Real-time fraud detection using delta live tables and machine learning

 Corresponding author: Kedarnath Goud Kothinti. Copyright © 2025 Author(s) retain the copyright of this article. This article is published under the terms of the Creative Commons Attribution License 4.0. Real-time fraud detection using delta live tables and machine learning Kedarnath Goud Kothinti * Liverpool John Moores University, UK. Global Journal of Engineering and Technology Advances, 2025, 23(01), 275-289 Publication history: Received on 14 March 2025; revised on 20 April 2025; accepted on 22 April 2025 Article DOI: https://doi.org/10.30574/gjeta.2025.23.1.0116 Abstract This article examines the transformative impact of Delta Live Tables (DLT) integrated with machine learning techniques on real-time fraud detection in financial institutions. Traditional batch processing approaches create critical vulnerabilities through delayed detection, while DLT offers a declarative framework that drastically reduces processing latency and improves detection accuracy. The article analyzes multiple dimensions of this technological shift, including architectural design, machine learning model performance, implementation strategies, and business benefits. Various machine learning approaches—from anomaly detection techniques like Isolation Forest and autoencoders to classification models such as Random Forest and neural networks—create a multi-layered defense system when deployed within DLT pipelines. The article outlines a comprehensive implementation architecture comprising data ingestion, feature engineering, scoring, decision-making, and feedback loop components. While highlighting significant business advantages including reduced fraud losses, decreased false positives, operational efficiency, improved regulatory compliance, and enhanced adaptability, the article also addresses implementation challenges related to model drift, feature latency, explainability requirements, and processing trade-offs. Keywords: Real-Time Fraud Detection; Delta Live Tables; Machine Learning; Anomaly Detection; Financial Security; Streaming Analytics 1. Introduction In today's digital banking landscape, the battle against financial fraud has evolved into a real-time challenge where every millisecond counts. Traditional batch processing approaches have become increasingly inadequate as fraudsters deploy sophisticated tactics that exploit the latency inherent in these systems. This article explores how Delta Live Tables (DLT) on Databricks, combined with advanced machine learning techniques, creates a powerful framework for real-time fraud detection and prevention. 1.1. The Growing Challenge of Financial Fraud Financial institutions worldwide are facing unprecedented challenges in combating fraud as digital transactions proliferate across banking channels. According to comprehensive research by Mohammed et al., global financial fraud reached an alarming $32.39 billion in 2023, with unauthorized transactions occurring across multiple channels, including online banking, mobile applications, and payment gateways [1]. Their study of 17 major financial institutions revealed that credit card fraud represents 38.6% of all identity theft reports, with affected customers experiencing a median loss of $1,500 per incident before detection. The research further demonstrated that traditional batch detection methods typically identify fraudulent transactions between 4-24 hours after occurrence, creating a critical vulnerability window during which 47% of stolen funds are transferred beyond recovery through sophisticated money laundering techniques involving cryptocurrency exchanges and cross-border transfers [1]. Global Journal of Engineering and Technology Advances, 2025, 23(01), 275-289 276 1.2. Delta Live Tables: Architecture for Real-Time Processing Delta Live Tables establishes a transformative declarative framework that processes transactional data with unprecedented speed and reliability. Mohammed et al. conducted extensive performance testing across six major banking implementations, documenting latencies as low as 100 milliseconds from transaction initiation to risk scoring, representing a 99.7% reduction compared to the traditional batch processing windows previously employed at these institutions [1]. Their longitudinal study spanning 18 months demonstrated that financial institutions implementing DLT-based fraud detection reported a statistically significant 42.8% decrease in false positives while simultaneously improving fraud capture rates by 37.5%, creating dual benefits for operational efficiency and financial protection. The researchers noted that DLT's ability to maintain transactional context across streaming data was particularly valuable for identifying sophisticated fraud scenarios that would otherwise escape detection in isolated processing approaches [1]. 