Abstract

This study establishes an unsupervised machine learning-based anomaly detection framework aimed at enhancing cybersecurity in Nigerian fintech startups, where swift digital growth, substantial transaction volumes, and a critical shortage of labeled fraud data hinder the efficacy of conventional rule-based and supervised detection methods. The study is driven by the necessity for adaptive and scalable security mechanisms that can detect emerging and novel fraud patterns while maintaining computational efficiency and operational feasibility in resource-limited fintech settings. The study uses real transaction data from a Nigerian fintech platform to model normal transactional behavior and find anomalies as deviations from this baseline. It does this through a structured process that includes data preprocessing, feature normalization, and behavioral analysis. Unsupervised methods, such as clustering-based methods, Isolation Forest, and reconstruction-based models (autoencoders and PCA) were used. Comparisons were done based on ROC-AUC, anomaly scores, false-positive rates, and how well they worked. The results show that models based on isolation and reconstruction always do better than simple clustering methods. Isolation Forest got a ROC-AUC of about 0.62 on validation data and kept its performance stable on the test set. Reconstruction-based models, on the other hand, showed better stability and lower false-positive rates. Overall, the results show that unsupervised anomaly detection is a useful and effective way to improve cybersecurity in fintech. It allows for early detection of suspicious activities without needing labeled data and provides a scalable framework that supports fraud prevention, regulatory compliance, and safe digital financial operations in the Nigerian fintech ecosystem.

Statistics