Real-Time Fraud Detection via Streaming Machine Learning: Architectures, Adaptive Models, and Production-Scale Inference
International Journal of Scientific Research in Computer Science Engineering and Information Technology · 20.06.2024 · 329
DOI: 10.32628/cseit24107453 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Scientific Research in Computer Science Engineering and Information Technology, online erschienen am 20.06.2024. Verfasst von Jaya Ram Menda Jaya Ram Menda. Die Arbeit nennt 15 Literaturangaben.
Abstract (Originalsprache)
The rapid digitization of financial transactions has led to an unprecedented explosion in real-time payment systems, e-commerce platforms, digital banking services, and instant peer-to-peer transfers, fundamentally transforming how monetary exchanges occur while simultaneously amplifying the scale and sophistication of fraudulent activities. As transaction volumes, velocities, and varieties continue to grow, fraud detection has become increasingly time-sensitive, requiring decisions to be made in milliseconds rather than minutes or hours. Traditional batch-based machine learning approaches, which rely on periodic data aggregation and offline model training, are no longer sufficient in this context, as they introduce unacceptable delays between transaction occurrence and fraud decision-making, leaving systems vulnerable to real-time exploitation. In response to these challenges, this paper presents a comprehensive and practical framework for low-latency fraud detection pipelines using streaming machine learning models, integrating modern distributed stream processing engines, adaptive online learning techniques, and production-grade model serving infrastructures. We argue that effective fraud detection in contemporary financial systems necessitates a tightly coupled, end-to-end architecture that seamlessly combines: (1) event-driven data ingestion capable of handling high-throughput transaction streams, (2) stateful stream processing that maintains contextual and historical transaction patterns in real time, (3) adaptive machine learning mechanisms that continuously learn from evolving data under concept drift, and (4) real-time model serving layers that ensure consistent, low-latency inference with strict performance guarantees. Drawing upon foundational work in stream processing frameworks such as Apache Kafka, Spark Streaming, and Apache Flink, incremental learning methods including Hoeffding Trees and ADWIN, and industrial-scale ML serving platforms like TensorFlow Serving, we propose a reference architecture that carefully balances latency, predictive accuracy, scalability, and operational reliability for modern fraud detection systems.
Zitieren
APA
Menda, J. R. M. J. R. (2024). Real-Time Fraud Detection via Streaming Machine Learning: Architectures, Adaptive Models, and Production-Scale Inference. International Journal of Scientific Research in Computer Science Engineering and Information Technology, 329. https://doi.org/10.32628/cseit24107453
Vancouver
Menda JRMJR. Real-Time Fraud Detection via Streaming Machine Learning: Architectures, Adaptive Models, and Production-Scale Inference. International Journal of Scientific Research in Computer Science Engineering and Information Technology. 2024:329. doi:10.32628/cseit24107453
BibTeX
@article{Menda2024,
author = {Menda, Jaya Ram Menda Jaya Ram},
title = {Real-Time Fraud Detection via Streaming Machine Learning: Architectures, Adaptive Models, and Production-Scale Inference},
journal = {International Journal of Scientific Research in Computer Science Engineering and Information Technology},
year = {2024},
pages = {329},
doi = {10.32628/cseit24107453}
}
EUVIMED