Finally, the sigmoid function outputs fraud probability:

$$p_j^{\left(i\right)}=\sigma \left({\mathbf{w}}_{out}^T\mathrm{ReLU}({\mathbf{W}}_{fusion}[{\mathbf{h}}_j^{\left(s\right)}\parallel {\mathbf{h}}_j^{\left(t\right)}]+{\mathbf{b}}_{fusion})+b_{out}\right)$$

(14)

Here, ${\mathbf{W}}_{fusion}\in {\mathbb{R}}^{(d_s+d_t)\times d_f}$ is the weight matrix of the fusion layer. ${\mathbf{w}}_{out}\in {\mathbb{R}}^{d_f}$ is the output weight vector. $p_j^{\left(i\right)}\in [0,1]$ denotes the probability that node j is a fraudulent account. The spatio-temporal graph attention network achieves complex feature extraction from transaction networks within single institutions through the complete process from Equations (8)–(14). However, this model leverages only local data. It fails to benefit from the fraud characteristics of other institutions. To further improve detection performance and meet the objectives in the problem statement, cross-institutional collaborative learning mechanisms must be introduced.

Theorem 1.

The prediction model ${\mathcal{M}}_i^{ST-GAT}$ of the spatio-temporal graph attention network for institution i effectively captures the multi-hop neighborhood structure in graph $G_i$ and temporal dependencies in transaction sequences. Let the empirical loss be

$${\mathcal{L}}_i^{ST-GAT}\left({\theta }_i^{ST}\right)={\displaystyle \frac{1}{|V_i^{train}|}}\sum _{j\in V_i^{train}}\ell (p_j^{\left(i\right)},a_j^{\left(i\right)})$$

(15)

Through spatial attention mechanisms and temporal GRU modeling, this model can significantly reduce empirical risk on standard datasets compared to methods based solely on node features.

Corollary 1.

The model parameters ${\theta }_i^{ST}$ of spatio-temporal graph attention network can be independently computed locally at each institution under privacy constraints. The parameter set can serve as the initial value for the subsequent federated learning stage. Parameter transferability is characterized by

$${\theta }_i^{ST}={\{{\mathbf{W}}^{\left(h\right)},{\mathbf{a}}^{\left(h\right)},{\mathbf{W}}_{GRU},{\mathbf{W}}_{fusion},\dots \}}_{h=1}^H$$

(16)

This enables each institution to independently optimize local models. It provides a feasible foundation for subsequent cross-institutional parameter aggregation.

2.3. Federated Learning and Parameter Aggregation

• Limitations of Centralized Methods: Traditional centralized fraud detection methods require financial institutions to aggregate raw transaction data to a central server for joint modeling. Although theoretically capable of achieving optimal global models, this approach severely violates privacy regulation requirements. It exposes sensitive customer information of institutions. It increases risks of data breaches and misuse [36]. Simultaneously, data aggregation incurs high costs and substantial security risks. Financial institutions cannot achieve true collaboration. Independent detection lacks cross-institutional synergy [37].
• Advantages of Federated Learning: Federated learning frameworks employ parameter rather than data aggregation. Each institution independently trains models using spatio-temporal graph attention networks locally. Only model parameters are uploaded to the central server for weighted aggregation. Global models are generated and redistributed to institutions. This “data-static, model-dynamic” paradigm protects original data privacy. Through adaptive weight-aggregation mechanisms, it handles differences across heterogeneous institutions. Every institution benefits from global knowledge. True collaborative fraud detection is achieved without sharing sensitive data (as shown in Figure 2).

Financial institutions independently train fraud detection models using spatio-temporal graph attention networks locally. They obtain local parameters ${\theta }_i^{ST}$. However, single-institution model performance is constrained by the scale of local data and feature diversity. To overcome this bottleneck while satisfying privacy protection constraints, this section proposes a federated learning framework for multi-institutional parameter collaborative aggregation. The core idea of federated learning is to train models without sharing raw data. In this research’s application scenario, institutions $i\in \{1,2,\dots ,M\}$ iteratively perform local model updates and central parameter aggregation according to federated rounds $t\in \{1,2,\dots ,T\}$. In round t, institution i receives global parameters ${\theta }_g^{(t-1)}$ from the central server. It performs stochastic gradient descent updates locally based on spatio-temporal GAT models:

$${\theta }_i^{\left(t\right)}={\theta }_g^{(t-1)}-{\eta }_i\nabla {\mathcal{L}}_i\left({\theta }_g^{(t-1)}\right)$$

(17)

Here, ${\eta }_i>0$ is the learning rate of institution i. $\nabla {\mathcal{L}}_i\left({\theta }_g^{(t-1)}\right)$ is the gradient based on the local dataset $D_i$. Gradient computation involves only institution-internal data. It strictly satisfies the privacy constraint in Equation (4). Transaction graphs of different institutions exhibit significant heterogeneity. Differences include node scale, edge density, and feature dimension. Direct parameter averaging causes improper knowledge fusion across heterogeneous institutions. It may cause negative transfer for certain institutions. To address this problem, an adaptive weight computation mechanism based on graph topological features is introduced. Compared to simple averaging aggregation that assigns equal weights to all institutions, the adaptive mechanism allows institutions with similar graph structures to exert greater mutual influence during aggregation, ensuring that knowledge transfer occurs primarily among compatible participants. Meanwhile, it reduces the contribution of highly heterogeneous institutions, preventing their distinct characteristics from dominating the global model and thereby mitigating negative transfer. First, extract graph feature vectors from each institution:

$${\varphi }_i=\left[{\displaystyle \frac{n_i}{n_{max}}},{\displaystyle \frac{|E_i|}{{\left|E\right|}_{max}}},{\overline{d}}_i,c_i,{\rho }_i\right]\in {\mathbb{R}}^5$$

