MACHINE LEARNING FOR THREAT DETECTION: ENHANCING CYBERSECURITY IN FINANCIAL INSTITUTIONS

Abstract

Financial institutions are increasingly targeted by cyber threats, necessitating advanced threat detection mechanisms beyond traditional rule-based and statistical anomaly detection systems. Machine learning (ML) offers a scalable, adaptive, and high-precision approach to cybersecurity, effectively identifying and mitigating evolving cyber risks. In the study, this paper presents an ML-based hybrid cyber security framework coupled with supervised learning, anomaly detection, and adversarial ML to improve financial security. Several models such as Deep Neural Networks (DNNs), Random Forest, Autoencoders, and Support Vector Machines (SVMs) were used to evaluate the framework. Accuracy, adversarial robustness, inference speed, and real-time detection efficiency were used for assessing the models. The highest accuracy (96.3%) is made by DNNs which is higher than Random Forest (93.1%), Autoencoders (92.4%), and traditional models as Logistic Regression (87.4%), and SVM (90.2%). Adversarial robustness testing was then performed; they tested whether an improved adversarial accuracy is reflected in a relative improvement in adversarial robustness, they found DNNs retained 84.7% accuracy under perturbation attacks, whereas SVM and Logistic Regression dropped below 75%. Real-time detection analysis showed that Random Forest gave the best tradeoff between accuracy and inference time (6.2ms) and was thus suitable for real-time applications. The results show that hybrid ML approaches greatly increase cybersecurity in financial institutions, being robust, adaptive, and precise. Future research should explore lightweight deep learning architectures, explainable AI (XAI), and federated learning to improve scalability and data privacy.