fraud detection system design

Rising e-commerce fraud has overwhelmed manual investigation processes, causing delayed responses and alert fatigue. This paper introduces the first multimodal large language model (MLLM)-based automation framework tailored for credit card fraud investigations. Our approach innovatively integrates task planning (Chain-of-Thought + ReAct), dynamic code execution, OCR, and visual reasoning over transaction graphs to enable end-to-end autonomous execution of a standardized seven-step investigative workflow: alert parsing → evidence collection → cross-source reasoning → report generation. The framework ensures interpretability and strict adherence to financial regulatory requirements. Evaluated on 500 real-world cases, it fully completes all seven analytical steps on average, achieves >92% accuracy in critical conclusions, significantly improves investigation efficiency, and substantially reduces both false negatives and analyst cognitive load.

This study addresses the lack of empirical evidence on key operational metrics—such as latency, cost, fairness, and adversarial robustness—for deploying large language models (LLMs) in trust and safety workflows like fraud detection and content moderation. Through a systematic literature review of 49 operationally relevant studies, the authors propose the FORTE role framework and a minimal deployment evidence checklist encompassing dimensions like latency budgets and per-decision costs to structurally evaluate LLM roles and evidence completeness in real-world settings. The analysis reveals a structural imbalance in existing work: while content moderation exhibits relatively comprehensive operational evidence, fraud detection suffers from severe gaps. The paper concludes by outlining critical directions for empirical research necessary to support reliable LLM deployment in high-stakes applications.

To address the limitations of existing approaches in credit card fraud detection—namely, insufficient accuracy, poor interpretability, and inadequate real-time performance—this paper proposes an end-to-end unsupervised learning framework. Methodologically, it integrates multi-source financial data to construct temporal behavioral features; introduces a novel composite risk scoring mechanism that fuses anomaly scores from Isolation Forest, One-Class SVM, and a deep autoencoder, augmented with burst- and frequency-based spending indicators; and employs PCA-based visualization coupled with DBSCAN density clustering to enable interpretable localization of fraudulent transactions. Evaluated on real-world transaction data, the framework achieves precise identification of 1–2% high-risk transactions, significantly improving detection priority for high-risk cardholders and merchants. It supports real-time anti-fraud decision-making while maintaining both high detection efficacy and operational interpretability.

The technical opacity of AI systems enables novel, covert, and systemic fraud that eludes detection by traditional fraud theories. Method: This paper proposes the “AI Fraud Diamond Model,” extending the classic Fraud Triangle with a fourth element—“technical opacity”—and develops a taxonomy of AI fraud encompassing five categories, including data manipulation and model misuse. It further introduces a diagnostic auditing paradigm tailored for automated systems, shifting audit focus from outcome verification to systematic vulnerability identification. Contribution/Results: Grounded in qualitative interviews with auditors from major consulting firms and domain experts, the study validates the model’s explanatory power in uncovering auditors’ technical skill gaps, interdisciplinary collaboration barriers, and constraints on system access. The framework advances AI governance and intelligent auditing by offering a theoretically rigorous yet practically implementable analytical tool.

Current automated detection tools struggle to meet regulatory practice demands due to insufficient transparency, interpretability, and the inability to map findings to specific legal provisions, resulting in a disconnect between academic research and enforcement applications. Through in-depth interviews with nine regulatory practitioners and an analysis integrating regulatory workflows with technical feasibility, this study systematically uncovers, from a regulatory perspective, the practical barriers to deploying automated tools for identifying deceptive designs. The work proposes a human-in-the-loop compliance review framework that is user-need-driven, supports the entire investigative workflow, and aligns both research and regulatory objectives, offering critical guidance for developing automated detection systems that genuinely meet real-world enforcement requirements.

This study addresses the challenge of detecting coordinated multi-account fraud under high transaction volumes by proposing a hierarchical, auditable fraud detection pipeline that integrates gradient-boosted trees, graph-based structural features, autoencoder-derived anomaly signals, TreeSHAP explanations, and an LLM-driven investigative agent, complemented by a disagreement-based escalation review mechanism. It presents the first systematic evaluation of the synergistic effectiveness of graph features, anomaly detection, and LLM agents in realistic fraud scenarios. Experimental results demonstrate that graph features substantially enhance fraud ranking performance in the medium-risk segment and achieve 100% recall in fraud ring detection. Although the LLM agent generates plausible explanations, its decision accuracy falls below that of threshold-based methods; however, its errors are effectively flagged through the disagreement-based escalation mechanism.

This study evaluates the risk of large language models (LLMs) being misused in sophisticated cybercrimes such as romance scams, CEO impersonation, and identity theft. To this end, we introduce the first reproducible, multi-turn interactive evaluation framework co-designed with law enforcement and policy experts. The framework decomposes malicious intent into seemingly benign queries to assess models’ ability to generate actionable information, benchmarking against standard web search and open-source LLMs with safety safeguards removed. Our findings indicate that mainstream closed-source models offer limited assistance for high-level criminal activities; however, their risk increases substantially when safety constraints are disabled. Moreover, multi-turn indirect requests prove more effective than explicit malicious prompts at bypassing current defenses, exposing critical limitations in existing safety strategies under complex adversarial scenarios.

This study addresses the challenges of dynamically evolving financial fraud and severe class imbalance in digital payment systems by conducting hypothesis-driven exploratory data analysis and feature engineering on the PaySim synthetic dataset, following the CRISP-DM methodology. To mitigate class imbalance, SMOTE oversampling is employed, and hyperparameter optimization is performed via GridSearchCV across multiple classifiers, including logistic regression, decision trees, random forests, and XGBoost. The resulting fraud detection framework achieves significantly enhanced detection performance while maintaining high scalability and robustness, thereby offering FinTech systems an efficient and reliable solution for real-time fraud prevention.

This study addresses the challenge of reduced predictive reliability in credit card fraud detection caused by severe class imbalance. The authors propose an optimized modeling approach based on the Explainable Boosting Machine (EBM), which systematically integrates Taguchi experimental design to fine-tune both data preprocessing sequences and model hyperparameters. By leveraging feature importance analysis and modeling interaction effects, the method enhances performance without resorting to conventional sampling techniques that may introduce bias. Evaluated on a standard credit card transaction dataset, the proposed approach achieves a ROC-AUC of 0.983, outperforming the baseline EBM (0.975) as well as established models including logistic regression, random forest, XGBoost, and decision trees. The framework successfully balances high predictive accuracy with strong interpretability, maintaining transparency without compromising performance.

This study addresses the dual challenges banks face from signature-based fraud—such as card-not-present transactions and account takeovers—and behavioral financial crimes, including layering money laundering and business email compromise, which traditional rule-based engines struggle to detect effectively. To tackle this, the authors propose an AI-powered security agent for both retail and corporate accounts, featuring a novel three-component architecture that fuses transaction streams and session streams in parallel. The framework integrates LSTM-based temporal modeling, statistical threshold monitoring, and account relationship graph networks to enable joint detection of multi-vector fraud and anti-money laundering threats. Experimental results on synthetic data demonstrate F1 scores of 0.787 and 0.867 for transaction and session streams, respectively, significantly outperforming rule-based baselines and standalone LSTM models, while achieving 96.6% authentication accuracy with critical response latency under 0.43 milliseconds.