<![CDATA[This study aims to evaluate the robustness of Convolutional Neural Networks (CNN) in face recognition systems under varying illumination conditions. The evaluation was conducted using a dataset comprising 36 subjects, with facial images captured under three distinct lighting scenarios: dim, normal, and bright. The research methodology involved training the CNN model using K-Fold Cross-Validation and assessing its stability against visual disturbances using artificial adversarial attacks based on the Fast Gradient Sign Method (FGSM). The novelty and main contribution of this study lie in the dual-evaluation approach, which simultaneously tests the model's resilience against natural illumination variations and artificial adversarial perturbations. Experimental results demonstrated that the CNN model achieved optimal face recognition performance at 50 epochs, maintaining an average accuracy rate of 81.48%. In conclusion, the evaluated CNN architecture is reliable and stable for face recognition in uncontrolled lighting environments, providing a solid foundation for developing more secure biometric systems against visual disturbances.]]>
Copyrights © 2026
