This paper proposes a Multi-Domain Incremental Learning (MDIL) framework for Face Presentation Attack Detection (PAD), which addresses the issue of catastrophic forgetting when learning from new domains. The framework introduces an Adaptive Domain-specific Experts (ADE) architecture that maintains the performance of previous domains while learning from new ones. The ADE framework includes a shared encoder and expert decoder, with the expert decoder being designed to learn domain-specific knowledge independently. Additionally, an asymmetric classifier is introduced to ensure consistency in predicted probability distributions across different domains, improving generalization ability. The ADE framework is combined with an Instance-wise Router (IwR) that selects the most appropriate expert branch for domain-agnostic instances. The IwR learns domain centers during training and uses similarity to these centers to determine the appropriate expert branch during inference. The asymmetric classifier is designed to handle the sparsity and discreteness of spoof samples, ensuring consistent predicted probability distributions across different domains. Extensive experiments on widely used benchmark datasets demonstrate that the proposed method achieves state-of-the-art performance compared to prior methods of incremental learning. Under more stringent settings, the method approximates or even outperforms DA/DG-based methods. The framework is effective in scenarios where data privacy constraints prevent access to previous domain data, making it suitable for PAD applications with potential domain transfer. The method is evaluated on five PAD datasets, including OULU-NPU, CASIA-MFSD, Idiap ReplayAttack, MSU-MFSD, and SiW. Results show that the proposed method significantly improves performance on previous domains and maintains optimal performance across different steps. The method also performs well in cross-domain and cross-attack-type settings, demonstrating its effectiveness in mitigating catastrophic forgetting and preserving generalization capability. The framework is visualized and analyzed to show the effectiveness of the proposed method in reducing performance degradation and maintaining consistent predicted probability distributions.This paper proposes a Multi-Domain Incremental Learning (MDIL) framework for Face Presentation Attack Detection (PAD), which addresses the issue of catastrophic forgetting when learning from new domains. The framework introduces an Adaptive Domain-specific Experts (ADE) architecture that maintains the performance of previous domains while learning from new ones. The ADE framework includes a shared encoder and expert decoder, with the expert decoder being designed to learn domain-specific knowledge independently. Additionally, an asymmetric classifier is introduced to ensure consistency in predicted probability distributions across different domains, improving generalization ability. The ADE framework is combined with an Instance-wise Router (IwR) that selects the most appropriate expert branch for domain-agnostic instances. The IwR learns domain centers during training and uses similarity to these centers to determine the appropriate expert branch during inference. The asymmetric classifier is designed to handle the sparsity and discreteness of spoof samples, ensuring consistent predicted probability distributions across different domains. Extensive experiments on widely used benchmark datasets demonstrate that the proposed method achieves state-of-the-art performance compared to prior methods of incremental learning. Under more stringent settings, the method approximates or even outperforms DA/DG-based methods. The framework is effective in scenarios where data privacy constraints prevent access to previous domain data, making it suitable for PAD applications with potential domain transfer. The method is evaluated on five PAD datasets, including OULU-NPU, CASIA-MFSD, Idiap ReplayAttack, MSU-MFSD, and SiW. Results show that the proposed method significantly improves performance on previous domains and maintains optimal performance across different steps. The method also performs well in cross-domain and cross-attack-type settings, demonstrating its effectiveness in mitigating catastrophic forgetting and preserving generalization capability. The framework is visualized and analyzed to show the effectiveness of the proposed method in reducing performance degradation and maintaining consistent predicted probability distributions.