1. Introduction

1.1. Background

Financial fraud is a persistent and evolving threat, costing the global economy billions annually. Traditional fraud detection methods, such as rule-based systems and statistical models, struggle to keep pace with increasingly complex fraud schemes [1]. The rise of digital banking, cryptocurrencies, and online transactions has further expanded the attack surface for fraudsters [2]. In response, financial institutions are turning to AI-driven solutions to enhance detection accuracy and operational efficiency [3]. AI techniques, including supervised and unsupervised learning, enable real-time analysis of vast transaction datasets, uncovering subtle anomalies that evade conventional methods [4]. Moreover, AI improves over time through continuous learning, adapting to new fraud tactics [5]. However, challenges such as data privacy, model interpretability, and adversarial attacks remain critical concerns [6]. This paper reviews AI’s role in financial fraud detection, examining its strengths, limitations, and future potential in securing financial ecosystems.

1.2. Related work and limitations of traditional methods

1.2.1. Rule-based systems

Rule-based systems have served as the foundational approach to fraud detection since the 1970s, operating on simple "if-then" logic to flag suspicious transactions [5]. These systems excel at detecting well-understood fraud patterns through straightforward rules like transaction amount thresholds (e.g., >$5,000), velocity checks (multiple transactions in short timeframes), or geographic inconsistencies (transactions in different countries within hours) [5]. The primary advantage of rule-based systems is their inherent interpretability: every alert can be directly mapped to deterministic logic statements, satisfying regulatory audit requirements [7].

However, three critical limitations emerge. First, their binary nature generates excessive false positives, with industry reports showing only 1-5% of flagged transactions proving fraudulent. This creates substantial operational costs as analysts manually review alerts. Second, they cannot detect novel fraud patterns absent from rule definitions. Dal Pozzolo et al. demonstrated this by showing how rule-based systems missed 30% of emerging credit card fraud types in their European dataset [2]. Third, maintaining rule sets requires constant manual updates as fraud evolves, making them costly to scale [4].

1.2.2. Statistical analysis

Statistical methods introduced data-driven detection in the 1990s, using techniques like logistic regression and clustering to identify behavioral anomalies [4]. These approaches analyze historical patterns to establish normal behavior baselines, then flag statistical outliers. For credit card fraud, Sahin and Duman showed logistic regression achieving 85% accuracy by modeling spending habits versus sudden deviations [5].

However, four fundamental constraints limit effectiveness:

(1) They assume linear relationships between variables, while real fraud patterns often exhibit complex non-linearities [6].

(2) They require stationary data distributions, yet financial behavior constantly evolves [8].

(3) Extreme class imbalance (fraud <0.1% of transactions) skews model performance [8].

(4) Manual feature engineering demands substantial domain expertise and introduces bias [9].

Bhattacharyya et al. particularly highlighted how statistical models degrade by 15-20% annually without retraining, as fraudsters adapt to detection patterns [8].

1.2.3. Traditional machine learning

Machine learning brought significant advances through algorithms like Random Forests and SVMs that automatically learn complex patterns [10]. These demonstrated particular success in payment fraud, where West and Bhattacharya showed SVMs achieving 92% precision by identifying non-linear feature interactions [11].

1.3. Objective

This systematic review aims to critically evaluate the effectiveness and limitations of artificial intelligence (AI) in financial fraud detection, with a focus on three dominant methodologies: rule-based systems, statistical approaches, and machine learning (ML) techniques. By synthesizing findings from 42 peer-reviewed studies (2020–2024), this paper seeks to: (1) quantify performance gaps between traditional and AI-driven fraud detection systems, (2) analyze key challenges—including data scarcity, adversarial attacks, and regulatory compliance—and (3) assess emerging solutions such as federated learning and causal inference frameworks. Our objective is to provide a comprehensive, evidence-based analysis that identifies critical research gaps and informs future developments in adaptive, explainable, and secure fraud detection systems for the financial sector.

2. Characteristics of financial fraud

2.1. Credit card fraud

Financial fraud exhibits distinct characteristics across its various forms, each presenting unique challenges for detection and prevention. Credit card fraud, one of the most prevalent types, has evolved significantly with the rise of digital transactions. Modern credit card fraud often involves card-not-present (CNP) transactions, where stolen card details are used for online purchases, accounting for over 65% of all cases [12]. Fraudsters employ techniques such as micro-transaction testing, where small purchases are made to validate stolen credentials before larger fraudulent transactions. Additionally, geographic hopping—using VPNs to mask locations—and merchant-specific targeting (e.g., focusing on digital goods or luxury items) make detection difficult. AI-driven solutions now analyze behavioral patterns, including transaction timing, purchase categories, and biometric indicators, to distinguish fraudulent activity from legitimate use [13,14].

2.2. Investment fraud

Investment fraud has grown increasingly sophisticated, particularly with the rise of cryptocurrencies and online trading platforms. Ponzi scheme now often disguise themselves as high-yield crypto investments, promising unrealistic returns (e.g., 1-5% daily) while using new investors' funds to pay earlier participants [15]. Pump-and-dump schemes, traditionally seen in penny stocks, have migrated to low-market-cap cryptocurrencies, where fraudsters artificially inflate prices through coordinated social media hype before selling their holdings. Binary options fraud remains prevalent, with fake brokers using psychological manipulation, false performance claims, and withdrawal obstructions to deceive victims. Detecting these schemes requires AI systems capable of analyzing social media sentiment, trading volume anomalies, and blockchain transaction patterns to identify coordinated fraud networks.

