Biometric technology utilizes unique physiological features of the human body, such as fingerprints, facial features, and iris patterns, to authenticate users’ identities. This technology offers both convenience and security and has found widespread applications in access control, mobile payment systems, airport security, unmanned stores, and various other fields. However, face recognition systems are particularly susceptible to presentation attacks (PAs), presenting a significant security challenge. Specifically, 3D mask presentation attacks involve an attacker wearing facial masks made of different materials in an attempt to mimic a legitimate user’s face and deceive the face recognition system. Unlike 2D attacks that rely on general images or videos, 3D mask presentation attacks pose a greater challenge for detection and defense. Consequently, these attacks represent a new and concerning security threat due to their increased difficulty in identification and prevention. We proposed a 3D mask presentation attack detection technology for facial recognition systems using meta-learning, specifically leveraging the MAML (Model-Agnostic Meta-Learning) algorithm. This approach offers superior performance and generalization capabilities, particularly beneficial for small sample datasets. Our optimized model, derived from the meta-learning framework, underwent experimentation using the HKBUMARs and MLFP databases to detect 3D mask presentation attacks. The experimental results demonstrated an impressive accuracy rate of 95.28%. This achievement signifies robust and reliable performance in identifying and mitigating 3D mask presentation attacks, thereby significantly enhancing the security and resilience of facial recognition systems.