1.3. Machine Learning Models in Fraud Detection When integrated with streaming data pipelines, modern machine learning approaches demonstrate remarkable performance improvements across various detection dimensions. Mohammed et al. conducted a comparative analysis of multiple algorithmic approaches using an anonymized dataset of 87.3 million transactions from three major European banks [1]. Their findings revealed that Isolation Forest algorithms detected 63.4% of novel fraud patterns with false positive rates of just 0.89%, providing essential protection against previously unseen attack vectors. The researchers' implementation of XGBoost classifiers achieved 91.3% accuracy in transaction classification with F1 scores of 0.874, significantly outperforming traditional rule-based systems, which averaged 76.8% accuracy on identical datasets. Perhaps most notably, their graph-based PageRank implementations identified coordinated fraud rings with 82.7% precision, increasing detection of sophisticated multi-account schemes by 3.2 times compared to standalone transaction monitoring approaches. The temporal pattern analysis capabilities of LSTM models analyzing sequential transaction behaviors improved predictive accuracy by 28.9% over static models by capturing evolving behavioral patterns and adapting to changing customer behaviors [1]. 1.4. Implementation Performance Metrics Organizations that have successfully deployed DLT-based fraud detection frameworks report significant operational improvements that extend beyond pure fraud prevention metrics. Mohammed et al. surveyed 12 financial institutions that completed DLT implementations between 2021-2023, documenting comprehensive performance improvements [1]. Their research showed average fraud detection time reduced from 6.3 hours to 2.4 minutes across these organizations, representing a 99.4% reduction in the critical window during which fraudsters can extract and transfer funds. The combination of improved detection accuracy and automated workflow integration decreased manual review workload by 41.2%, allowing fraud analysts to redirect their expertise toward complex cases requiring human judgment. Customer experience metrics showed meaningful improvement as false positive rates declined from 7.8% to 4.3% of legitimate transactions, reducing unnecessary friction for genuine customers. The cumulative financial impact was substantial, with overall fraud losses reduced by 29.7% in the first year of implementation, despite fraud attempts increasing by 23.4% during the same period, demonstrating the system's effectiveness against evolving attacks [1]. 1.5. Case Study: Major Financial Institution Implementation Mohammed et al. documented a comprehensive case study of a top-10 North American bank that implemented a DLTbased fraud detection system processing an extraordinary volume of 38,500 transactions per second during peak periods [1]. The implementation architecture ingested data from 17 distinct transaction sources, including card payments, wire transfers, ACH transactions, and online banking activities. Their system calculated 246 real-time fraud indicators for each transaction, ranging from customer behavioral biometrics to network connection patterns and merchant risk profiles. The production environment deployed 8 different machine learning models working in ensemble configuration, with a weighted voting mechanism determining final risk assessment. This sophisticated architecture delivered comprehensive risk scores with a median latency of just 187 milliseconds from transaction initiation to decision response. Financial performance analysis conducted by the researchers revealed an exceptional ROI of 641% within 18 months, primarily achieved by preventing $143.7 million in fraud losses that would have otherwise occurred based on historical patterns and control group comparison [1]. By combining Delta Live Tables with advanced machine learning techniques, financial institutions can transform their fraud prevention approach from reactive to proactive. This shift not only reduces financial losses but also enhances customer trust through seamless protection that operates invisibly in the background. As digital transactions continue to grow in volume and complexity, real-time fraud detection capabilities will increasingly become a competitive differentiator in the financial services industry. Global Journal of Engineering and Technology Advances, 2025, 23(01), 275-289 277 Table 1 Real-Time Financial Fraud Detection: Key Performance Metrics Before and After DLT Implementation [1] Metric Traditional Approach DLT with ML Approach Improvement (%) Fraud Detection Time 6.3 hours 2.4 minutes 99.40% False Positive Rate 7.80% 4.30% 44.90% Fraud Capture Rate Baseline Baseline + 37.5% 37.50% Manual Review Workload Baseline Reduced by 41.2% 41.20% Processing Latency Batch(hours) 100 milliseconds 99.70% Novel Pattern Detection (Isolation Forest) Limited 63.40% 20% Transaction Classification Accuracy (XGBoost) 76.80% 91.30% 18.90% Fraud Loss Reduction Baseline Reduced by 29.7% 29.70% 2. The Limitations of Traditional Fraud Detection Conventional fraud detection systems have historically relied on batch ETL (Extract, Transform, Load) processes for analyzing financial transactions. These systems process data in scheduled intervals, typically examining transactions hours or even days after they occur. According to the groundbreaking research conducted by Anchoori et al., traditional batch processing introduces an average detection latency of 17.3 hours from transaction initiation to fraud alert generation across the financial sector. Their comprehensive analysis of 23 global banking institutions revealed that during this critical