Adaptive Fraud Detection Under Concept Drift: Risks and Challenges of AI-Based Financial Fraud Detection

Abstract

This study has examined how concept drift affects the performance and reliability of artificial intelligence-based financial fraud detection systems, and it has developed and tested a framework for building resilient, adaptive fraud detection in evolving fraud environments. The central problem has been that fraud is a non-stationary, adversarial phenomenon in which the statistical relationship between transaction features and fraudulent behavior changes over time, a condition known as concept drift, so that a model trained on historical data degrades silently as fraud patterns evolve, and static artificial-intelligence models deployed without adaptation therefore lose accuracy, miss novel fraud, and expose institutions to rising loss and compliance risk. Guided by five research questions, concerning how concept drift affects detection performance, which drift types most affect accuracy, which adaptive techniques best mitigate drift, what risks static models carry, and how adaptive governance, Explainable AI, and continuous monitoring improve resilience, the study has proposed a framework comprising five contributing constructs: drift detection and dynamic adaptation; adaptive machine-learning technique maturity; Explainable AI and transparency; governance, oversight and continuous monitoring; and data quality and feature engineering, with detection resilience as the outcome. A quantitative, cross-sectional design supplemented by a prototype drift-simulation evaluation has been used, and data have been collected from fraud-analytics engineers, data scientists, model-risk and compliance officers, fraud investigators, and financial-technology architects. Out of 284 distributed questionnaires, 242 valid responses have been retained, producing an 85.2% valid response rate. The analysis has included descriptive statistics, reliability testing, correlation analysis, multiple regression, hypothesis testing, and a prototype evaluation of detection-performance decay and recovery under simulated drift. The findings have shown that all five constructs were significantly and positively associated with detection resilience, that the regression model was significant, F(5, 236) = 52.41, p < .001, explaining 52.6% of the variance, and that drift detection and dynamic adaptation and adaptive machine-learning technique maturity were the strongest predictors. The prototype evaluation showed that static models suffered the largest accuracy losses under adversarial and sudden drift, while adaptive ensembles coupled with drift detectors recovered the most performance. The study concludes that reliable AI-based fraud detection requires treating fraud as a non-stationary adversarial process and building drift-aware adaptation, explainability, and continuous governance into the system, rather than deploying static models that decay unseen.