Sistema de reconocimiento facial para control de acceso con Deep Learning en la Cooperativa de Ahorro y Crédito Sol en el área de cajas/bóveda.

Abstract

Access control to the cash register and vault areas of Cooperativa de Ahorro y Crédito SOL presented security vulnerabilities due to the reliance on manual logs, which generated risks of identity impersonation, recording errors, and lack of traceability. Therefore, the main objective of this research was to develop an automated access validation application based on facial recognition using convolutional neural networks. The study applied a field and applied methodology with a mixed approach, using surveys, interviews, and observation sheets addressed to a sample of 12 employees in order to diagnose the initial situation. As a result, the proposed application mitigated four high-severity risks and fulfilled seven operational requirements. Furthermore, when comparing two technological models, the transfer learning program implemented with the DeepFace library showed better comparative performance than the local supervised learning classifier, since it used 100 reference photographs per authorized employee, achieving an accuracy of 98 percent, a precision of 96.15 percent, and a sensitivity of 100 percent. In contrast, the local model required 300 photographs and obtained 55 percent accuracy with 89 false negatives. Thus, it was concluded that the implementation of neural networks reduced dependence on physical records and automated biometric verification in real time, providing a more secure, traceable, and auditable access management process for the financial institution.