delay window, fraudsters successfully execute an average of 8.2 additional unauthorized transactions per compromised account, significantly amplifying financial losses before detection mechanisms can intervene. The researchers documented that this delayed response capability results in financial institutions absorbing approximately 72.6% higher fraud losses compared to those implementing real-time detection frameworks [2]. The fundamental problem of static rule sets presents another significant vulnerability in traditional fraud detection architectures. Anchoori's team conducted an extensive two-year longitudinal study encompassing 4.7 million confirmed fraud cases across multiple banking platforms and payment ecosystems. Their findings revealed that conventional rulebased detection systems correctly identified only 63.8% of fraudulent transactions, with the remaining 36.2% of cases escaping detection entirely until customer-reported disputes were filed. The research team documented that these rulebased detection parameters remained unchanged for an average of 73 days between updates across the surveyed institutions, creating extended windows of vulnerability during which sophisticated fraud rings could operate with minimal detection risk. When rule modifications were eventually implemented, Anchoori et al. found that each adjustment required an average of 14.2 engineering hours to design, test, and deploy, creating substantial operational overhead for financial institutions already struggling with resource constraints. The researchers further noted that 67.3% of these rule adjustments were reactive responses to identified fraud patterns rather than proactive defenses against emerging threats [2]. Scalability limitations represent a critical technical barrier for batch-oriented systems operating in today's high-velocity transaction environment. The systematic performance testing conducted by Anchoori et al. across six major banking platforms documented that traditional fraud detection architectures experienced significant system degradation once transaction volumes exceeded 12,300 transactions per second, with processing times increasing exponentially beyond this threshold. Their technical benchmarking revealed that batch processing frameworks required an average of 46.7 minutes of additional processing time for each 10% increase in transaction volume during peak periods—a limitation that becomes particularly problematic during seasonal shopping events and promotional periods. The research team observed particularly severe performance degradation during high-transaction events such as Black Friday, when fraud attempts simultaneously increased by 317% while system responsiveness decreased by 43.8%, creating a perfect storm of vulnerability precisely when protection was most needed. Anchoori's analysis further demonstrated that 78.3% of financial institutions relying on batch processing were forced to implement emergency capacity expansions at least four times annually to manage transaction spikes, incurring average additional infrastructure costs of $3.7 million per institution annually [2]. Global Journal of Engineering and Technology Advances, 2025, 23(01), 275-289 278 The compliance and regulatory implications of delayed detection represent perhaps the most concerning aspect of traditional batch processing for fraud detection. Anchoori and colleagues analyzed regulatory filings and enforcement actions across North American, European, and Asia-Pacific markets, documenting those financial institutions utilizing batch processing frameworks faced 2.8 times higher regulatory penalties related to compliance failures compared to those implementing real-time detection capabilities. Their detailed analysis of 34 enforcement cases revealed average penalties of $13.7 million per incident, with regulatory authorities specifically citing detection latency as a contributing factor in 76.2% of these cases. The research further established that institutions relying on batch processing encountered significant challenges in meeting reporting requirements, with 41.3% of required suspicious activity reports filed after mandated deadlines due to inherent detection delays, creating additional regulatory exposure. Particularly concerning was the finding that delayed detection significantly hampered data completeness in regulatory filings, with 57.8% of batch-processed reports lacking critical transaction context that had been lost due to processing gaps, further compromising compliance quality [2]. Anchoori's research team conducted comprehensive technical assessments comparing batch processing with emerging real-time architectures, documenting stark performance differences across multiple operational dimensions. Their controlled experiments revealed that batch systems required an average of 326.7 GB of additional storage per million transactions processed compared to streaming alternatives, creating substantial infrastructure overhead. Furthermore, the research established that batch systems generated 3.7 times more false positives during high-volume processing periods, creating significant operational burdens as fraud analysts were required to review thousands of legitimate transactions flagged incorrectly. Perhaps most importantly, the researchers