Fraud Detection

Fraud detection refers to the set of methods, systems, and processes used to identify deceptive, unauthorised, or criminal activities in financial transactions, digital platforms, insurance claims, and other commercial contexts. In modern practice, fraud detection systems are built primarily on machine learning (ML) and artificial intelligence (AI), replacing or augmenting earlier rule-based approaches with adaptive models that can identify novel fraud patterns and operate at the speed and scale required by digital payment systems. Financial fraud represents one of the most economically significant application domains for AI. Global losses to payment fraud alone exceeded US0 billion in 2023, and this figure does not include cybercrime, insurance fraud, and identity theft. As of 2025, 90 percent of global banks report using AI and ML for fraud prevention and detection, according to Feedzai's 2025 AI Trends in Fraud and Financial Crime report.

Types of Fraud

Fraud manifests in numerous forms across different sectors. Payment fraud encompasses card-not-present (CNP) fraud in online transactions, account takeover (ATO) attacks where fraudsters gain control of legitimate accounts, and authorised push payment (APP) fraud where victims are deceived into transferring funds to criminals. Identity fraud involves using stolen or synthetic identities to open accounts or obtain credit. In insurance, fraud includes inflated or fabricated claims, staged accidents, and provider billing fraud. In e-commerce, it encompasses return fraud, coupon abuse, and fraudulent seller activity. Money laundering, while distinct from transactional fraud, overlaps with fraud detection in the anti-money laundering (AML) domain, where suspicious transaction patterns must be flagged regardless of whether an individual transaction appears fraudulent in isolation.

Traditional Rule-Based Systems

The earliest automated fraud detection systems used manually constructed rules: if a transaction exceeds a certain amount, or is made from an unusual geographic location, flag it for review. Rule-based systems are interpretable and fast to deploy, but they suffer from critical limitations. Fraud patterns evolve as criminals adapt to known rules. False positive rates are high, generating large volumes of legitimate transaction alerts that frustrate customers and strain operations teams. Rules cannot capture the complex combinatorial patterns characteristic of sophisticated fraud.

Machine Learning Approaches

Modern fraud detection systems layer multiple ML techniques. Supervised classification models — including gradient boosted decision trees (GBDTs), random forests, and deep neural networks — are trained on labelled historical transaction data to predict fraud probability for each new transaction. Models such as XGBoost and LightGBM are widely used in production due to their accuracy, interpretability through feature importance scores, and resistance to overfitting. GBDT models have demonstrated fraud detection precision rates above 90 percent on mobile financial services transaction datasets. A core challenge is class imbalance: fraudulent transactions typically represent less than 0.1 percent of all transactions. Techniques including oversampling (SMOTE), undersampling, and cost-sensitive learning are used to ensure models do not trivially classify all transactions as legitimate. Unsupervised anomaly detection techniques identify transactions that deviate significantly from established patterns without requiring labelled fraud examples. This is particularly valuable for detecting novel fraud types not seen in historical data.

Graph-Based Methods

Many fraud patterns are best captured through the relationships between entities rather than the properties of individual transactions. Money mule networks, for instance, involve chains of accounts used to move and obfuscate funds. Graph neural networks (GNNs) and graph database analytics can identify suspicious structural patterns in transaction networks, flagging clusters of accounts with unusual connectivity. Financial institutions have integrated graph databases with ML models to improve detection of organised fraud rings that would be invisible to transaction-level models.

Large Language Models in Fraud

Emerging applications use large language models for fraud detection in unstructured data — for instance, analysing the narrative descriptions in insurance claims for inconsistencies, or processing customer service conversation transcripts to detect social engineering attacks.

Real-Time Scoring Infrastructure

Production fraud detection systems must evaluate transactions in milliseconds to avoid delaying payment authorisation. This requires a serving infrastructure that can invoke ML models, retrieve customer behaviour profiles, and apply business rules within a response time budget of 100 to 300 milliseconds. Feature stores — databases optimised for low-latency retrieval of pre-computed customer and account features — are a key architectural component of real-time fraud systems.

Explainability and Regulatory Requirements

In regulated financial services, fraud detection decisions that affect customers — such as blocking a transaction or freezing an account — must be explainable. Regulators in many jurisdictions require that automated decisions affecting consumers be accompanied by a human-understandable explanation. Explainability techniques including SHAP (SHapley Additive exPlanations) values and LIME (Local Interpretable Model-agnostic Explanations) are used to generate per-decision explanations from black-box ML models. Malaysia has experienced significant growth in online financial fraud, with losses reaching RM1.12 billion in the first half of 2025 alone, following a record RM1.58 billion in reported losses in 2024. This scale of losses has made AI-powered fraud detection a national priority for financial sector regulators and institutions. Bank Negara Malaysia (BNM) and Payments Network Malaysia (PayNet) are jointly developing an AI-based fraud detection system slated for rollout in 2026. The system will leverage machine learning, large language models, and predictive analytics to identify unusual or high-risk activities in real time, such as sudden large transfers to previously unknown accounts or transactions matching known money mule account profiles. When the system detects a threat, it will alert the customer and force a pause before the transaction is processed. PayNet is also developing a federated AI model called FinancialGPT, trained entirely on Malaysian financial transaction data, to strengthen fraud detection across the national payment network. The federated architecture allows participating banks to contribute to model training without sharing raw customer data, addressing privacy and data sovereignty concerns. Major Malaysian banks have already deployed their own AI fraud systems. Maybank, CIMB, and RHB have significantly expanded AI-driven fraud monitoring, and Malaysian banks collectively reported blocking RM399 million in fraud attempts in 2024 through enhanced security measures including AI. These systems monitor transaction patterns, device fingerprints, and behavioural biometrics to flag suspicious activity. BNM's Risk Management in Technology (RMiT) policy document, updated in 2023, sets requirements for financial institutions deploying AI and ML in risk management and fraud detection, including model governance, explainability, and ongoing performance monitoring. Compliance with RMiT requires banks to maintain documented model inventories and conduct regular model validation for all material AI models used in financial crime detection.

See Also

1. Feedzai. (2025). AI Trends in Fraud and Financial Crime Report 2025. Feedzai Research.
2. West, J., and Bhanu, B. (2024). Fraud Detection in Financial Services with Graph Neural Networks. Journal of Machine Learning Research, 25(1).
3. Bank Negara Malaysia. (2023). Risk Management in Technology (RMiT) Policy Document. Kuala Lumpur: BNM.
4. PayNet. (2025). National Payments Fraud Report H1 2025. Payments Network Malaysia Sdn Bhd.
5. Lundberg, S.M., and Lee, S.I. (2017). A Unified Approach to Interpreting Model Predictions. Advances in Neural Information Processing Systems 30 (NeurIPS 2017).