AI-Powered Anti-Money Laundering (AML) and fraud detection - enhancing financial security through intelligent fraud detection

 Corresponding author: Osarense Dorothy Iguodala Copyright © 2025 Author(s) retain the copyright of this article. This article is published under the terms of the Creative Commons Attribution Liscense 4.0. AI-Powered Anti-Money Laundering (AML) and fraud detection - enhancing financial security through intelligent fraud detection Osarense Dorothy Iguodala 1, * and Aghogho Oyiborhoro 2 1 Department of Accounting, University of Lagos. 2 KPMG United States. World Journal of Advanced Research and Reviews, 2025, 26(02), 3702–3714 Publication history: Received on 17 January 2025; revised on 24 February 2025; accepted on 27 February 2025 Article DOI: https://doi.org/10.30574/wjarr.2025.26.2.0637 Abstract The increasing sophistication of financial crimes, including money laundering and fraud, necessitates advanced technological solutions to enhance financial security. Artificial Intelligence (AI)-powered Anti-Money Laundering (AML) and fraud detection systems have emerged as transformative tools in the financial sector, enabling proactive threat identification and risk mitigation. This study explores the integration of AI techniques—such as machine learning (ML), deep learning, and natural language processing (NLP)—in detecting fraudulent activities and identifying suspicious transactions in real time. AI-driven AML frameworks leverage predictive analytics and anomaly detection models to enhance compliance with regulatory frameworks while reducing false positives. This research highlights key AI-based methodologies in fraud detection, including supervised and unsupervised learning models, neural networks, and reinforcement learning. Moreover, it examines the role of explainable AI (XAI) in improving transparency and trust in financial security operations. The integration of AI with blockchain technology is also discussed, showcasing its potential to enhance transaction traceability and prevent illicit activities. Despite its advantages, AI-driven AML systems face challenges, including data privacy concerns, adversarial attacks, and regulatory compliance issues. This study emphasizes the need for a balanced approach that combines AI innovation with ethical and legal considerations. By leveraging AI-powered AML and fraud detection, financial institutions can significantly improve their ability to combat financial crime, ensuring a more secure and resilient global financial ecosystem. Keywords: AI-driven AML; Fraud detection; Machine learning; Financial security; Predictive analytics; Regulatory compliance 1. Introduction Financial crimes, particularly money laundering and fraud, have become increasingly complex due to the rapid evolution of digital transactions and global financial networks. The widespread adoption of online banking, digital payment systems, and cryptocurrency exchanges has inadvertently created opportunities for illicit financial activities, challenging traditional AntiMoney Laundering (AML) and fraud detection frameworks. Regulatory agencies and financial institutions are continuously striving to strengthen their risk assessment methodologies and compliance mechanisms to combat financial crimes effectively. However, conventional rule-based approaches often fall short in identifying sophisticated fraud patterns, leading to inefficiencies such as high false positive rates and increased operational costs. In response to these limitations, Artificial Intelligence (AI) has emerged as a transformative force in financial security, offering advanced machine learning (ML) and deep learning techniques to detect, prevent, and mitigate money laundering risks and fraudulent activities [1]. AI-powered AML systems leverage predictive analytics, anomaly detection, and network analysis to enhance the identification of suspicious transactions in real time. Unlike traditional methods, AI-driven models can adapt to evolving fraud tactics by continuously learning from vast and World Journal of Advanced Research and Reviews, 2025, 26(02), 3702–3714 3703 diverse financial data. Supervised and unsupervised learning techniques allow these systems to classify fraudulent behaviors based on historical data while also detecting emerging threats without prior knowledge of illicit activities. Deep learning architectures, including convolutional neural networks (CNNs) and recurrent neural networks (RNNs), further enhance fraud detection by processing complex financial patterns, transactional sequences, and user behaviors. This study presents an in-depth analysis of AI-powered fraud detection and AML strategies, focusing on their efficacy, challenges, and integration within the financial sector [2]. By systematically reviewing existing literature and conducting empirical analyses, this research evaluates the impact of AI-driven AML models on financial security. It explores how financial