documented that traditional batch architectures demonstrated virtually no capability for detecting coordinated, multi-channel fraud attacks that spanned different transaction types—a critical weakness exploited by sophisticated fraud rings that deliberately structure activities to avoid detection. Anchoori et al. concluded that financial institutions relying primarily on batch processing were operating with a fundamental technological disadvantage in the increasingly sophisticated fraud landscape, where millisecond response capabilities have become essential for effective defense [2]. Table 2 Key Limitations of Batch ETL Systems in Modern Fraud Detection: Quantitative Analysis [2] Performance Metric Traditional Batch Processing Real-Time Processing Difference (%) Detection Latency 17.3 hours Near real-time ~100% reduction Additional Fraudulent Transactions 8.2 per compromised account Minimal ~100% reduction Increased Fraud Losses 72.6% higher Baseline 72.60% Correct Fraud Identification 63.80% Near 100% 36.2% gap Rule Update Frequency 73 days average Continuous ~100% improvement Rule Modification Effort 14.2 engineering hours Automated ~100% reduction Transaction Volume Threshold 12,300 TPS Significantly higher Not specified Processing Time Increase 46.7 min per 10% volume increase Minimal ~100% reduction False Positive Rate 3.7x higher Baseline 370% difference Regulatory Penalties 2.8x higher Baseline 280% difference Average Penalty Amount $13.7 million per incident Significantly lower Not specified Storage Requirements 326.7 GB more per million transactions Baseline Significant reduction 3. Delta Live Tables: A Game-Changer for Real-Time Processing Delta Live Tables represents a paradigm shift in how financial institutions approach streaming data processing. As a declarative framework built on Delta Lake, DLT simplifies the development and maintenance of real-time data pipelines while providing enterprise-grade reliability. According to comprehensive research published by Polimetla, financial Global Journal of Engineering and Technology Advances, 2025, 23(01), 275-289 279 institutions implementing DLT have experienced transformative improvements in their data engineering efficiency and fraud detection capabilities. His detailed analysis of 19 major banks and financial services organizations documented average development time reductions of 78.3% when implementing real-time fraud detection pipelines using DLT compared to traditional custom-built streaming architectures. The research further revealed that these institutions achieved a 91.7% reduction in pipeline failures after transitioning to DLT, alongside an impressive 86.4% decrease in engineering maintenance hours required to sustain these mission-critical systems. Polimetla emphasizes that this combination of reduced development time and improved operational reliability creates a compelling business case for DLT adoption in fraud detection use cases, where both time-to-market and system reliability directly impact financial outcomes [3]. The continuous data ingestion capabilities of Delta Live Tables provide a robust foundation for real-time fraud detection in financial services environments. Polimetla's technical benchmarking across various financial transaction ecosystems documented that DLT implementations successfully processed an average of 27,500 transactions per second with 99.997% reliability, even during periods of extreme volume fluctuation that typically challenge traditional data processing architectures. His detailed assessment revealed that DLT's declarative SQL and Python APIs reduced code complexity by 67.8% compared to custom streaming implementations, making these pipelines more maintainable and allowing financial institutions to focus engineering resources on fraud detection logic rather than infrastructure maintenance. Particularly significant for fraud detection applications, Polimetla documented that DLT successfully maintained sub-200ms end-to-end processing latency even when simultaneously ingesting data from 23 different transaction sources, ranging from traditional card payments to emerging cryptocurrency exchanges. This multi-source processing capability proved especially valuable for detecting sophisticated fraud patterns that deliberately span multiple channels to evade detection, with financial institutions reporting a 43.7% increase in cross-channel fraud identification after implementing DLT-based solutions [3]. The incremental processing capabilities of DLT deliver particular advantages for fraud detection use cases where computational efficiency directly impacts detection speed and cost-effectiveness. Polimetla's research highlighted that traditional batch-oriented fraud detection systems typically reprocess entire transaction datasets during each analytical cycle, creating substantial computational overhead and introducing unnecessary latency. His comparative analysis revealed that DLT's sophisticated change data capture and incremental processing mechanisms reduced computational requirements by 87.6% while simultaneously improving data freshness by 94.3% compared to traditional approaches. Through detailed performance testing across multiple cloud platforms, Polimetla demonstrated that DLT-powered systems could process 1.2 million transactions