Face presentation attack detection (PAD) is critical for securing face recognition systems against presentation attacks. Recent advances have explored synthetic data as a promising alternative to mitigate privacy concerns and the limited availability of real data. However, models trained solely on synthetic data often suffer from performance degradation due to distribution shifts when applied to real domains. In this work, we propose an entropy-regularized multi-level feature alignment framework that leverages labeled synthetic data and unlabeled real samples for domain-adaptive training. By encouraging feature consistency across domains and promoting confident predictions on real inputs, the proposed method effectively bridges the synthetic-to-real gap. Extensive experiments on public benchmarks demonstrate superior generalization performance compared to existing approaches, highlighting its potential for practical deployment in real-world face PAD systems.