Synthetic Data Generation for Face Anti-Spoofing Using Conditional Generative ModelsSubhan, Ali (Avtor) Peer, Peter (Mentor) Emeršič, Žiga (Komentor) Sabadin, Jernej (Komentor) deep learningbiometricsface anti-spoofinggenerative modelsdiffusion modelsconditional generationsynthetic datasetsFace anti-spoofing (FAS) systems achieve strong performance when train ing and evaluation data originate from the same dataset, yet their accuracy often deteriorates sharply under domain shift. Existing domain-generalisation research has primarily addressed this problem through new architectures and learning objectives, while comparatively little attention has been given to improving the diversity and composition of the training data. Moreover, prior generative approaches may require the identities used for synthesis to have been observed during generator training. This thesis investigates a data-centric approach based on instruction-guided image-to-image diffusion. Dataset-specific LoRA adapters are trained on FLUX.1-Kontext to learn presentation-attack transformations from paired live and spoof images. The learned transformations are then applied to live faces from other datasets, generating identity-preserving cross-domain spoofs that combine the identity and capture conditions of one domain with the attack characteristics of another. Synthetic samples are added exclusively to the source domains of the standard leave-one-domain-out OCIM protocols, thereby preserving the domain-generalisation setting. The proposed augmentation strategy is evaluated at several synthetic-to-real-spoof ratios using three architecturally distinct FAS models: the semantic, CLIP-based FLIP-MCL and GD-FAS models, and the texture-oriented SSAN model. The results reveal a clear representation-dependent effect. Semantic models generally benefit from synthetic augmentation, particularly on difficult transfer protocols; for ex ample, GD-FAS reduces mean HTER on OCM→I from 10.50% to 3.50%. Small augmentation ratios of 10–25% are typically more effective than larger doses, whereas SSAN does not improve consistently. These findings show that instruction-guided diffusion can strengthen cross-domain FAS, provided that the quantity of synthetic data and the detector’s underlying representation are carefully considered.20262026-11-03 10:00:01Magistrsko delo190757sl