Recent advances in palmprint recognition models using deep learning have shown promise, but performance often degrades significantly when applied to unseen target domains or new datasets. To address this, we propose a palmprint recognition method based on unsupervised domain adaptation that aims to minimize the performance drop across different datasets. Our approach aligns features, followed by a competitive network to enhance feature extraction and model learning, and is further optimized through a tailored loss function. Extensive experiments show that our method maintains strong generalization capabilities on unseen target domains, outperforming existing models in terms of accuracy and robustness across multiple datasets.