Generative Fingerprint Augmentation against Membership Inference Attacks

This thesis aspires to provide a privacy protection mechanism for neural networks concerning fingerprints. Biometric identifiers, especially fingerprints, have become crucial in the last several years, from banking operations to daily smartphone usage. Using generative adversarial networks (GANs), we train models specialized in compressing and decompressing (Codecs) images in order to augment the data these models used during the learning process to provide additional privacy preservation over the identity of the fingerprints found in such datasets. We test and analyze our framework with custom membership inference attacks (MIA) to assess the quality of our defensive mechanism.