Edge-Based Real-Time Fraud Detection Using Federated Learning Across Distributed Payment Gateways

8 9 . 4 . 4 Edge-Based Real-Time Fraud Detection Using Federated Learning Across Distributed Payment Gateways Author: Nikhil Kassetty Date: 23 August, 2025 Abstract This study investigates an edge-based, real-time fraud detection framework leveraging federated learning (FL) across distributed payment gateways. The purpose is to address challenges of data privacy, latency, and scalability in modern financial ecosystems that rely on high-velocity, geographically distributed transaction streams. The methodology integrates an FL-enabled gradient-boosted decision model and an edge-level anomaly detection pipeline deployed across eight distributed payment nodes, evaluated with a synthetic yet industryrealistic transaction dataset. The key findings show that the proposed architecture reduces detection latency by 41%, improves precision by 12% compared to centralized baselines, and ensures full data locality at the payment gateway. The conclusion demonstrates that federated learning at the network edge can significantly improve both fraud detection accuracy and system efficiency while preserving data confidentiality. Keywords: • Federated Learning • Edge Computing • Fraud Detection • Distributed Systems • Payment Gateways Introduction Background and Context 8 9 . 4 . 4 Financial transaction networks have evolved into high-volume, real-time digital infrastructures supporting cross-border and multi-channel payments. With this evolution comes a growing exposure to sophisticated fraud typologies, including coordinated attacks, identity theft, bot-driven micro-fraud, and money laundering patterns hidden within massive data streams. Traditional fraud detection pipelines rely on centralized model training, which requires the aggregation of sensitive transaction data into centralized servers. This approach introduces latency, increases operational cost, and raises regulatory and privacy concerns under frameworks such as GDPR, PSD2, and evolving national data privacy laws. Edge computing has emerged as a strategic approach for executing computations closer to data sources. By deploying fraud detection models directly at payment gateways, institutions can react to anomalies with reduced latency while minimizing bandwidth consumption. Federated learning strengthens this paradigm by enabling collaborative model training across distributed data silos without requiring raw data exchange. Thus, FL provides both privacy preservation and global model generalizability. 2. Literature Review Early fraud detection research largely focused on centralized machine learning models applied to tabular banking data (Bhattacharyya et al., 2020; Whitrow et al., 2009). More recent studies have incorporated deep learning architectures such as graph neural networks and temporal networks (Wang et al., 2022; Chen & Liu, 2021). Parallel work in federated learning has demonstrated potential for privacy-preserving collaborative model training in healthcare, IoT, and mobile applications (McMahan et al., 2017; Yang et al., 2019). Within financial services, federated approaches are gaining attention: several studies explore FL for credit scoring (Hardy et al., 2019), anomaly detection (Ramaswamy et al., 2021), and distributed bank fraud detection (Zhao et al., 2022). However, limited literature investigates FL specifically deployed at the network edge, where resource constraints, latency requirements, and real-time decisioning are critical. Additionally, few studies integrate federated frameworks with edge inference pipelines tuned for transaction-level fraud scoring. Research Questions or Hypotheses RQ1: Can an edge-based federated learning architecture improve real-time fraud detection performance across distributed payment gateways? 8 9 . 4 . 4 RQ2: Does FL reduce detection latency relative to centralized training approaches? RQ3: Can global model accuracy be achieved without compromising data privacy? Hypotheses: H1: Federated training across edge gateways yields higher detection accuracy than isolated local models. H2: Edge-level inference significantly decreases end-to-end fraud scoring latency. H3: Federated optimization achieves comparable or superior model performance relative to centralized training. Significance of the Study This research contributes to the growing intersection of edge computing, federated learning, and financial fraud detection. It offers a scalable solution for real-time detection in distributed networks, addresses privacy concerns, and provides a performance-oriented evaluation that bridges theoretical FL capabilities with practical deployment considerations. The findings may inform regulatory bodies, financial institutions, and payment processors seeking nextgeneration fraud detection strategies. 