per minute using just 22.7% of the computational resources required by traditional batch systems handling equivalent workloads. This efficiency translated directly to operational savings, with the studied institutions reporting an average 68.3% reduction in cloud infrastructure costs for their fraud detection platforms after DLT implementation. Polimetla noted that these cost savings often justified the entire investment in DLT migration within the first 9-14 months of operation, providing both immediate operational benefits and long-term financial advantages [3]. Schema enforcement and evolution capabilities represent another crucial advantage of DLT for fraud detection applications that must adapt to constantly changing transaction formats and data structures. Polimetla's comprehensive assessment documented that financial institutions typically manage 187.3 schema changes annually across their transaction processing systems, with each change previously requiring an average of 12.4 hours of engineering work to propagate through analytical pipelines. His detailed study of DLT implementations revealed that automatic schema evolution handling reduced this engineering overhead by 93.7%, allowing fraud detection models to adapt to changing data structures without pipeline disruptions that could create detection blind spots. Additionally, Polimetla found that DLT's robust schema enforcement eliminated 97.2% of data quality issues that previously required manual intervention, resulting in 42.8% more reliable model performance and 38.9% fewer false positives in fraud classification. The research highlighted that this improved data quality was particularly valuable for machine learningbased fraud detection, with models trained on DLT-processed data demonstrating 27.4% higher precision and 31.8% higher recall compared to identical models trained on data processed through traditional pipelines with their inherent quality variations [3]. The built-in data governance features of DLT deliver particular value for financial institutions operating in highly regulated environments where audit capabilities directly impact compliance costs and risk exposure. Polimetla conducted extensive compliance assessments across implementations in 7 different regulatory jurisdictions, finding that DLT's automatic versioning, lineage tracking, and comprehensive audit logging capabilities reduced compliance validation efforts by 76.3% compared to traditional data architectures requiring separate governance tooling. His detailed analysis of regulatory reporting workflows revealed that institutions implementing DLT reduced audit preparation time from an average of 37.2 person-days to just 8.9 person-days per quarterly review cycle, creating Global Journal of Engineering and Technology Advances, 2025, 23(01), 275-289 280 substantial operational savings while improving compliance quality. Most significantly, Polimetla documented that DLTs' built-in governance capabilities helped financial institutions achieve 99.7% audit compliance rates without requiring specialized tooling or dedicated compliance engineering resources, representing a critical operational advantage in increasingly stringent regulatory environments. Financial institutions reported particular value in DLT's ability to provide point-in-time reproducibility of fraud detection results, with compliance teams able to reconstruct the exact state of data and detection logic for any historical transaction—a capability that proved invaluable during regulatory examinations and fraud investigations [3]. Figure 1 DLT Implementation Impact on Performance Metrics [3] 4. Machine Learning Models for Fraud Detection in DLT Systems When integrated with Distributed Ledger Technology (DLT), various machine learning approaches create a powerful multi-layered defense system against fraudulent activities. This integration combines the immutability and transparency of blockchain with the predictive capabilities of AI, resulting in more robust protection mechanisms. 4.1. Anomaly Detection 4.1.1. Isolation Forest Isolation Forest algorithms have demonstrated exceptional performance in identifying fraudulent transactions with minimal false positives. According to Waspada et al. [4], their comprehensive analysis of credit card transaction data revealed that Isolation Forest achieved an accuracy of 97.69% with an F1-score of 0.82 in fraud detection scenarios. Their experiments utilized a dataset containing 284,807 transactions, of which only 0.172% (492 cases) represented fraudulent activities, highlighting the algorithm's effectiveness in handling highly imbalanced datasets. The researchers found that Isolation Forest significantly outperformed traditional methods like Local Outlier Factor (LOF) and One-Class SVM in this domain, with LOF achieving only 0.64 F1-score and One-Class SVM reaching 0.72. Furthermore, their implementation demonstrated remarkable efficiency, processing 10,000 transactions in just 3.2 seconds, making it suitable for real-time applications. A particularly noteworthy finding was that Isolation Forest maintained robust performance even when the contamination rate (percentage of anomalies in the training data) varied between 0.1% and 1%, showing adaptability to different fraud prevalence scenarios. The algorithm's natural ability to handle highdimensional data without requiring extensive feature engineering made it especially valuable for financial transaction analysis, where the researchers utilized 28 principal components derived from the original transaction features. 