institutions deploy advanced AI algorithms to detect fraudulent activities, comply with regulatory standards, and minimize the risks associated with money laundering schemes. Moreover, this study examines the role of Explainable AI (XAI) in addressing the transparency and interpretability concerns surrounding AIbased decision-making in AML processes. Since regulatory compliance is a crucial aspect of AML implementation, this paper also discusses the evolving legal frameworks and governance structures that shape AI adoption in financial crime prevention. Furthermore, the integration of AI with blockchain technology presents new opportunities for enhancing financial security. Blockchain’s decentralized and immutable ledger system can complement AI-powered AML strategies by improving transaction traceability and reducing the risk of data manipulation. Figure 1 Concept of fraud Detection system By leveraging smart contracts and cryptographic techniques, AI-enhanced blockchain systems can strengthen compliance protocols and ensure accountability in financial transactions. However, despite these advancements, AIdriven AML solutions face significant challenges, including adversarial attacks, data privacy concerns, and the need for extensive computational resources [3], [4]. This paper addresses these challenges and proposes potential solutions, emphasizing the importance of ethical AI deployment in financial security applications. By analyzing the scientific and technical dimensions of AI-powered AML and fraud detection, this research contributes to the broader discourse on financial crime prevention in the digital age. It underscores the need for a multidisciplinary approach that combines AI innovation, regulatory compliance, and ethical considerations to develop robust financial security mechanisms. Ultimately, this study aims to provide valuable insights into the future of AI-driven AML solutions, offering recommendations for policymakers, financial institutions, and technology developers to enhance the effectiveness of fraud detection and financial crime mitigation strategies. Despite the advancements in AI-powered AML and fraud detection, several challenges must be addressed to ensure their effective deployment. Adversarial attacks on AI models, in which malicious actors manipulate data inputs to evade detection, pose a significant threat to the reliability of fraud detection systems. Additionally, data privacy regulations, such as the General Data Protection Regulation (GDPR) and the California Consumer Privacy Act (CCPA), impose strict limitations on the collection and processing of financial data, requiring AI-driven AML solutions to balance security with legal compliance. Furthermore, World Journal of Advanced Research and Reviews, 2025, 26(02), 3702–3714 3704 the high computational requirements of deep learning models present scalability concerns, necessitating the development of efficient and resource-conscious AI architectures. 2. Literature Review The application of Artificial Intelligence (AI) in Anti-Money Laundering (AML) and fraud detection has been extensively studied in recent years, with researchers exploring various machine learning (ML) and deep learning techniques to enhance financial security. Traditional rule-based AML systems have been widely criticized for their inability to adapt to evolving fraud patterns, leading to inefficiencies such as high false positive rates and excessive manual investigations (Ngai et al., 2011). As a result, AI-driven solutions have gained prominence for their ability to analyze vast amounts of financial data, detect anomalies, and identify suspicious transactions in real time [5]. In a seminal study, Balasubramanian et al. (2019) demonstrated that supervised learning models, such as decision trees and random forests, outperform traditional rule-based approaches in fraud detection, reducing false positives by 30% and improving overall detection accuracy. Similarly, Königstorfer & Thalmann (2020) highlighted the role of neural networks in enhancing AML processes, showcasing how deep learning models can detect complex money laundering schemes through sequential transaction analysis. One of the most significant advancements in AI-powered AML has been the use of unsupervised learning techniques, particularly clustering and anomaly detection models. Scharfstein et al. (2018) investigated the application of self-organizing maps (SOM) and autoencoders in detecting fraudulent transactions, revealing that these models can successfully identify previously unknown fraud patterns. Furthermore, Zang & Zhou (2021) compared the effectiveness of k-means clustering and density-based spatial clustering (DBSCAN) in AML, concluding that DBSCAN is more effective in detecting irregular transaction behaviors, particularly in highvolume financial networks. These findings align with the work of Gai et al. (2022), who demonstrated that a hybrid approach combining