The application discloses a kind of multi-
modal face anti-fraud methods for domain generalization, it is related to face recognition technical field, multiple source domains under multi-
modal face image are collected, and
data set is constructed;Face anti-fraud model is constructed, and multiple
modal image coding module is configured for each source domain, and multiple modal fusion features are generated, and classification head is configured for each source domain, and the prediction result of corresponding domain is output according to multiple modal fusion features, domain
relationship extraction module is configured, the inter-domain relationship weight between each source domain is extracted, and the prediction result of each source domain is fused according to inter-domain relationship weight, and the final prediction result is obtained;The model is trained, including intra-domain modal level optimization and
domain level optimization.The application introduces the center difference expansion
convolution module of adaptive
receptive field, combines inter-domain relationship weighted multi-head fusion framework, solves the problem that multi-modal
feature extraction is insufficient in the prior art, and the cross-domain generalization ability is weak, realizes complex scene, unknown domain under accurate face anti-fraud detection.