4.1.2. Autoencoders Deep learning-based autoencoders have revolutionized normal transaction pattern learning in DLT environments. According to Sooriyarachchi and Gangichetty [5], implementing autoencoders within a Delta Live Tables (DLT) pipeline enables near real-time anomaly detection with remarkable efficiency. Their system architecture, designed for a major financial services company, processed transaction data through a streaming pipeline that scored events within Global Journal of Engineering and Technology Advances, 2025, 23(01), 275-289 281 milliseconds of their occurrence. The authors implemented a sophisticated autoencoder model with encoding layers of [64, 32, 16, 8] neurons and corresponding decoding layers, trained on over 12 months of historical transaction data with approximately 1.2 billion records. The model was deployed using Databricks Machine Learning, achieving a processing throughput of over 200,000 transactions per minute with an average latency of just 178 milliseconds. Their novel approach combined both reconstruction error and prediction error metrics, establishing dynamic threshold boundaries based on statistical properties of the error distribution. This dual-metric approach reduced false positives by 67% compared to single-metric methods while maintaining a 98.2% detection rate for known fraud patterns. A key innovation in their implementation was the continuous model monitoring system that tracked concept drift using population stability index (PSI) metrics, automatically triggering retraining when PSI exceeded 0.2, thus ensuring the model remained effective against evolving fraud tactics. The pipeline's integration with alerting systems enabled their security team to investigate suspicious transactions within an average of 4.3 minutes from occurrence, significantly reducing potential financial losses. 4.2. Classification Models 4.2.1. Random Forest and XGBoost Ensemble methods have become standard in the industry due to their robust performance across diverse transaction types. Ashraf et al. [6] conducted extensive research on fraudulent cryptocurrency transaction prediction using ensemble deep learning approaches. Their study analyzed 23,678 Bitcoin transactions, including 2,916 fraudulent cases identified through known scam addresses. The researchers implemented multiple ensemble strategies, with a stacked ensemble combining XGBoost, Random Forest, and gradient boosting achieving the highest performance with 95.73% accuracy and an impressive AUC (Area Under Curve) of 0.983. Their feature engineering process extracted 21 key transaction attributes, including temporal patterns, network characteristics, and value-based metrics, with feature importance analysis revealing that transaction velocity, temporal spacing, and network centrality provided the strongest signals for fraud detection. The Random Forest model within their ensemble demonstrated particular strength in handling the non-linear relationships between features, achieving 93.92% accuracy independently, while XGBoost excelled at capturing subtle interactions between features with 94.65% accuracy. Performance benchmarking showed that their ensemble approach reduced false positives by 37.2% compared to the best single-model approach, translating to approximately 58 fewer false alerts per 1,000 transactions. The researchers also noted that computational efficiency varied significantly, with Random Forest completing inference in 0.86 seconds for 1,000 transactions compared to 1.26 seconds for XGBoost on identical hardware, an important consideration for high-throughput blockchain systems processing millions of daily transactions. 4.2.2. Neural Networks Deep neural networks excel at identifying complex non-linear patterns in transaction data, particularly when temporal and spatial dimensions are involved. Kamisetty et al. [7] developed an artificial intelligence-based strategy for fraud detection in Bitcoin transactions that leveraged deep learning architectures to identify suspicious patterns. Their research utilized a dataset comprising 2.9 million Bitcoin transactions, including approximately 30,000 labeled fraudulent cases gathered from established blacklists and scam addresses. The team implemented a multi-layer neural network with three hidden layers (256-128-64 neurons) using ReLU activation functions and dropout regularization (rate=0.3) to prevent overfitting. This architecture achieved 94.2% accuracy and 91.8% recall on the test dataset, significantly outperforming logistic regression baselines that reached only 83.6% accuracy. A key innovation in their approach was the incorporation of graph-based features derived from transaction networks, including measures of address clustering, transaction flow patterns, and temporal behavior across the blockchain. These network-derived features improved model performance by 8.7% compared to models using only transaction metadata. The researchers noted that their model was particularly effective at detecting specific fraud typologies, achieving 97.3% detection rates for Ponzi schemes while maintaining 93.1% accuracy for money laundering patterns. Implementation challenges included the computational intensity of feature extraction, which required processing the entire Bitcoin blockchain and took approximately 76 hours on their research infrastructure. However, once deployed, the model could evaluate new transactions in near real-time, with an average processing time of 246 milliseconds per transaction, making it suitable for integration with blockchain validation systems. 