(18)

Here $n_i$ is the node count of institution i. $|E_i|$ is the edge count. ${\overline{d}}_i={\displaystyle \frac{2|E_i|}{n_i}}$ is the average degree. $c_i$ is the global clustering coefficient. ${\rho }_i$ is the network density. To avoid dimension effects, all features are normalized. $n_{max}$ and ${\left|E\right|}_{max}$ are the maximum node count and maximum edge count across all institutions. These five topological features are selected because they collectively characterize the structural properties of transaction networks from complementary perspectives. Node count and edge count reflect the overall scale of institutional transaction activity, which directly influences the amount of information available for model training. Average degree captures the typical connectivity pattern of accounts, indicating whether transactions are concentrated among a few accounts or distributed across many participants. The clustering coefficient measures the tendency of accounts to form tightly connected groups, which is relevant for detecting organized fraud rings. Network density provides a normalized measure of graph sparsity that enables meaningful comparison across institutions of different scales. Together, these features provide a compact yet informative representation of institutional graph characteristics that influence model behavior during federated aggregation. Based on graph feature vectors, compute topological similarity between any two institutions. Cosine similarity is employed as the similarity metric:

$$s_{ij}={\displaystyle \frac{{\varphi }_i·{\varphi }_j}{{\parallel {\varphi }_i\parallel }_2{\parallel {\varphi }_j\parallel }_2}},\forall i,j\in \{1,2,\dots ,M\}$$

(19)

Cosine similarity is adopted as the similarity metric because it measures the angular difference between feature vectors independent of their magnitudes, making it robust to scale variations across institutions. Alternative measures, such as Euclidean distance, are sensitive to absolute feature values and may be dominated by large-scale institutions, while Pearson correlation requires centered data and assumes linear relationships. Cosine similarity provides a bounded measure between 0 and 1 that directly reflects structural similarity patterns without requiring additional normalization or assumptions about feature distributions. Here · denotes the vector dot product. ${\parallel ·\parallel }_2$ denotes $L_2$ norm. The similarity matrix $\mathbf{S}={\left(s_{ij}\right)}_{M\times M}$ reflects the degree of graph-structure similarity across institutions. Based on the similarity matrix, assign adaptive weights to local parameters of each institution in global aggregation. The weight design principle is as follows: institutions with graph structures similar to those of the target institutions receive larger parameter-contribution weights. This ensures knowledge from similar institutions is more incorporated into global models:

$$w_i^{\left(t\right)}={\displaystyle \frac{n_i{\sum }_{j=1}^M{s}_{ij}}{{\sum }_{k=1}^M\left(n_k{\sum }_{j=1}^M{s}_{kj}\right)}}$$

(20)

Here $w_i^{\left(t\right)}>0$ and ${\sum }_{i=1}^M{w}_i^{\left(t\right)}=1$. Weights comprehensively consider the institution sample scale ($n_i$) and topological similarity with other institutions (${\sum }_{j=1}^M{s}_{ij}$). The central server receives local parameter updates ${\left\{{\theta }_i^{\left(t\right)}\right\}}_{i=1}^M$ from all institutions. It performs weighted aggregation to generate global parameters:

$${\theta }_g^{\left(t\right)}=\sum _{i=1}^M{w}_i^{\left(t\right)}{\theta }_i^{\left(t\right)}$$

(21)

Global parameters are subsequently redistributed to all institutions. They serve as initial values for the next federated iteration. After aggregation, the comprehensive performance of global models on multi-institutional data distributions can be expressed as weighted loss functions:

$${\mathcal{L}}_{global}^{\left(t\right)}=\sum _{i=1}^M{w}_i^{\left(t\right)}{\mathcal{L}}_i\left({\theta }_g^{\left(t\right)}\right)$$

(22)

Here, the local loss ${\mathcal{L}}_i\left({\theta }_g^{\left(t\right)}\right)$ of each institution is weighted according to weights in Equation (7). It reflects the comprehensive performance of global models across all institutions. Federated iteration continues until global parameters converge or reach preset iteration counts. In each iteration, the change in global loss is defined as:

$$\Delta {\mathcal{L}}_{global}^{\left(t\right)}=\left|{\mathcal{L}}_{global}^{\left(t\right)}-{\mathcal{L}}_{global}^{(t-1)}\right|<ϵ$$

(23)

Here, $ϵ>0$ is the convergence threshold. When convergence conditions are satisfied, all institutions obtain unified global model parameters ${\theta }_g^{*}$. Original data remains locally stored in institutions, satisfying privacy protection constraints. This federated framework achieves collaborative knowledge fusion for multi-institutional fraud detection through parameter rather than data flow.

Theorem 2.

