Predicting multi-factor authentication uptake using machine learning and the UTAUT framework Ronald Kato [1] , Aggrey Obbo [1] , Richard Kimera* [1]

https://www.academia.edu/academia-ai-and-applications/2/1/10.20935/AcadAI8107 This study investigates a machine learning-driven framework for predicting multi-factor authentication (MFA) adoption in Uganda’s financial services ecosystem, a context increasingly exposed to cybersecurity risks as digital finance expands. This research aims to (i) identify key behavioral, technological and contextual determinants influencing MFA uptake, (ii) develop and validate an interpretable predictive model aligned with the Unified Theory of Acceptance and Use of Technology (UTAUT) and (iii) compare multiple classification algorithms, including classical ensembles and custom neural architectures, to establish an optimal approach for low-resource settings. Using the nationally representative FinScope Uganda 2023 dataset (survey responses = 3176), we engineered features from UTAUT constructs, security behavior indicators and digital access patterns. A binarized proxy for MFA adoption was derived from validated, high-loading security perception items. Methodologically, we implemented an experimental pipeline involving stratified train–test splitting, SMOTE applied within each cross-validation fold to avoid data leakage and repeated (n = 30) experiments to ensure stability of estimates. Six predictive models, Logistic Regression, Random Forest, Gradient Boosting and XGBoost alongside custom-built Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) architectures, were trained and optimized. Gradient Boosting achieved the strongest performance (mean accuracy = 0.838; F1-score = 0.835; AUC-ROC = 0.928), outperforming both linear baselines and complex neural models, which struggled with recall and F1-scores on tabular survey data. Timing analysis showed that Gradient Boosting balanced computational efficiency with predictive accuracy, making it suitable for low-bandwidth, resource-constrained environments. SHAP-based interpretability revealed that trust in digital security, prior exposure to mobile services, perceived effort and peer influence were the most influential drivers of MFA adoption. The findings advance current knowledge by integrating UTAUT constructs with explainable AI, strengthening behavioral prediction models in sub-Saharan Africa, where empirical MFA studies remain limited. This study contributes a reproducible, theory-grounded modeling pipeline, detailed comparative analysis between classical and neural network approaches and evidence-based policy recommendations.