REGISTRO DOI: 10.69849/revistaft/ra10202509071822
Daniela Cristina Abreu Jové de Araújo
Abstract
Artificial intelligence (AI) and machine learning (ML) are reshaping fraud detection by enabling continuous auditing, predictive modeling, and real-time anomaly identification. Unlike traditional auditing methods limited by manual sampling, AI-based systems process vast data streams and uncover subtle irregularities across financial and operational records. Practical applications include credit-card fraud detection, procurement monitoring, and journal-entry analysis, where supervised and unsupervised models—augmented by graph-based approaches—detect anomalies and adapt to adversarial behaviors. While these techniques enhance efficiency and detection accuracy, they also raise significant ethical and governance concerns, particularly regarding algorithmic bias, transparency, privacy, and regulatory compliance. A balanced approach combining technological advances with ethical safeguards, model interpretability, and robust governance structures is essential to realize the benefits of AI in fraud detection. This integration promises to shift auditing from episodic assurance toward continuous, adaptive systems that strengthen organizational resilience and public trust.
Keywords: Artificial intelligence; Machine learning; Fraud detection; Continuous auditing; Predictive models; Anomaly detection; Algorithmic bias; Ethics; Governance.
Artificial intelligence (AI) and machine learning (ML) are transforming the detection of financial and operational fraud by enabling continuous, data-driven auditing that scales far beyond the capacities of manual inspection. Where traditional forensic techniques relied on sampling and rule-based flags, modern ML pipelines ingest transactional streams, event logs, text records and relational metadata to generate predictive risk scores and surface anomalies in near real time. Advances in deep learning, graph algorithms and streaming analytics permit models to learn complex temporal and relational patterns of legitimate behavior and thus distinguish subtle deviations that often precede or accompany fraudulent schemes. These capabilities reduce detection latency, broaden the scope of observable signals, and allow auditors to prioritize investigations by estimated risk rather than by random or periodic sample alone (Cook, Misirli & Fan, 2022; Hilal, Gadsden & Yawney, 2022).
In practice, continuous auditing systems combine feature engineering, supervised learning, unsupervised anomaly detection, and graph-based methods. Supervised classifiers such as tree ensembles and gradient-boosted machines remain workhorses when labeled fraud histories exist; they offer strong performance on imbalanced transactional datasets and are interpretable via feature importance methods. Unsupervised approaches—autoencoders, isolation forests, and clustering—are especially useful when labeled fraud examples are scarce or when new types of misuse emerge. Graph neural networks and relational techniques can model actor–entity relationships (e.g., payees, accounts, IP addresses) and detect coordinated behaviors that evade transaction-level thresholds. In streaming settings, lightweight online models and incremental learning methods update risk estimates continuously, enabling immediate triage and automated responses while preserving audit trails for subsequent human review (Hilal, Gadsden & Yawney, 2022; Zhang et al., 2018).
One practical manifestation of these techniques is in credit-card and payments fraud detection, where low-latency classification must decide whether to authorize, decline, or route a transaction for additional authentication. Architectures here often pair engineered features (amount, merchant category, geolocation patterns, device fingerprint) with sequence models (LSTM, transformer variants) and graph-based signal fusion; production systems further embed feedback loops so that confirmed fraud labels retrain or recalibrate models, reducing false positives and adapting to adversary behavior. In audit contexts beyond payments—expense reimbursements, procurement, payroll, regulatory reporting—continuous monitoring platforms expose anomalies such as duplicate invoices, abnormal vendor relationships, or suspicious journal-entry patterns, allowing audit teams to move from reactive forensics to proactive risk management (Cook, Misirli & Fan, 2022; Zhang et al., 2018).