3. Methodology Research Design This study employs a quantitative experimental research design evaluating performance metrics of fraud detection models under three conditions: 1. Centralized training, 2. Isolated edge-local training, and 3. Federated training across distributed edges. Participants or Datasets The dataset is a high-fidelity synthetic dataset generated to mimic global transaction gateways. It comprises 24 million transactions with attributes including amount, geolocation, merchant category, device identifiers, inter-arrival times, and user behavioral history. Approximately 1.2% of transactions reflect fraudulent patterns designed based on industry whitepapers and prior fraud detection studies. 8 9 . 4 . 4 Eight distributed payment gateways serve as "participants" in federated training, each holding a partition of the dataset representative of unique geographical regions. Data Collection Methods Data streams were simulated in real time and fed into edge servers equipped with lightweight fraud scoring modules. Each gateway maintained local logs, model parameters, and update histories. No raw transaction data left its originating node. Federated learning rounds were executed using a FedAvg variant supporting asynchronous updates. Data Analysis Procedures Model performance was evaluated through accuracy, precision, recall, F1-score, AUC-ROC, and inference latency. Statistical analyses included paired t-tests comparing centralized vs. federated vs. local models, with α = .05 as the significance threshold. Latency measurements included model inference time, FL communication overhead, and end-to-end scoring duration. All results were averaged across 20 replications. Ethical Considerations Although this study uses synthetic data, the research design acknowledges major ethical challenges in financial machine learning: privacy, fairness, explainability, and auditability. Federated learning was chosen specifically to eliminate the need to centralize sensitive user data. The framework adheres to GDPR principles of data minimization and purpose limitation. No personally identifiable real-world data was used. 4. Results Presentation of Findings 8 9 . 4 . 4 Table 1. Model Performance Across Training Paradigms Metric Local Edge Model Centralized Model Federated Model Accuracy 0.928 0.953 0.958 Precision 0.781 0.842 0.945 Recall 0.611 0.902 0.887 F1-Score 0.687 0.868 0.912 AUC-ROC 0.925 0.964 0.972 Latency statistics are presented in Table 2: Table 2. Latency Comparison Measurement Centralized Inference (ms) Edge FL Inference (ms) Feature extraction 11.2 11.5 Model scoring 42.8 18.6 Network overhead 55.1 3.2 End-to-end latency 109.1 64.3 Statistical or Analytical Results Paired t-tests show: • Federated models significantly outperform local models on all metrics (t = 4.77, p < .01). • Federated and centralized models have no significant difference in accuracy (t = 1.11, p = .21). • Edge inference reduces latency by 41% relative to centralized inference (t = 6.29, p < .001). 8 9 . 4 . 4 Present Key Results Without Interpretation 1. Federated learning achieved the highest precision (0.945) and highest AUC-ROC (0.972). 2. Centralized training achieved the highest recall (0.902). 3. Federated inference reduced latency to 64.3 ms from 109.1 ms. 4. Statistical tests confirm significant improvements in federated performance over local baselines. 5. No significant performance gap exists between federated and centralized models. 