The paper presents a Multi-Domain Incremental Learning (MDIL) method for Face Presentation Attack Detection (PAD) to address the issue of catastrophic forgetting when adapting to new domains. The proposed method, called Adaptive Domain-specific Experts (ADE), aims to maintain performance in previous domains while learning from new ones. Key components include: 1. **ADE Framework**: This framework consists of ADE blocks, an Instance-wise Router (IwR), and an asymmetric classifier. ADE blocks handle domain-specific knowledge, IwR selects the appropriate expert for domain-agnostic instances, and the asymmetric classifier ensures consistent predicted probability distributions across different classifiers. 2. **IwR Module**: The IwR module learns domain centers and predicts gate signals to choose the most suitable expert branch during inference, maintaining domain-specific parameter isolation and sharing domain-invariant parameters. 3. **Asymmetric Classifier**: This classifier consolidates multiple classifiers into a single network, ensuring consistent predicted probability distributions and improving generalization. Experiments on various PAD datasets demonstrate that the proposed method outperforms state-of-the-art incremental learning (IL) methods and DA/DG-based methods, achieving state-of-the-art performance in both new and previous domains. The method effectively mitigates catastrophic forgetting and generalizes well to new domains, making it suitable for real-world applications.The paper presents a Multi-Domain Incremental Learning (MDIL) method for Face Presentation Attack Detection (PAD) to address the issue of catastrophic forgetting when adapting to new domains. The proposed method, called Adaptive Domain-specific Experts (ADE), aims to maintain performance in previous domains while learning from new ones. Key components include: 1. **ADE Framework**: This framework consists of ADE blocks, an Instance-wise Router (IwR), and an asymmetric classifier. ADE blocks handle domain-specific knowledge, IwR selects the appropriate expert for domain-agnostic instances, and the asymmetric classifier ensures consistent predicted probability distributions across different classifiers. 2. **IwR Module**: The IwR module learns domain centers and predicts gate signals to choose the most suitable expert branch during inference, maintaining domain-specific parameter isolation and sharing domain-invariant parameters. 3. **Asymmetric Classifier**: This classifier consolidates multiple classifiers into a single network, ensuring consistent predicted probability distributions and improving generalization. Experiments on various PAD datasets demonstrate that the proposed method outperforms state-of-the-art incremental learning (IL) methods and DA/DG-based methods, achieving state-of-the-art performance in both new and previous domains. The method effectively mitigates catastrophic forgetting and generalizes well to new domains, making it suitable for real-world applications.