Under the federated learning framework, raw data $D_i$ of each institution remains in local storage throughout the training process. The central server receives and processes only model parameters ${\theta }_i^{\left(t\right)}$ rather than raw data. Since model parameters represent aggregated statistical patterns learned from transaction data rather than individual transaction records, direct reconstruction of raw data from parameters is infeasible under standard federated protocols. This privacy-preserving property is expressed as:

$$D_i\not\subset \mathcal{F}\left(\{{\theta }_1^{\left(t\right)},\dots ,{\theta }_M^{\left(t\right)}\}\right),\forall t,\forall i$$

(24)

Here $\mathcal{F}(·)$ denotes any deterministic reconstruction function. This formulation indicates that raw institutional data cannot be deterministically recovered from shared parameters alone. Additionally, through the adaptive weight aggregation mechanism in Equation (20), the risk of negative transfer across heterogeneous institutions is reduced by assigning lower aggregation weights to institutions with dissimilar graph structures.

Corollary 2.

Under the convergence state of the federated learning framework with adaptive weight aggregation, global model parameters ${\theta }_g^{*}$ are designed to improve detection performance for participating institutions. Let the independent training model of institution i be ${\theta }_i^{solo}$. Let the model after federated learning be ${\theta }_i^{fed}$. The design objective is expressed as follows:

$${\mathcal{L}}_i\left({\theta }_g^{*}\right)\le {\mathcal{L}}_i\left({\theta }_i^{solo}\right),\forall i\in \{1,2,\dots ,M\}$$

(25)

The extent to which this objective is achieved depends on the degree of complementary fraud patterns across institutions and the effectiveness of the similarity-based weighting scheme in mitigating negative transfer.

2.4. FedGAT Framework Design

Space Complexity: Space consumption of the FedGAT algorithm stems from multiple storage requirements. First, institutions must store the complete representation of transaction graphs locally. This includes node feature matrix $X_i\in {\mathbb{R}}^{n_i\times d}$ and adjacency information. Space required is $O(n_i·d+|E_i\left|\right)$. Second, during spatio-temporal graph attention network computation, multiple intermediate representations must be stored. These include attention weight matrices ${\alpha }^{\left(h\right)}\in {\mathbb{R}}^{|E_i|\times H}$ needed by multi-head attention mechanisms and hidden state sequences ${\left\{{\mathbf{h}}_m^{\left(i\right)}\right\}}_{m=1}^L$ when GRU processes transaction sequences. Space required is $O\left(\right|E_i|·H+L·d_t)$. Additionally, storing the local model parameters ${\theta }_i$ during federated iteration round t requires $O\left(\right|\theta \left|\right)$ space. Here $\left|\theta \right|$ is the total model parameter count. This relates to network depth and feature dimension. On the central server, a copy of global parameters ${\theta }_g^{\left(t\right)}$ must be maintained. A similarity matrix $\mathbf{S}\in {\mathbb{R}}^{M\times M}$ is used for similarity metrics across institutions. Therefore, total system space complexity is $O(M·(n_i·d+|E_i|+\left|\theta \right|)+M^2)$. In typical financial applications, since $\left|\theta \right|$ is far smaller than the feature matrix and adjacency information scale, dominant terms are usually $O(M·n_i·d)$.

Time Complexity: Time complexity analysis requires considering various computational stages of the algorithm layer by layer. During local training at single institutions, each forward propagation of spatio-temporal graph attention networks involves multiple operations: multi-head attention computation in the spatial layer performs attention operations on each edge with complexity $O\left(\right|E_i|·d·H)$. Each attention head involves d-dimensional feature comparison and softmax computation. Spatial feature aggregation through neighborhood aggregation has complexity $O({\sum }_{j\in V_i}|{\mathcal{N}}_j|·d)=O\left(\right|E_i|·d)$. In the temporal layer, transaction sequence encoding and GRU recursion require $O(L·d_c)$ and $O(L·d_t^2)$ time, respectively. Here L is the average transaction sequence length, and $d_t$ is the GRU hidden-state dimension. Spatio-temporal fusion prediction involves $O(n_i·d)$ operations. Therefore, the time for single forward propagation is $O\left(\right|E_i|·d·H+L·d_t^2)$. Considering backward propagation has the same complexity order, local training for E rounds requires $O(E·\left|D_i\right|·\left(\right|E_i|·d·H+L·d_t^2\left)\right)$ time. During the federated phase, local training across all institutions is performed in parallel, not increasing total time. Subsequent central server operations are executed sequentially: computing graph features for all institutions requires $O(M·\left(n_i+\right|E_i|\left)\right)$, similarity matrix computation requires $O(M^2·5)=O\left(M^2\right)$, weight computation requires $O\left(M^2\right)$, and parameter aggregation requires $O\left(\right|\theta \left|\right)$. Federated iteration proceeds for T rounds. Total time complexity is thus:

$$O\left(T·M·E·\left|D_i\right|·\left(\right|E_i|·d·H+L·d_t^2)+T·(M·(n_i+|E_i|)+M^2+\left|\theta \right|)\right)$$

(26)

In practical applications, communication overhead and central aggregation operations during the federated phase (second term) are relatively small. Dominant terms are local training complexity. Through parallel computation across institutions, actual wall-clock time can be significantly reduced. The complete procedure of the FedGAT framework is summarized in Algorithm 1.

