The widespread use of facial masks poses a major challenge for face recognition systems, often leading to drastic performance degradation. This paper presents a systematic, multi-stage framework to build a highly robust face recognition model against such occlusions. We perform extensive ablation experiments by gradually introducing three main enhancements over a strong Triplet Loss baseline: (1) Mask-Aware Sampling to explicitly learn cross-mask invariances; (2) a Spatial Attention module to adaptively focus on un-occluded facial regions; and (3) the state-of-the-art ArcFace loss to maximize the embedding’s discrimination power. Extensive experiments demonstrate that our final model not only achieves an outstanding F1-score exceeding 97% across all verification scenarios (unmasked-unmasked, masked-masked, and cross-mask) but, more critically, exhibits exceptional robustness, reducing the performance variance between the easiest and hardest scenarios from 5.9% in the baseline to a mere 2.0%. Furthermore, we evaluate our models also on the Labeled Faces in the Wild (LFW) which is a standard face benchmark and our final model achieved an efficient improvement of +8.9% accuracy improvement over the baseline in the general case of LFW, demonstrating that our model acts effectively in general when it may not be necessarily occlusive situation as well.