The paper addresses the challenge of unified physical-digital face attack detection, which has been a significant gap in existing research. The authors propose a dataset called UniAttackData, which combines physical and digital attacks, ensuring ID consistency for each subject. This dataset includes 1,800 participants with 2 physical and 12 digital attacks each, resulting in a total of 29,706 videos. To address the computational burden and the difficulty in learning a compact feature space for both types of attacks, the authors develop UniAttackDetection, a unified attack detection framework based on Vision-Language Models (VLMs). The framework consists of three main modules: the Teacher-Student Prompts (TSP) module, the Unified Knowledge Mining (UKM) module, and the Sample-Level Prompt Interaction (SLPI) module. These modules work together to learn a robust and unified feature space for detecting both physical and digital attacks. Extensive experiments on UniAttackData and other datasets demonstrate the effectiveness and superiority of the proposed approach in unified face attack detection. The contributions of the paper include the development of the UniAttackData dataset and the UniAttackDetection framework, which significantly enhance the performance of unified attack detection.The paper addresses the challenge of unified physical-digital face attack detection, which has been a significant gap in existing research. The authors propose a dataset called UniAttackData, which combines physical and digital attacks, ensuring ID consistency for each subject. This dataset includes 1,800 participants with 2 physical and 12 digital attacks each, resulting in a total of 29,706 videos. To address the computational burden and the difficulty in learning a compact feature space for both types of attacks, the authors develop UniAttackDetection, a unified attack detection framework based on Vision-Language Models (VLMs). The framework consists of three main modules: the Teacher-Student Prompts (TSP) module, the Unified Knowledge Mining (UKM) module, and the Sample-Level Prompt Interaction (SLPI) module. These modules work together to learn a robust and unified feature space for detecting both physical and digital attacks. Extensive experiments on UniAttackData and other datasets demonstrate the effectiveness and superiority of the proposed approach in unified face attack detection. The contributions of the paper include the development of the UniAttackData dataset and the UniAttackDetection framework, which significantly enhance the performance of unified attack detection.