Algorithm 1: Adaptive Federated Graph Attention Network (FedGAT)

3. Experiments and Analysis

3.1. Experimental Setup and Datasets

This study employs two public financial network datasets to verify the effectiveness of the FedGAT framework. The first dataset is the Elliptic Bitcoin transaction network dataset, released by Elliptic, a blockchain analytics company, in collaboration with the MIT-IBM Watson AI Lab [38]. This dataset comprises anonymous transaction graphs collected from the Bitcoin blockchain. In this graph, each node represents a transaction, and edges between nodes represent Bitcoin flow relationships between transactions. The dataset contains 203,769 nodes and 234,355 edges, with each node equipped with a 166-dimensional feature vector. Node labels are categorized into three classes: illicit transactions comprise 2% (4545 nodes), licit transactions comprise 21% (42,019 nodes), and the remaining 77% of transactions are unlabeled. The transaction network feature design fully reflects the complexity of Bitcoin transactions. The first 94 features represent local transaction information, including transaction timestamps, input/output quantities, transaction fees, output volumes, and aggregated statistics related to inputs and outputs. The final 72 features are aggregated features computed from one-hop neighborhood transaction information, including the maximum, minimum, standard deviation, and correlation coefficients of neighborhood transactions across the same data dimensions. Temporally, the dataset contains 49 time steps, with each time step spanning approximately two-week intervals. Transactions within each time step form a single connected component, and no edges connect across different time steps. This characteristic naturally endows the dataset with temporal features.

The second dataset is the IEEE-CIS fraud detection dataset, released by the IEEE Computational Intelligence Society in partnership with Vesta Corporation for the IEEE-CIS Fraud Detection competition [39]. It originates from real payment transaction environments and contains abundant annotated transaction records. The dataset provides rich transaction attribute information, including merchant features, device information, transaction geographic location, and other multi-dimensional features that reflect complex transaction scenarios in actual financial payment systems. Unlike the Elliptic dataset, which emphasizes network structure, the IEEE-CIS dataset provides perspectives for understanding fraud characteristics from multi-source attribute angles. This dataset is used in experiments to provide supplementary validation of the FedGAT framework’s capabilities for handling multi-source heterogeneous data. In particular, it validates the effectiveness of spatio-temporal graph attention networks in extracting valid representations from high-dimensional transaction features.

The combined use of both datasets enables validation of the proposed methods from different perspectives. The graph structure of the elliptic dataset allows direct validation of graph attention mechanisms for identifying network topology patterns. The IEEE-CIS dataset provides an assessment of multi-dimensional feature processing capabilities in real financial scenarios. Through experiments on both datasets, the comprehensive advantages of the FedGAT framework in processing graph-structured data, temporal features, and multi-institutional heterogeneity can be demonstrated.

All experiments in this research were conducted on a mid-to-high configuration Windows 11 system environment. The hyperparameter settings used in all experiments are summarized in

Table 2. Hyperparameter configurations of the FedGAT framework.

Parameter Name	Value	Parameter Name	Value
Spatio-Temporal Graph Attention Network
Multi-head attention heads	8	Graph feature projection dimension	32
Hidden feature dimension	64	Spatial feature output dimension	64
Temporal feature output dimension	64	GRU hidden state dimension	64
Fusion layer hidden dimension	128	Activation function type	ReLU
Temporal decay coefficient $\lambda$	0.1	Node feature dimension	166
Attention weight scaling factor	0.125	Gradient clipping threshold	1.0
Federated Learning Framework
Total federated iteration rounds T	100	Local training rounds E	5
Total financial institutions M	5	Global learning rate	0.01
Local learning rate $\eta$	0.005	Learning rate decay coefficient	0.95
Convergence loss threshold $ϵ$	$1\times {10}^{-4}$	Early stopping patience	20
Weight decay coefficient	$1\times {10}^{-5}$	Momentum value	0.9
Graph Similarity and Aggregation
Similarity metric method	Cosine similarity	Aggregation weight normalization	Softmax
Graph feature extraction method	Topological statistical features	Node count normalization factor	Dataset maximum
Edge count normalization factor	Dataset maximum	Heterogeneity measurement dimension	5-dimensional
Model Training
Batch size	32	Optimizer type	Adam
Loss function	Binary cross-entropy	Regularization type	L2 regularization
Dropout ratio	0.3	Initialization method	Xavier initialization
Maximum training epochs	100	Validation set ratio	0.2
Data Processing
Maximum transaction sequence length L	20	Neighborhood sampling size	10
Multi-hop neighborhood depth	2	Subgraph sampling ratio	0.8
Feature normalization method	Standardization	Missing value handling	Mean imputation
Class imbalance handling	Weighted loss	Positive class sample weight	10.0

. Hardware configuration includes an Intel Core i7-12700K processor with processing capability exceeding 12 cores. Storage is configured with 64 GB of DDR4/DDR5 memory to meet multi-institutional data-parallel processing and model-training requirements. Computational acceleration employs NVIDIA RTX 3080 supporting CUDA 11.8 and above. This configuration supports the simultaneous processing of Elliptic and IEEE-CIS datasets, as well as local model training across multiple institutions in the federated learning framework. Under this hardware environment, a single federated iteration runtime is approximately 15–30 min. Complete experimental workflow (including 100 federated iterations and ablation studies) can be completed within 5–8 h.

3.2. Single-Institution Fraud Detection Performance Comparison

To verify the effectiveness of the spatio-temporal graph attention network in single-institution environments, comparative experiments include four models: baseline model XGBoost, a variant without the graph attention mechanism (without GAT), a variant without the GRU temporal module (without GRU), and the complete FedGAT model. Experiments are conducted on the Elliptic Bitcoin transaction network dataset and the IEEE-CIS fraud detection dataset to validate the method’s generalization.