supervised and unsupervised learning significantly enhances fraud detection efficiency by leveraging historical fraud data while identifying new fraudulent patterns. Deep learning techniques, including Long Short-Term Memory (LSTM) networks and Transformer-based models, have also gained traction in AML research. Goodfellow et al. (2017) introduced the concept of adversarial training in deep learning, illustrating how fraudsters attempt to manipulate AI models to evade detection. Building on this, Zhou et al. (2020) proposed an LSTM-based approach to detect transaction anomalies, achieving a 92% accuracy rate in identifying suspicious activities. More recently, Rahman et al. (2023) examined the potential of Transformer-based architectures, such as BERT and GPT models, in financial fraud detection, finding that these models can effectively analyze sequential transaction data and extract complex fraud patterns with minimal human intervention. These studies emphasize the growing importance of deep learning in AML, particularly in analyzing temporal transaction behaviors and reducing false negatives [6]. Natural Language Processing (NLP) has also emerged as a critical tool in AI-driven AML frameworks, enabling financial institutions to extract valuable insights from unstructured data sources. Chen et al. (2019) explored the application of sentiment analysis and topic modeling in AML, demonstrating that NLP techniques can enhance Know Your Customer (KYC) processes by analyzing adverse media reports and suspicious activity narratives. Similarly, Zhang & Liu (2021) developed an NLP-based system for monitoring regulatory compliance, using entity recognition and sentiment analysis to detect financial crime risks in textual data. Their findings align with the work of Kroll et al. (2022), who emphasized the importance of AI-driven text analysis in identifying financial entities involved in money laundering activities. These studies highlight the potential of NLP in complementing traditional transaction monitoring systems, providing a more holistic approach to financial crime detection. Explainable AI (XAI) has become a growing area of interest in AML research, addressing concerns regarding the transparency and interpretability of AI-driven fraud detection systems. Rudin et al. (2018) argued that the opacity of deep learning models presents challenges for regulatory compliance, as financial institutions must provide justifications for AML decisions. In response, Ribeiro et al. (2020) developed an XAI framework that enhances the interpretability of fraud detection models using SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). Their research demonstrated that incorporating explainability techniques not only improves regulatory compliance but also enhances trust in AI-powered AML solutions. More recently, Liao et al. (2023) explored the role of XAI in improving auditor oversight, finding that interpretable AI models facilitate better collaboration between compliance teams and regulatory authorities. These findings underscore the necessity of balancing AI innovation with ethical and legal considerations to ensure responsible AI deployment in financial security [7]. World Journal of Advanced Research and Reviews, 2025, 26(02), 3702–3714 3705 Figure 2 Illustration of AI and Blockchain Integration Another emerging trend in AI-powered AML is the integration of blockchain technology to enhance transparency and traceability in financial transactions. Nakamoto (2008) originally introduced blockchain as a decentralized ledger system, and since then, researchers have explored its applications in AML. For instance, Feng et al. (2020) proposed a hybrid AI-blockchain framework for fraud detection, leveraging smart contracts to automate compliance processes and detect illicit transactions in real time [8]. Similarly, Liang & Hu (2021) examined the potential of blockchain analytics in monitoring cryptocurrency transactions, demonstrating that AI-driven network analysis can effectively trace suspicious financial flows in decentralized finance (DeFi) ecosystems. Despite these advancements, Böhme et al. (2022) cautioned that blockchain's pseudonymity poses challenges for AML enforcement, as criminals increasingly exploit privacyenhancing technologies to obscure their financial activities. These studies highlight both the opportunities and limitations of blockchain in AI-powered AML strategies, emphasizing the need for regulatory adaptations to address emerging risks. Despite the significant progress in AI-driven AML and fraud detection, several challenges remain. One of the primary concerns is the vulnerability of AI models to adversarial attacks, in which malicious actors manipulate transaction data to bypass detection mechanisms. Huang et al. (2019) investigated adversarial machine learning in financial security, demonstrating that fraudsters can exploit model weaknesses by injecting adversarial noise into transactional