Tom Tom Founders Festival 2018

This talk introduces and discusses a neural network architecture involving a convolutional autoencoder subnetwork capable of perturbing input data such that multiple objectives are optimized. In particular, the proposed architecture is applied to an important challenge in the field of biometrics: protecting privacy in face images while retaining the face recognition utility of biometric identification and verification systems. This novel optimization scheme, semi-adversarial training, is facilitated by a submodule consisting of an auxiliary gender-classifier and an auxiliary face matcher such that an objective function with three terms can be utilized for training: (1) to confound gender attributes of face images, (2) to ensure that the perturbed face images look realistic, and (3), to ensure that biometric face recognition performance is not substantially impacted. Extensive experiments confirm that the resulting model generalizes to multiple, independent face image databases as well as different gender classification and face matching software, showing that the proposed architecture comprises state-of-the-art performance for extending gender privacy to face images. Further, methods for extending this architecture to multiple soft biometric attributes are discussed as well as methods for improving the performance and interpretability of semi-adversarial neural network models.