As shown in Figure 3, the recall rate directly reflects the capability to detect fraudulent transactions. On the Elliptic dataset, the XGBoost baseline model’s recall is 0.58. This aligns with the current status described in the research background. The GAT model achieves 0.72. This is 5.9% higher than without GRU at 0.68. This difference reveals an important phenomenon: in the Elliptic dataset, network topology structure contributes more to fraud identification than temporal features. This stems from the inherent characteristics of Bitcoin transactions. Fraud behavior often manifests as specific account association patterns (such as multi-level transfer chains). These patterns are more prominent in the graph structure than in the time series. While without GRU captures the dynamics of individual transactions, it cannot identify abnormal relationships spanning multiple transaction steps. This limits performance. The complete FedGAT model achieves 0.85. This is 13% improvement compared to without GAT. This indicates that even when the structure is fully utilized, the temporal dimension still improves performance through fine-tuning of weights (via a temporal decay mechanism). On the IEEE-CIS dataset, XGBoost is 0.61, without GAT is 0.75, without GRU is 0.70, and FedGAT is 0.87. Notably, the GAT improvement compared to baseline (23%) is lower than the Elliptic improvement (24%). Without GRU gap increases from 5.9% on Elliptic to 7.1% on IEEE-CIS. This indicates the IEEE-CIS dataset contains richer feature information (including multi-source data from devices and merchants). This makes temporal modeling relatively more important. But network relationships remain core.

From Figure 4, false positive rate reflects model conservativeness and accuracy trade-off. XGBoost’s false-positive rate on Elliptic is 0.145. On IEEE-CIS, it is 0.138. The difference between datasets is only 0.7%. This indicates that traditional feature engineering methods have low sensitivity to data sources. Ironically, this also indicates a weak ability to capture complex patterns. Without GAT decreases from 0.082 to without GRU at 0.095 (Elliptic). The difference is 16%. This indicates that graph structure plays a significant role in reducing false positives. This is because graph attention mechanisms learn true relationship patterns of normal transactions. This avoids misclassifying novel but legal transaction relationships as fraud. Conversely, models with only temporal dimensions easily fall into the “novelty equals anomaly” trap. They over-sensitize to abnormal temporal patterns. FedGAT reduces the false positive rate to 0.038 (Elliptic) and 0.032 (IEEE-CIS). The absolute difference between datasets is only 0.006. This indicates spatio-temporal fusion representations exhibit high stability across different data domains. This stability stems from a combination of graph-structure generalizability (capturing universal fraud topologies) and temporal mechanism detail-orientation (adapting to specific evolutionary characteristics).

As shown in Figure 5, the F1 score comprehensively measures precision and recall balance. On Elliptic, without GAT (0.76) advantage compared to without GRU (0.71) stems from better false positive control (0.082 vs. 0.095). However, on IEEE-CIS, without GAT (0.78) and without GRU (0.73), the model maintains approximately a 5% gap. Gap extent aligns with recall difference. This indicates both modules have relative advantages under different data characteristics. FedGAT achieves 0.84 and 0.86 on the respective datasets. Improvement compared to without GAT is approximately 10.5% and 10.3%. This highly consistent improvement rate indicates that synergistic effects possess stability. From error bar analysis, FedGAT error bars (0.82–0.86 on Elliptic) are significantly smaller than without GAT (0.73–0.79). This reflects that spatio-temporal fusion not only improves average performance. More importantly, it enhances model robustness. Without GRU disadvantage compared to without GAT is more apparent on Elliptic (6.6% difference) and slightly narrower on IEEE-CIS (6.8% difference). This suggests that in scenarios with rich graph structure information, the relative importance of temporal modeling is lower. In scenarios with multi-dimensional features, temporal dynamics become more critical.

Quantitative results summarized in

Table 3. Comparative performance analysis on Elliptic and IEEE-CIS datasets.

Model	Elliptic Dataset			IEEE-CIS Dataset
	Recall	FPR	F1-Score	Recall	FPR	F1-Score
XGBoost	$0.58\pm 0.008$	$0.145\pm 0.005$	$0.62\pm 0.007$	$0.61\pm 0.009$	$0.138\pm 0.004$	$0.64\pm 0.008$
Without GAT	$0.72\pm 0.018$	$0.082\pm 0.006$	$0.76\pm 0.016$	$0.75\pm 0.021$	$0.076\pm 0.007$	$0.78\pm 0.019$
Without GRU	$0.68\pm 0.015$	$0.095\pm 0.008$	$0.71\pm 0.014$	$0.70\pm 0.018$	$0.089\pm 0.006$	$0.73\pm 0.016$
FedGAT	$0.85\pm 0.012$ *	$0.038\pm 0.003$ *	$0.84\pm 0.011$ *	$0.87\pm 0.010$ *	$0.032\pm 0.003$ *	$0.86\pm 0.009$ *

* Statistically significant improvement over the best-performing baseline ($p<0.05$).

are presented as mean ± standard deviation derived from ten independent experimental trials. To evaluate the empirical robustness of the framework, a two-tailed paired t-test was conducted between FedGAT and the best-performing baseline. The asterisk (*) denotes statistically significant improvements with a p-value less than 0.05. The experimental data indicate that FedGAT consistently achieves the highest performance metrics across both financial datasets. The narrow standard deviations, specifically 0.011 and 0.009 for the F1-score on the Elliptic and IEEE-CIS datasets, respectively, demonstrate the stability of the proposed spatio-temporal fusion mechanism. These statistically significant gains confirm that the integration of graph attention and temporal decay effectively captures complex fraud patterns while mitigating the impact of stochastic initialization.

