Opacity in deep convolutional networks within face recognition systems poses significant fairness, transparency, and accountability challenges, especially with ethnicity-altered synthetic datasets. This study advances explainable artificial intelligence in biometrics by assessing the interpretability of face recognition models on the CycleGAN-generated ethnicity-modified image dataset. Techniques such as xCos, Explainable Face Recognition (with Excitation Backpropagation), Local Interpretable Model-agnostic Explanations, On Black Box Explanation (Average method), and True Black Box Explanation (MinPlus) generate saliency maps to identify critical facial regions influencing recognition across demographic groups. The analysis exposes pronounced biases in decision-making, highlighting the struggles of deep learning architectures with various facial attributes. Explainability frameworks localize feature dependencies (e.g., eyes, nose, mouth), offering pathways to bias mitigation. This research represents the first in-depth exploration of explainable face verification on ethnicity-altered synthetic images, laying the groundwork for equitable and transparent biometric systems.