data. To mitigate these risks, He et al. (2022) proposed a robust AI framework that incorporates adversarial training and anomaly detection techniques, significantly improving fraud detection resilience. Additionally, data privacy regulations, such as the General Data Protection Regulation (GDPR) and the California Consumer Privacy Act (CCPA), impose strict limitations on the collection and processing of financial data (Sweeney et al., 2020). Researchers have proposed privacy-preserving AI techniques, such as federated learning and homomorphic encryption, to address these challenges (Yang et al., 2021). These findings indicate the need for a multidisciplinary approach that combines AI advancements with legal and ethical considerations to ensure effective AML implementation. In summary, the literature suggests that AI-powered AML and fraud detection systems have significantly enhanced financial security by leveraging ML, deep learning, NLP, and blockchain analytics [9], [10]. However, challenges related to transparency, adversarial attacks, and regulatory compliance remain critical areas for future research. The integration of XAI, privacypreserving AI techniques, and hybrid AI-blockchain solutions represents promising directions for improving AML effectiveness while maintaining ethical standards. This review underscores the importance of ongoing innovation and collaboration between financial institutions, regulatory agencies, and AI researchers to develop robust and responsible AML strategies in the digital era. 3. Methodology This study employs a multidisciplinary approach to investigate the efficacy of Artificial Intelligence (AI) in Anti-Money Laundering (AML) and fraud detection. The methodology is structured around a combination of quantitative data analysis, machine learning (ML) model evaluation, and qualitative assessment of regulatory compliance frameworks. A systematic approach is adopted to ensure the robustness and validity of findings, integrating empirical data, algorithmic World Journal of Advanced Research and Reviews, 2025, 26(02), 3702–3714 3706 performance metrics, and compliance considerations. The study follows a structured research design that includes data collection, preprocessing, model selection, evaluation criteria, and regulatory assessment to provide a holistic understanding of AI-powered AML mechanisms. 3.1. Data Collection and Preprocessing The dataset used in this study comprises transactional records from financial institutions, sourced from publicly available AML datasets, simulated financial transactions, and real-world suspicious activity reports (SARs). The primary dataset includes attributes such as transaction amount, frequency, sender and receiver details, geographic location, and timestamps. Additionally, adverse media reports, regulatory filings, and financial institution compliance reports are incorporated to enrich the dataset with unstructured text data for Natural Language Processing (NLP) analysis. The data is preprocessed to remove inconsistencies, normalize transaction amounts, and handle missing values using imputation techniques [11], [12]. Anomalous transactions are labeled based on historical fraud reports, enabling supervised learning model training. Data privacy measures are strictly adhered to, ensuring compliance with regulatory standards such as GDPR and the California Consumer Privacy Act (CCPA). 3.2. Machine Learning and Deep Learning Models A comparative analysis of various AI techniques is conducted to assess their effectiveness in fraud detection. Supervised learning models, including Logistic Regression (LR), Decision Trees (DT), Random Forest (RF), Gradient Boosting Machines (GBM), and Support Vector Machines (SVM), are deployed to classify transactions as fraudulent or nonfraudulent. Additionally, deep learning models, including Long Short-Term Memory (LSTM) networks and Transformerbased architectures, are implemented to analyze sequential transaction behaviors. Anomaly detection techniques such as Isolation Forests, Autoencoders, and Variational Autoencoders (VAEs) are used for unsupervised learning to identify emerging fraud patterns. For NLP-based AML analysis, pre-trained BERT and GPT models are fine-tuned to extract insights from textual data sources, such as SAR narratives and regulatory filings [13]. 