Based on the temporal evolution data in Figure 6, temporal stability reflects the model’s ability to adapt to long-term transaction flows. XGBoost variance coefficient throughout the period is approximately 3.4%. It exhibits high stability but low performance characteristics. This is because static feature-based methods inherently lack sensitivity to temporal changes. Without GRU variance coefficient is approximately 0.8%. Stability is highest, conversely indicating complete neglect of the fraud pattern temporal evolution. Without GAT variance coefficient is 5.2%. Mid-period fluctuations are significant. Particularly around time step 10–25, the decline occurs from 0.75 to 0.69. This reflects pure temporal modeling over-sensitivity to periodic changes. In contrast, FedGAT exhibits ideal characteristics: early period (1–20) maintains 0.79–0.82 baseline level. Mid-period stabilizes at 0.81–0.83. Late period (30–49) steadily increases to 0.84–0.88. Variance coefficient is 2.3%. This upward trend directly validates the effectiveness of the temporal decay mechanism. Anomaly-detection capability for recent transactions continues to improve. In the critical late-period window, FedGAT improvement compared to without GAT is approximately 14 percentage points. This indicates model improvement by learning from recent time segments to detect fraud. Meanwhile, FedGAT’s average advantage over the period without GRU reaches 16 percentage points. This indicates that the graph structure’s long-term memory capability compensates for the temporal model’s neglect of historical information. This temporal evolution pattern indicates that effective fraud detection requires capturing both persistent topological anomalies and adapting to short-term dynamic changes. Single-dimensional modeling leads to significant performance compromise.

3.3. Multi-Institutional Collaborative Detection Performance Comparison

To verify the effectiveness of the FedGAT framework for collaborative learning with multi-institutional privacy protection, the Elliptic dataset is partitioned into five virtual financial institutions to simulate real cross-institutional fraud detection scenarios. The partitioning procedure is designed to reflect realistic heterogeneity among financial institutions while preserving essential data characteristics.

The dataset is partitioned based on temporal segmentation combined with graph connectivity. Specifically, the 49 time steps in the Elliptic dataset are divided into five non-overlapping groups, with each group assigned to one virtual institution. Institution 1 receives time steps 1–10, Institution 2 receives time steps 11–20, Institution 3 receives time steps 21–30, Institution 4 receives time steps 31–40, and Institution 5 receives time steps 41–49. This partitioning strategy preserves temporal ordering within each institution, as transactions within each time step form connected components with no cross-time-step edges in the original dataset. The resulting virtual institutions exhibit natural heterogeneity in terms of transaction volume, graph density, and fraud prevalence, reflecting the diversity observed in real-world financial ecosystems. Class imbalance is handled consistently across all institutions using a weighted binary cross-entropy loss, with a positive class weight of 10.0, as specified in the parameter table. This weighting scheme compensates for the low fraud prevalence (approximately 2%) in the dataset and ensures that the model does not trivially predict all transactions as legitimate.

To ensure the statistical reliability of the reported results, all experiments are conducted with five independent runs, each using different random seeds for parameter initialization and batch sampling. The reported performance metrics represent the mean values across these runs. The standard deviations are included in the error bars shown in the figures. For single-institution experiments, the coefficient of variation across runs remains below 3% for all metrics, indicating stable model behavior. For multi-institution federated experiments, variance across runs is similarly low, with standard deviations typically within 0.02 for recall, false positive rate, and F1 score. Experiments compare three scenarios: independent institution training, simple parameter averaging aggregation, and the proposed adaptive weight aggregation. This comprehensively evaluates the federated framework’s collaborative capability.

As shown in Figure 7, the federated learning framework brings significant performance improvements to each institution. In the recall rate, Institution 1 improves from 0.71 in independent training to 0.85 after federated learning. Improvement magnitude is 19.1%. Institution 2 improves from 0.72 to 0.84. Improvement is 16.7%. Institution 3 improves from 0.75 to 0.87. Improvement is 16.0%. Institution 4 improves from 0.70 to 0.82. Improvement is 17.1%. Institution 5 improves from 0.65 to 0.79. Improvement is 21.5%. This synergistic effect stems from shared fraud knowledge across institutions. Although transaction graph structures across institutions exhibit significant differences, they share some universal fraud patterns. Examples include multi-level transfer chains and high-frequency transactions. Through federated aggregation, the global model learns these common features and feeds back to each institution. Small institutions or those with scarce fraud samples (such as Institution 5) benefit most. In the false positive rate, the average decreases from 0.090 in independent training to 0.045 after federated learning. Reduction is approximately 50%. This indicates that collaborative learning improves detection rates while avoiding false-positive judgments of novel legal transactions by learning each institution’s normal transaction patterns. F1 score improves from an average of 0.76 to 0.86. Improvement is 10.3%. This embodies balanced improvement in detection precision and coverage.

From Figure 8, the federated learning convergence process exhibits three distinct phases. In the early phase (rounds 1–25), the global weighted loss rapidly decreases from 0.35 to 0.20. Reduction is 43%. This reflects the global model’s rapid learning of multi-institutional common features. In mid-phase (rounds 25–75), loss continues a steady decline from 0.20 to 0.05. Slope gradually becomes gentler. This indicates the model is fine-tuning to adapt to institution-specific characteristics. In the late phase (after round 75), loss maintains fluctuation within the 0.04–0.05 range. Variance coefficient is below 2%. This indicates the algorithm converges to a stable state. The convergence point appears at round 75. Compared to the preset maximum of 100 iterations, this completes 25% earlier. This indicates the algorithm possesses good convergence efficiency. Convergence realization benefits from the introduction of an adaptive weight mechanism. If simple parameter averaging were employed (without distinguishing among institutions), gradient directions across heterogeneous institutions would conflict. This would slow convergence speed or cause oscillation. Adaptive weights achieve faster convergence by assigning stronger influence to similar institutions. This makes gradient update directions more consistent.

