This study proposes a robust ox face recognition system based on the Swin Transformer (Tiny) architecture combined with ArcFace loss to generate highly discriminative 512-dimensional embeddings. As a contribution a custom dataset of 427 oxen with 3,843 training and 2,562 validation images is created and augmented to reflect real-world scenarios. The proposed model achieved 98.44% closed-set recognition accuracy which out performs ViT-Small and ResNet-50, with precision and ROC-AUC of 98.57% and 99.79%, respectively. Additionally, this study also supports both closed-set identification and open-set verification using cosine similarity thresholds, addressing the challenges related to scalability, dataset scarcity, and cross-pose variation thus offering a practical solution for automated ox identification and secure livestock management.