4.3. Network Analysis Graph-based algorithms provide critical insights by examining relationships between different entities in the transaction ecosystem. In their research on credit card fraud detection, Waspada et al. [4] extended their analysis to include network-based approaches, creating transaction graphs where nodes represented customer accounts and merchants while edges captured payment flows. Their implementation of a customized PageRank algorithm assigned risk scores based on proximity to known fraudulent nodes, achieving an 86.5% success rate in identifying previously Global Journal of Engineering and Technology Advances, 2025, 23(01), 275-289 282 undetected fraud rings. The researchers discovered that accounts within three transaction hops of confirmed fraudulent entities were 22 times more likely to be involved in suspicious activities themselves. Their graph construction process incorporated 14 months of transaction history, resulting in a network with over 3.2 million nodes and 78 million edges that revealed distinct community structures corresponding to legitimate versus fraudulent behavioral patterns. By combining Isolation Forest anomaly scores with graph centrality metrics, the researchers created a hybrid scoring mechanism that improved precision by 7.9% compared to either approach used independently, demonstrating the complementary nature of statistical and network-based fraud detection methods. Entity embedding techniques transform entities into vector spaces where proximity indicates relationships. Sooriyarachchi and Gangichetty [5] implemented entity embeddings as part of their Delta Live Tables pipeline, creating 64-dimensional vector representations for customers, merchants, and devices based on their transaction histories. Their approach processed over 230,000 unique entities, generating embeddings that captured behavioral similarities even when no direct connections existed between accounts. The technique proved particularly effective in identifying synthetic identity fraud cases, where perpetrators carefully constructed seemingly unrelated accounts that nevertheless displayed subtle behavioral patterns when projected into the embedding space. The researchers utilized cosine similarity metrics to establish relationship scores, with similarity thresholds above 0.85 triggering enhanced verification measures. Their implementation of t-SNE dimensionality reduction for visualization enabled analysts to identify suspicious clusters, leading to the discovery of 14 previously unknown fraud rings operating across their customer base. The embedding model was recomputed weekly using a sliding 90-day window of transaction data, requiring approximately 5.3 hours of processing time on their distributed computing infrastructure but providing crucial adaptability to evolving criminal strategies and customer behavior patterns. 4.4. Sequence Models Temporal pattern analysis through LSTMs and Transformer-based models captures suspicious transaction sequences that might appear normal when viewed individually. Kamisetty et al. [7] extended their deep learning approach with a specialized LSTM network that processed chronological sequences of Bitcoin transactions to identify temporal anomalies. Their sequence model utilized a bidirectional LSTM architecture with 128 memory units followed by attention mechanisms that highlighted suspicious transactions within user histories. This approach achieved remarkable results in detecting account takeovers, with 92.6% accuracy in identifying compromised wallets based solely on behavioral changes over time. The researchers found that sequence length significantly impacted performance, with optimal results achieved using 30-50 transaction histories per address, balancing pattern recognition capabilities against computational requirements. Their model successfully identified subtle precursors to larger fraudulent activities, such as test transactions and graduations in transaction size, with an average early detection window of 3.9 days before major fraud attempts. Implementation in a production environment processed transaction sequences in 433 milliseconds on average, enabling near real-time risk assessment as transactions were submitted to the blockchain network. Figure 2 Comparative Analysis of Machine Learning Models Across Different Fraud Detection Scenarios [4,5,6,7] Global Journal of Engineering and Technology Advances, 2025, 23(01), 275-289 283 Transformer-based models have shown superior performance for longer sequence analysis. Ashraf et al. [6] incorporated transformer architectures in their ensemble approach, utilizing a 4-layer transformer with 6 attention heads that maintained context across extended transaction histories. When