Figure 9 quantifies the heterogeneity degree across institutions through the topology similarity matrix. Diagonal values are all 1.0 (institution self-similarity). Non-diagonal elements reflect the true extent of the difference between institutions. The similarity between Institution 1 and Institution 2 is highest at 0.82. This indicates these two institutions have similar transaction graph structures. They may both represent medium-scale payment platforms. The similarity between Institution 1 and Institution 5 is lowest at only 0.35. This indicates that these two institutions have significant differences in their graph features. They may respectively represent large exchanges and small wallet services. Average non-diagonal similarity is 0.61. This indicates that multi-institutions possess medium-level heterogeneity. This heterogeneity existence is core reason for introducing adaptive weights. If equal weights (0.2 for each institution) were employed, different characteristics of Institution 1 and Institution 5 would “interfere” mutually. This would cause the global model performance to be suboptimal on both. Conversely, an adaptive weight mechanism through similarity matrix computation enables similar institutions to exert greater influence in global aggregation. This avoids negative impacts across heterogeneous institutions. Specifically, higher institution similarity (such as 0.82) should assign higher weights during aggregation; lower similarity (such as 0.35) should assign correspondingly lower weights. This makes the global model tend to learn multi-institution common features rather than being dominated by a few heterogeneous institutions.

As shown in Figure 10, aggregation weights for each institution are not static constants. They dynamically vary throughout the federated iteration process. Early weight allocation reflects initial similarity based on graph features: higher-weight institutions (Institutions 1, 2) have initial weights in the 0.21–0.22 range. Lower-weight institutions (Institution 5) have initial weights of only 0.19. During iteration, Institution 2’s weight gradually increases from an initial 0.21 to 0.215. It maintains relative stability. This indicates institution characteristics similar to the majority of institutions. It provides continuous positive contributions to global learning. Institution 1 weight exhibits a declining trend from 0.215 to 0.19. This indicates the model discovers Institution 1 possesses unique features during training. Lowering its weight helps prevent these specificities from over-influencing the global model. Institution 3’s weight maintains high stability around 0.205. This indicates that the institution’s feature compatibility with all institutions is best. The weights for Institutions 4 and 5 are relatively lower and tend to stabilize. This reflects their distance from mainstream features. The necessity of weight evolution stems from the fact that even if two institutions have high initial similarity, over the long term, through federated iteration, they may gradually diverge due to different local data distributions. Dynamic weight adjustment can reflect these changes in real time. This ensures aggregation always optimizes for current global and local characteristics rather than relying on fixed initial similarity.

Figure 11 clearly demonstrates the negative transfer problem and its solution. Negative transfer refers to a phenomenon in federated learning in which the global model, due to fusing inappropriate heterogeneous data from institutions, causes certain institutions’ performance to decrease. When only simple parameter averaging aggregation is employed, Institution 3’s recall is 0.74. Compared to independent training at 0.75, it actually decreases 1.3%. Institution 4 improves from 0.70 to 0.73, though only marginally. Institution 5 still decreases from 0.65 to 0.62. Decrease is 4.6%. These negative transfer phenomena precisely result from Institution 5’s excessive heterogeneity relative to the other four institutions. Simple averaging causes Institution 5 global model to be “submerged” by mainstream features. This causes its local specific features to be neglected. In contrast, after adopting adaptive weight aggregation, all five institutions achieve positive effects. Institution 1 improves from independent 0.68 to 0.81. Improvement is 19%. Institution 2 improves from 0.72 to 0.85. Improvement is 18%. Institution 3 improves from 0.75 to 0.87. Improvement is 16%. Institution 4 improves from 0.70 to 0.82. Improvement is 17%. Institution 5 improves from 0.65 to 0.79. Improvement is 21.5%. The most heterogeneous institution benefits the most. This indicates an adaptive weight mechanism through lowering Institution 5 aggregation weight protects its specific features from averaging. Meanwhile, this institution still learns universal features from the global model through a lower weight. This is adaptive-aggregation sophistication: it allows heterogeneous institutions to maintain their identities while reaping collaborative learning benefits. Compared to simple averaging, producing 2.9% average negative transfer risk, adaptive aggregation completely eliminates negative transfer. All institution performance improvements exceed 15%.

3.4. Comparison with Prior Studies and Limitations

To systematically evaluate the contributions of FedGAT relative to existing approaches,

Table 4. Methodological and performance comparison with representative prior studies.

Method	Graph	Temporal	Adaptive Fed.	Recall	F1-Score
GraphFA [25]	✓	×	×	0.70	0.72
FraudGNN-RL [26]	✓	✓	×	0.74	0.77
Fraud Shield [27]	×	✓	×	0.67	0.69
FedGAT (Ours)	✓	✓	✓	0.85	0.84

Note: ✓ indicates the method supports this capability; × indicates it does not. Fed. = Federated Learning.

summarizes the key characteristics and performance metrics of representative recent studies. To ensure a fair discussion, performance discrepancies arising from differences in dataset scale and class distributions are explicitly analyzed.

