Аннотация
This paper addresses a critical challenge in face anti-spoofing by developing a unified deep neural network architecture that simultaneously performs genuine/fake face classification and depth estimation. The purpose of this study is to overcome the limitations of conventional methods that treat these tasks in isolation—often leading to suboptimal exploitation of complementary geometric and semantic features. The problem centers on the need for a cohesive framework that can jointly capture high-level semantic cues for spoof detection and low-level depth information to improve reliability under diverse testing conditions. The novelty of our approach lies in the integration of a state-of-the-art ConvNeXt backbone with a custom U-Net–style decoder and a dedicated classification head, which together enable multi-scale feature extraction and joint modeling of both tasks. The network is trained end-to-end using a joint loss function that combines cross-entropy loss for classification with mean squared error loss for depth estimation. Experiments on academic datasets such as OULU-NPU and CASIA across multiple protocols reveal that our method not only achieves competitive performance relative to state-of-the-art techniques, but also effectively sets the connection between geometric and semantic facial attributes.