3.3. Evaluation Metrics and Performance Assessment To ensure an objective assessment of AI models, multiple evaluation metrics are utilized, including Accuracy, Precision, Recall, F1-score, and Area Under the Receiver Operating Characteristic (AUROC) curve. Precision and Recall are particularly emphasized to minimize false positives and false negatives, as high false positive rates lead to excessive manual investigations, while false negatives pose risks of undetected financial crime. Additionally, explainability metrics such as Shapley Additive explanations (SHAP) values and Local Interpretable Model-agnostic Explanations (LIME) are employed to assess model transparency and regulatory compliance. Computational efficiency, measured in terms of training time and inference speed, is also considered to evaluate the feasibility of AI-driven AML systems for real-time deployment [14]. The study critically evaluates AI-powered AML models within the context of regulatory frameworks, including the Financial Action Task Force (FATF) recommendations, the European Union’s Fifth AntiMoney Laundering Directive (5AMLD), and the Bank Secrecy Act (BSA). Compliance risk is assessed by examining model interpretability, auditability, and adherence to financial crime regulations. Ethical considerations, including bias mitigation and fairness in AI decision-making, are analyzed using fairness-aware algorithms to detect and correct potential discriminatory outcomes in fraud detection. The integration of Explainable AI (XAI) is explored to enhance regulatory transparency and ensure that AI-generated risk assessments align with legal standards. 3.4. Experimental Setup and Validation The experiments are conducted in a controlled environment using Python-based machine learning frameworks, including Scikit-learn, TensorFlow, and PyTorch. The dataset is split into training (70%), validation (15%), and testing (15%) subsets to evaluate model generalization. Hyperparameter tuning is performed using Bayesian optimization to optimize model performance. To ensure robustness, k-fold cross-validation (k=10) is employed, reducing the risk of overfitting. The study also compares AI-powered AML models against traditional rule-based AML systems to quantify improvements in fraud detection efficiency. Furthermore, case studies of real-world AML implementations are reviewed to validate the practical applicability of AI-driven financial security strategies. While this study provides a comprehensive evaluation of AI-powered AML solutions, certain limitations must be acknowledged. First, the availability of high-quality labeled fraud data remains a challenge due to the confidentiality of financial transactions. To address this, synthetic data generation techniques, including Generative Adversarial Networks (GANs), are explored. Second, adversarial attacks on AI models remain an ongoing concern, requiring further research into robust adversarial defense mechanisms. Lastly, the evolving landscape of financial crime necessitates continuous model adaptation, highlighting the need for AI-driven AML systems that incorporate reinforcement learning for real-time threat detection [15]. Future research will focus on the integration of AI with blockchain analytics for enhanced financial crime monitoring, as well as the development of privacy-preserving AI techniques to comply with data protection regulations. World Journal of Advanced Research and Reviews, 2025, 26(02), 3702–3714 3707 3.5. Data Collection and Preprocessing The dataset used in this study comprises real-world and synthetic financial transaction records obtained from multiple sources, Suspicious Activity Reports (SARs) from regulatory bodies such as the Financial Crimes Enforcement Network (FinCEN). Kaggle, European AML databases, and the IEEE-CIS fraud dataset. Cryptocurrency transaction data extracted from Ethereum and Bitcoin public ledgers using GraphQL APIs. Generated using Gaussian Mixture Models (GMM) and Generative Adversarial Networks (GANs) to augment rare fraud instances. 3.6. Each transaction record contains attributes such as: Where Tid = Transaction ID, Uid= Unique User ID, Asrc,Adest= Sender and receiver accounts, Camount= Transaction amount, Fflag= Label (1 for fraudulent, 0 for non-fraudulent), Ttime= Timestamp of the transaction, Llocation= Transaction location, Rrisk= Risk score based on historical behavior 3.7. Data Preprocessing Missing values were imputed using mean values for continuous features and mode for categorical variables. Transaction frequency (TF) for each user was calculated: 3.8. Transaction deviation score (TD): Min-Max scaling was applied to ensure uniform feature ranges. The dataset was highly imbalanced (fraud cases ~2%). SMOTE (Synthetic Minority Over-sampling Technique) was used to generate additional fraudulent samples. 3.9. AI Model Development The study implements multiple ML models to classify transactions as fraudulent or non-fraudulent. 3.9.1. Logistic Regression (Baseline Model) Where Y=1Y=1Y=1 indicates fraud, Xi represents transaction features, and βi are model coefficients. 3.9.2. Random Forest (RF) A decision-tree ensemble model where the final prediction is obtained by: where T is the number of decision trees. 3.9.3. Gradient Boosting (XGBoost) An ensemble learning approach that minimizes the loss function using: Where ℓ is the loss function and λ is the regularization parameter. 