As shown in Table 4, existing methods typically address only subsets of the key challenges in financial fraud detection. GraphFA employs graph-based camouflage detection [25], but its reliance on static graph snapshots limits the capture of evolving fraud patterns. FedGAT achieves a recall of 0.85 compared to 0.70 for GraphFA through the integration of temporal attention mechanisms.FraudGNN-RL combines graph neural networks with reinforcement learning and incorporates federated learning for privacy preservation [26]. However, its uniform aggregation strategy does not account for institutional heterogeneity. FedGAT employs adaptive weight aggregation based on graph topological similarity, achieving an improved F1-score of 0.84 compared to 0.77 for FraudGNN-RL. Fraud Shield utilizes CNN and LSTM for temporal modeling [27], but does not exploit transaction network topology. By leveraging spatio-temporal graph attention, FedGAT achieves an 18 percentage point improvement in recall.

Several limitations should be noted. The evaluation relies on virtual institutions created via temporal partitioning of public datasets. Additionally, the framework does not yet incorporate explicit differential privacy or adversarial defense mechanisms. Future research will explore scalability in larger federated networks.

3.5. Discussion

• Spatio-Temporal Fusion Achievement of Research Objectives: Experimental results fully validate the effectiveness of the proposed FedGAT framework in addressing key problems raised in the research background. In the single-institution detection stage, FedGAT achieves 0.85 recall and a false positive rate of 0.038 on the Elliptic dataset. It achieves 0.87 and 0.032, respectively, on the IEEE-CIS dataset. This completely surpasses the performance bottlenecks of existing methods (recall below 60%, false positive rate reaching 15%). A spatio-temporal attention network, by simultaneously modeling account relationships and transaction temporal evolution, enables the model to identify complex fraud patterns that traditional methods cannot detect. In particular, sustained recall increases in late temporal analysis validate the effectiveness of the temporal decay mechanism. In the multi-institutional collaborative phase, the adaptive weight aggregation framework extends performance improvements to all participating institutions. Even Institution 5, with maximum heterogeneity, improves from 0.65 to 0.79. Compared to simple averaging aggregation of negative transfer risk, adaptive aggregation eliminates this problem. This ensures all institutional performance improvements.
• Heterogeneity Processing Effectiveness: Heterogeneity processing is a core challenge in cross-institutional federated learning. An adaptive weight-aggregation mechanism provides a systematic solution using quantitative graph-topology similarity metrics. The experimental observation of institution similarity, ranging from 0.82 to 0.35, and the distribution’s unevenness fully indicate a significant degree of heterogeneity. Adaptive weight mechanism improves heterogeneity processing through: similarity computation ensures that similar institutions receive higher weights; weights dynamically evolve during iteration to reflect the consistency degree between each institution and the global objective during training; and a meta-learning framework that separates parameterized meta-features prevents averaging heterogeneous institution-specific features. This design makes FedGAT perform stably across multiple datasets (Elliptic F1 = 0.84, IEEE-CIS F1 = 0.86). The current method computes similarity based on five graph topology features. Although simple and effective, an increase in the number of institutions or the joining of new institutions requires recomputation. This may introduce additional overhead in dynamic environments.
• Privacy Protection and Future Research Directions: The FedGAT framework achieves privacy-friendly collaborative learning through parameter rather than raw data uploading. Original transaction data always remains local. This complies with GDPR and other regulatory requirements. However, a theoretical possibility exists of inferring individual data through gradient inversion attacks using only a federated framework. Future work should explore combining differential privacy mechanisms with adaptive aggregation. This maintains acceptable performance levels under strict privacy guarantees. Additionally, the current framework assumes a trusted central server. Future work could consider introducing blockchain or distributed aggregation technologies to achieve fully decentralized federated learning. This further enhances system robustness and privacy protection capabilities.

4. Conclusions

This research proposes an Adaptive Federated Graph Attention Network (FedGAT) framework. It effectively resolves the trade-off between privacy protection and detection performance in cross-institutional collaborative detection of financial fraud. By combining a spatio-temporal graph attention network with adaptive federated learning, the framework achieves systematic modeling of complex fraud patterns while protecting institutional data privacy. In a single-institution environment, the spatio-temporal fusion mechanism enables fraud detection recall surpassing 85%. The false-positive rate is kept below 5%. This represents improvements exceeding 40% and 70%, respectively, compared to existing methods. In multi-institutional collaborative scenarios, an adaptive weight-aggregation mechanism based on graph-topology similarity effectively addresses differences among heterogeneous institutions. It eliminates negative transfer caused by traditional averaging aggregation. All participating institutions achieve performance improvements exceeding 15%. Experimental validation demonstrates that even institutions with maximum heterogeneity benefit through a federated framework. This fully proves solution effectiveness in handling diverse participants in complex financial ecosystems. This research not only breaks the single-institution bottleneck in detection performance. More importantly, it provides a privacy-friendly and scalable cross-institutional fraud control system. It establishes a technical foundation for collaborative risk control among banks, payment platforms, and other financial institutions. Based on these findings, financial institutions should prioritize constructing transaction graph representations and adopt adaptive aggregation mechanisms when collaborating with heterogeneous partners. Regulatory bodies are encouraged to establish standardized protocols for privacy-preserving collaborative detection that comply with data protection regulations. Future work should integrate differential privacy mechanisms to provide formal privacy guarantees and explore decentralized aggregation protocols to enhance system robustness. Application to other financial crime domains, such as insurance fraud and securities manipulation, also warrants further investigation.