Face Image Reconstruction from Deep Templates

State-of-the-art face recognition systems are based on deep (convolutional) neural networks. Therefore, it is imperative to determine to what extent face templates derived from deep networks can be inverted to obtain the original face image. In this paper, we discuss the vulnerabilities of a face recognition system based on deep templates, extracted by deep networks under image reconstruction attack. We propose a de-convolutional neural network (D-CNN) to reconstruct images of faces from their deep templates. In our experiments, we did not assume any knowledge about the target subject nor the deep network. To train the D-CNN reconstruction models, we augmented existing face datasets with a large collection of images synthesized using a face generator. The proposed reconstruction method was evaluated using type-I (comparing the reconstructed images against the original face images used to generate the deep template) and type-II (comparing the reconstructed images against a different face image of the same subject) attacks. We conducted a three-trial attack for each target face image using three face images reconstructed from three different D-CNNs. Each D-CNN was trained on a different dataset (VGG-Face, CASIA-Webface, or Multi-PIE). The type-I attack achieved a true accept rate (TAR) of 85.48% at a false accept rate (FAR) of 0.1% on the LFW dataset. The corresponding TAR for the type-II attack is 14.71%. Our experimental results demonstrate the need to secure deep templates in face recognition systems.