Despite technical promise, deploying AI/ML for fraud detection raises significant ethical and governance issues that auditors, firms, and regulators must address. Algorithmic bias may produce disparate impacts if training data encode historical inequities or measurement errors—leading to disproportionate scrutiny of certain customers, vendors, or demographic groups. The “black box” nature of many high-performing models complicates explanations demanded by internal governance or external regulation and can erode trust among stakeholders. Data privacy and confidentiality constraints intersect with detection objectives: intrusive feature sets may heighten detection power but create compliance risks under privacy laws and contractual obligations. Moreover, adversaries may attempt to manipulate models via poisoning attacks, data obfuscation, or behavior designed to mimic benign patterns, which calls for robust model validation, monitoring for concept drift, and adversarial testing as part of the audit lifecycle (Raji & Buolamwini, 2019; Hilal, Gadsden & Yawney, 2022).
To manage these trade-offs, organizations should adopt a layered approach that integrates technical, procedural and legal controls. Technically, model cards, counterfactual explanations and interpretable surrogate models can provide auditors with readable rationales for high-risk alerts; uncertainty quantification and calibrated risk thresholds reduce overconfidence in model outputs. Procedurally, continuous auditing requires well-defined escalation pathways, human-in-the-loop decision points for critical actions, documented model governance, and periodic independent audits of both data pipelines and model behavior. Legally and ethically, organizations should map features against privacy and fairness constraints, perform pre-deployment impact assessments, and maintain transparent communication with affected parties when automated decisions materially affect them. Cross-disciplinary teams—combining data scientists, auditors, compliance officers and ethicists—are essential to translate model performance into trustworthy operational practice (Raji & Buolamwini, 2019; Cook, Misirli & Fan, 2022).
The flowchart illustrates the end-to-end process of AI and ML-driven fraud detection in financial and operational contexts. It begins with diverse data inputs—transactional streams, event logs, text records, and relational metadata—that feed into the ML pipeline, where techniques such as feature engineering, supervised and unsupervised learning, anomaly detection, and graph-based methods are applied. The outputs generate risk scoring and anomaly detection, enabling predictive assessments of fraudulent activities. These insights support continuous auditing, allowing real-time monitoring while highlighting challenges like algorithmic bias and data privacy. Finally, the process integrates ethical and governance considerations, ensuring that detection systems remain transparent, fair, and trustworthy.
Figure 1. AI and ML-Driven Continuous Auditing for Fraud Detection.
Source: Created by author.
Looking forward, research trends point to several promising directions: hybrid models that combine symbolic rules with neural representations to improve interpretability and rule-consistency; continued development of graph-based and contrastive learning methods to detect coordinated, low-signal fraud; and standardized benchmarks and audit processes for model fairness and robustness. Regulatory momentum toward required algorithmic impact assessments and independent audits will likely shape how continuous auditing systems are designed and disclosed. As adversaries become more sophisticated, defenders must invest not only in more powerful detection algorithms but in resilient architectures that include monitoring for model degradation, mechanisms for privacy-preserving analytics, and readiness to incorporate third-party audits. When implemented with rigorous governance and ethical safeguards, AI and ML can shift auditing from episodic assurance to an ongoing, adaptive defense that protects value, reduces losses, and strengthens public trust in financial systems (Hilal, Gadsden & Yawney, 2022).
References
Cook, A. A., Misirli, G., & Fan, Z. (2022). Anomaly detection for IoT time–series data: A survey. IEEE Internet of Things Journal, 9(10), 7403–7421.
Hilal, W., Gadsden, S. A., & Yawney, J. (2022). Financial fraud: a review of anomaly detection techniques and recent advances. Expert Systems with Applications, 193, 116429.
Raji, I. D., & Buolamwini, J. (2019). Actionable auditing: Investigating the impact of publicly naming biased performance results of commercial AI products. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 429–435.
Zhang, Y., Zhang, J., Qin, L., Zhang, W., & Lin, X. (2018). Efficiently finding top-k shortest paths in large graphs. Proceedings of the ACM SIGMOD International Conference on Management of Data, 457–472.