5. Discussion Interpretation of Results The results indicate that federated learning at the network edge substantially improves fraud detection accuracy while optimizing latency. The high precision suggests that FL facilitates the aggregation of global fraud patterns without compromising data privacy. Similarly, faster edge inference supports real-time decisioning for high-throughput payment environments. Comparison with Existing Literature These findings align with McMahan et al. (2017) in demonstrating that FL can approximate centralized model performance. They also extend prior work in financial FL (Hardy et al., 2019; Zhao et al., 2022) by demonstrating deployment feasibility in edge environments, where resource limitations and latency constraints require specialized optimizations. Compared to earlier fraud detection studies (Wang et al., 2022; Whitrow et al., 2009), the present work emphasizes distributed and privacy-preserving architectures. Implications The implications span technological, regulatory, and operational domains: • Technological: Payment gateways can integrate real-time fraud detection without reliance on centralized servers. 8 9 . 4 . 4 • Regulatory: Federated learning supports compliance with privacy laws requiring data localization. • Operational: Reduced latency and improved precision can lower false-positive rates and financial losses. Limitations of the Study 1. Synthetic datasets, although realistic, may not fully capture emerging fraud patterns. 2. Experiments assume stable network connections; real-world deployments may introduce noise or update failures. 3. Only a single FL algorithm (FedAvg variant) was tested. 4. Edge devices in production may have more severe hardware constraints. Future Research Directions • Integrating graph neural networks into federated frameworks. • Studying FL robustness to adversarial attacks. • Exploring cross-organization federated collaboration among banks. • Developing explainability tools tailored for federated models. 6. Conclusion This study demonstrates that edge-based federated learning provides a powerful solution for real-time fraud detection across distributed payment gateways. Federated models matched or surpassed centralized performance while reducing latency and preserving data privacy. These results suggest that combining FL with edge computing offers a practical and regulatoryaligned framework for modern payment networks. Future work should refine the architecture with more diverse datasets, advanced model families, and cross-institutional collaborations. 8 9 . 4 . 4 REFERENCES 1. Brown, S. L., & Adams, W. T. (2021). Behavioral analytics in financial fraud detection. Journal of Financial Security, 14(2), 55–72. 2. Chen, Y., & Park, J. (2023). Deep learning architectures for multi-modal anomaly detection. ACM Computing Surveys, 55(4), 1–32. 3. Davis, R., & Osei, K. (2020). Transactional feature engineering for real-time fraud scoring. IEEE Transactions on Information Forensics, 15, 2841–2856. 4. Dutta, A., & Singh, P. (2022). Advances in graph neural networks for fraud detection. Data Mining Letters, 8(3), 77–95. 5. Fernandez, J., & Yu, M. (2021). Sequential modeling of customer behavior in digital payment systems. Journal of Machine Learning Applications, 9(1), 33–49. 6. Gao, L., & Thompson, D. (2023). Multi-modal deep learning for financial security. International Journal of FinTech Research, 5(2), 66–84. 7. Ahmed, W. U., Uddin, M. R., Sadat, Q. T., Das, P., & Hasan, M. (2020, June). Performance assessment of a small-scale vertical axis single-stage savonius wind turbine by using artificial wind. In 2020 IEEE Region 10 Symposium (TENSYMP) (pp. 1816.1819). IEEE. 8. Allen, B. (2022). Cloud-scale response orchestration. ACM Computing Review, 66(7), 1–12. 9. Das, J., Halder, D., Uddin, M. R., Sadat, Q. T., & Hasan, M. (2020, June). Design and Analysis of Soft Switching PWM DC-DC Power Converter with High-Frequency Transformer Link for Portable Arc Welding Machine. In 2020 IEEE Region 10 Symposium (TENSYMP) (pp. 1820-1823). IEEE. 10. Bennett, T., & Rogers, K. (2021). Deep Q-learning applications in cyber defense. Cyber AI Research, 8(2), 99–115. 11. Chen, D. (2022). Incident response automation for Kubernetes clusters. CloudSys Journal, 5(1), 22–41. 12. Cheng, W., & Li, Y. (2024). Anomaly detection using deep learning in cloud traffic. Computer Security Today, 47(2), 70–84. 8 9 . 4 . 4 13. N. Kassetty, K. Alang, V. Paruchuru, S. Sharma, P. Goel and S. Kumar, "Cloud Security Management: Advanced AI Techniques for Anomaly Detection and Response Automation," 2025 International Conference on Networks and Cryptology (NETCRYPT), New Delhi, India, 2025, pp. 1620-1624, doi: 10.1109/NETCRYPT65877.2025.11102438.