3.10. Deep Learning Models Two deep learning models were employed: World Journal of Advanced Research and Reviews, 2025, 26(02), 3702–3714 3708 3.10.1. Long Short-Term Memory (LSTM) Designed for sequential transaction behavior analysis. The hidden state ht is updated as follows: where ft,it,otf, are forget, input, and output gates respectively. 3.10.2. Transformer-Based Model (BERT for NLP AML Detection) To analyze unstructured textual data from SARs, BERT embeddings were generated using: E=BERT(S) where S represents suspicious report narratives, and E is the transformed feature vector. 3.11. Performance Evaluation and Analysis Evaluation Metrics The models were assessed using: 3.11.1. Accuracy 3.11.2. Precision 3.11.3. Recall 3.11.4. F1-Score AUROC (Area Under the Receiver Operating Characteristic Curve) to measure discrimination power. 3.12. Experimental Setup and Model Comparison All experiments were conducted in a controlled environment using Python (Scikit-learn, TensorFlow, and PyTorch). The dataset was split into 70% training, 15% validation, and 15% testing. Bayesian Optimization was used for hyperparameter tuning. World Journal of Advanced Research and Reviews, 2025, 26(02), 3702–3714 3709 Figure 3 Results and observation The results indicate that deep learning models, particularly LSTM and BERT, outperform traditional ML models in detecting financial fraud as show in chart 1. The high recall values of deep learning models demonstrate their ability to minimize false negatives, which is critical in AML applications. 3.13. Regulatory Compliance and Ethical Analysis To ensure transparency, (SHAP) was applied, revealing that high transaction amounts and frequent international transfers were the most influential fraud indicators. Additionally, fairness-aware algorithms detected potential biases in models, requiring reweighting of certain demographic features to mitigate discriminatory patterns. The AI models were mapped to regulatory guidelines from the Financial Action Task Force (FATF) and EU AML Directives. The findings indicate that AI-powered systems align with AML requirements but require explainability enhancements to satisfy regulatory scrutiny. This study demonstrates the effectiveness of AI-powered AML systems, with deep learning models exhibiting superior fraud detection capabilities [16], [17]. However, challenges related to model explainability and adversarial robustness remain. Future work will explore federated learning for privacy-preserving AML detection and blockchain analytics for enhanced financial traceability. 3.14. Results and Analysis The results of this study are divided into multiple sections, including statistical analysis, model performance evaluation, feature importance ranking, and regulatory compliance validation [18]. The findings are presented with mathematical interpretations, tables, and complex formulas derived from financial fraud detection metrics. A preliminary statistical analysis was conducted on the transaction dataset to understand the distribution of fraudulent and non-fraudulent transactions. 3.15. Descriptive Statistics The dataset contained 3,500,000 transactions, out of which 72,400 transactions (2.07%) were labeled as fraudulent Table 1 Descriptive Statistics Feature Mean Median Standard Deviation Min Max Transaction Amount ($) 1,850.23 720.15 5,610.72 0.01 3,500,000 Transaction Frequency (per user) 12.4 7 23.1 1 200 Fraud Label (1=Fraud, 0=NonFraud) 0.0207 0 0.14 0 1 World Journal of Advanced Research and Reviews, 2025, 26(02), 3702–3714 3710 The distribution of transaction amounts was highly skewed, with fraudulent transactions having a higher median transaction value. The performance of various AI models was evaluated using multiple metrics, including Accuracy, Precision, Recall, F1-score, and AUROC (Area Under the Receiver Operating Characteristic Curve). Figure 4 Performance Metrics From the chart 2, it is evident that deep learning models outperform traditional machine learning models in fraud detection. The BERT-based NLP model achieved the highest AUROC score (0.97), indicating superior discrimination between fraudulent and legitimate transactions. 3.16. Mathematical Interpretation of Model Results To further analyze the model's efficiency, we derived key performance formulas and evaluated them against our dataset. Let P(F) be the probability of a transaction being fraudulent, and P(T) be the total number of transactions. The posterior probability that a transaction is fraudulent given observed features X is computed using Bayes' Theorem: Using dataset statistics: • Prior probability of fraud: P(F)=0.0207P(F) = 0.0207P(F)=0.0207 • Likelihood of fraud given high transaction amount: P(X∣F) =0.78 • Overall likelihood of high transaction amounts: P(X)=0.25P(X) = 0.25P(X)=0.25 Thus, This result indicates that a transaction with a high amount has a 6.45% probability of being fraudulent more than three times the baseline probability of fraud. 3.17. Feature Importance Analysis The contribution of different transaction features to fraud detection was analyzed using Shapley Additive Explanations (SHAP).