Difficulties in Using Synthetic Data for Presentation Attack Detection in Finger Vein Recognition: The Role of Model Fingerprints

Four distinct GAN-based I2I translation techniques are employed for the synthesis of biometric finger vein presentation attack instrument (PAI) samples corresponding to three public presentation attack datasets. The PAD training using these synthetic PAI samples reveals weaknesses for a small share of settings (in terms of datasets and GAN types). Removing the GAN model fingerprints (with three technical variants) from the synthetic data is not resolving these problematic PAD results, in contrary, this strategy creates more problematic results than it was intended to resolve. Finally, we show that PAI samples generated with different GAN types can be easily discriminated, even when using identical GAN types but only different parameter setups the resulting synthetic data can still be differentiated. This indicates that the present GAN model fingerprints are stronger than often believed, eventually caused by the significant redundancy present in our biometric datasets as compared to natural data as typically used in GAN fingerprint assessments. Overall, in the generation of synthetic PAI samples, CycleGAN as well as StarGANv2 generated specimens turn out to be highly useful to train finger vein PAD systems.