Semi-Adversarial Networks: Convolutional Autoencoders for Imparting Privacyto Face Images
share
In this paper, we design and evaluate a convolutional autoencoder that perturbs an input face image to impart privacy to a subject. Specifically, the proposed autoencoder transforms an input face image such that the output image can be successfully used for face recognition but not for gender classification. The proposed network uses a new type of learning scheme developed in this work -- referred to as semi-adversarial training -- in order to accomplish the stated purpose. The network has three subnetworks: a convolutional auto-encoder, a pseudo gender classifier, and a pseudo face matcher. A two-stage training scheme is used. The objective function utilized for training this network has three terms: one to ensure that the perturbed image is a realistic face image; another to ensure that the gender attributes of the face are confounded; and a third to ensure that biometric recognition performance due to the perturbed image is not impacted. Experiments confirm the efficacy of the proposed architecture in extending gender privacy to face images.
READ FULL TEXT