Liveness detection plays a crucial role in protecting facial recognition systems from spoofing attempts using printed images, video replays, or 3D masks. We propose a novel U-Net-based architecture that captures both local texture anomalies and global facial structures to distinguish live faces from spoofed ones. Unlike conventional CNN-based methods that often lose fine-grained spatial information, our model preserves critical cues through multi-scale feature extraction and skip connections. The encoder extracts high-level semantic information, whereas the decoder restores spatial details that are sensitive to spoofing cues like reflectance irregularities and edge distortions. Tests on public datasets demonstrate that our approach achieves strong performance and generalizes effectively to novel attack types and across different datasets. Additive angular margin softmax further boosts generalization. The architecture is lightweight and suitable for real-time deployment in consumer-grade facial authentication systems. The source code for this work is available at: https://github.com/UV72/deep-liveness-2025 .