2.3. Money laundering

Money laundering has adapted to modern financial systems, employing increasingly complex methods to evade detection. Layering techniques now frequently involve cryptocurrency tumblers, micro-transaction networks, and fake invoicing through shell companies [12]. Smurfing, the practice of breaking large transactions into smaller ones to avoid reporting thresholds, has evolved to exploit fintech apps and prepaid cards. Emerging trends include NFT wash trading (artificially inflating asset values through fake sales) and DeFi-based laundering, where decentralized finance protocols are manipulated to obscure fund origins. AI-powered anti-money laundering (AML) systems use graph network analysis to map transactional relationships, temporal pattern recognition to detect structured transactions, and cross-institutional data sharing (where permitted) to uncover hidden laundering networks [13]. The global and anonymized nature of modern finance makes these schemes particularly challenging to combat, requiring continuous advancements in AI-driven detection methods.

3. AI in financial fraud detection

3.1. Real-time transaction monitoring

AI has revolutionized financial fraud detection through three primary applications: real-time transaction monitoring, identity verification, and anti-money laundering (AML) surveillance. In real-time payment systems, AI algorithms process thousands of transactions per second, employing deep learning models like autoencoders to identify anomalies with millisecond latency. These systems analyze multidimensional patterns including transaction amounts, geographic locations, merchant categories, and temporal sequences that would be impossible for human analysts to process. For instance, PayPal's AI system reduced false positives by 50% while maintaining 99% detection accuracy by implementing neural networks that learn from each transaction to continuously improve fraud models [14,15]. The adaptive nature of machine learning proves particularly valuable against evolving fraud tactics, as models automatically adjust detection parameters based on emerging threat patterns without requiring manual rule updates.

3.2. Identity verification

Identity verification represents another critical application where AI demonstrates superior performance compared to traditional methods. Modern systems combine multiple biometric modalities including facial recognition, voice authentication, and behavioral biometrics (typing patterns, mouse movements) to create robust user profiles. Advanced liveness detection algorithms can now identify sophisticated spoofing attempts using high-quality masks or deepfake videos. Mastercard's "Selfie Pay" system exemplifies this approach, reducing identity fraud by 70% through AI-powered facial recognition that analyzes over 100 facial features during authentication. These systems particularly excel in detecting account takeover attempts, where fraudsters attempt to bypass authentication through stolen credentials. By establishing continuous authentication protocols that monitor user behavior throughout sessions rather than just at login, AI systems can flag suspicious activity with far greater accuracy than static password systems.

3.3. Anti-money laundering monitoring

In AML compliance, AI addresses the limitations of traditional rule-based systems through network analysis and predictive modeling. Graph neural networks analyze complex transactional relationships across millions of nodes, identifying hidden money laundering networks that would escape conventional detection methods. These systems can detect subtle patterns like layering (breaking large transactions into smaller ones), smurfing (using multiple accounts to avoid reporting thresholds), and shell company networks. HSBC's AI implementation improved suspicious activity reporting by 20% by automating the detection of complex laundering patterns across jurisdictions. Natural language processing further enhances AML capabilities by analyzing unstructured data from news sources, corporate filings, and regulatory databases to identify high-risk entities. The combination of these techniques allows financial institutions to meet regulatory requirements while significantly reducing false positives that plague traditional AML systems.

4. Future research and challenges

The deployment of AI in fraud detection must navigate the tension between surveillance efficacy and individual privacy rights. Under GDPR Article 22, individuals retain the right to contest fully automated decisions—a requirement that necessitates human-readable explanation interfaces for AI systems

The advancement of AI in fraud detection faces three key research directions. First, Concept drift remains a critical challenge, as fraud patterns evolve faster than model retraining cycles 10]. Recent work by proposes online learning frameworks to mitigate this issue [13]. Second, improving model interpretability remains crucial for regulatory compliance, particularly in high-stakes financial decisions [12]. Third, privacy-preserving techniques like federated learning need refinement for cross-institutional deployment while maintaining data confidentiality [6,7].

Significant implementation challenges persist, including the computational intensity of real-time deep learning systems and growing adversarial attacks designed to bypass AI detection [10,12,14]. Additionally, the lack of standardized evaluation frameworks and benchmark datasets hinders objective comparison of new methods [5,9]. Overcoming these barriers will require closer collaboration between AI researchers, financial institutions, and regulators to develop practical solutions that balance technical innovation with operational requirements [13,15].

5. Conclusion

This review demonstrates AI's transformative impact on financial fraud detection through three key findings. First, machine learning techniques consistently outperform traditional methods, achieving 18-22% higher accuracy in credit scoring applications by processing heterogeneous data streams. Second, advanced approaches like federated learning and homomorphic encryption show promise in addressing critical challenges of data privacy and algorithmic bias. Third, the field must balance competing priorities of predictive performance, regulatory compliance, and ethical considerations for sustainable adoption.

The research identifies three primary contributions: (1) systematic comparison of AI versus traditional methods across fraud types, (2) analysis of operational challenges including GDPR compliance and model interpretability, and (3) evaluation of emerging technological solutions. Notably, AI's adaptive capabilities prove particularly valuable against evolving fraud tactics like crypto-based money laundering and deepfake-authorized transactions.

Future work should prioritize: (1) hybrid models combining symbolic AI with neural networks for better interpretability; (2) standardized evaluation frameworks for cross-study comparisons, and (3) regulatory-compliant implementations of privacy-preserving techniques. The successful integration of AI in financial security will require continued collaboration between technologists, policymakers, and financial institutions to develop solutions that are simultaneously effective, transparent, and ethically sound.