evaluated against sequences containing up to 100 transactions per address, their transformer model achieved 96.2% accuracy in identifying sophisticated fraud operations that established legitimacy over time before executing fraudulent transactions. The model's self-attention mechanism effectively distinguished between routine transactions and significant pattern deviations, maintaining a false positive rate of just 3.8% even when analyzing dormant accounts that suddenly became active. The researchers noted that their transformer implementation required substantial computational resources during training, necessitating 37 hours on a system with eight NVIDIA V100 GPUs, but inference speed remained practical at 267 milliseconds per transaction sequence. A particularly valuable finding was the model's effectiveness in detecting crosschain fraud attempts where perpetrators moved assets between different cryptocurrencies in attempts to obfuscate their activities, with the transformer achieving 89.7% detection accuracy for these complex scenarios compared to 76.3% for traditional classification approaches. 5. Implementing a Real-Time Fraud Detection Pipeline with DLT Recent research by Abbassi et al. provides a comprehensive framework for implementing real-time fraud detection pipelines using Delta Live Tables (DLT). Their extensive study, which analyzed 14 financial institutions across North America, Europe, and Asia that successfully deployed such architectures, offers both detailed technical insights and quantifiable metrics demonstrating the effectiveness of this approach for combating financial fraud in the digital age [8]. 5.1. Data Ingestion Layer The data ingestion layer serves as the critical foundation for real-time fraud detection capabilities, responsible for capturing high-velocity transaction data from diverse sources with minimal latency. Abbassi's research team documented that effective implementations simultaneously process transaction data from an average of 13.7 distinct channels, with the most sophisticated systems integrating up to 26 different data sources, including mobile banking applications, online payment gateways, ATM networks, and third-party fraud intelligence feeds. Their detailed technical architecture evaluation revealed that optimized DLT pipelines achieved sustained ingestion rates of 37,500 events per second with 99.998% reliability, maintaining this performance even during 300% transaction volume spikes that typically occur during peak shopping periods. Most impressively, these systems delivered end-to-end latency averaging just 84 milliseconds from event creation to availability for analysis, with 99th percentile latency not exceeding 157 milliseconds. The researchers found that this massive performance improvement—representing a 99.3% reduction in ingestion latency compared to traditional batch systems that averaged 12.7 minutes—directly translated to fraud mitigation outcomes. Financial institutions participating in the study reported that every 100ms reduction in ingestion latency corresponded to approximately $1.87 million in prevented annual fraud losses due to faster intervention capabilities, with one global bank documenting fraud losses decreasing by 42.3% within six months of deploying their optimized ingestion layer [8]. 5.2. Feature Engineering Layer The feature engineering layer represents the most computationally intensive component of fraud detection pipelines, responsible for transforming raw transaction data into meaningful risk indicators that enable accurate fraud detection. Abbassi et al. identified that high-performing systems calculate an average of 217 distinct risk indicators per transaction, spanning multiple risk domains to create a comprehensive risk profile. Their detailed analysis showed that transaction velocity features, which tracked patterns like transaction frequency, amount variations, and merchant category transitions, contributed most significantly to detection accuracy (31.7% of model importance across analyzed implementations), with sudden changes in transaction intervals providing particularly strong fraud signals. Geospatial anomalies that identified impossible travel patterns—such as transactions in different countries within impossibly short timeframes—accounted for 27.3% of overall importance according to the researchers' feature contribution analysis. Their study found that these geospatial models typically maintained databases of 2.7 billion location points to accurately assess travel feasibility between transaction locations. Behavioral biometrics, including keystroke dynamics, mouse movement patterns, and mobile device handling characteristics, accounted for 21.8% of detection capability, with the researchers noting that these signals were particularly valuable for detecting account takeover scenarios where legitimate credentials were being used by unauthorized individuals. Network features examining connections to known fraud cases provided the remaining 18.2% of signal strength, with relationship graphs typically containing between 17 million and 43 million nodes in the studied implementations. The researchers documented that these sophisticated feature calculations occurred with average latencies of just 137 milliseconds in optimized DLT implementations,