Deep Face Decoder: Towards understanding the embedding space of convolutional networks through visual reconstruction of deep face templates.

Advances in deep learning and convolutional neural networks (ConvNets) have driven remarkable face recognition (FR) progress recently. However, the black-box nature of modern ConvNet-based face recognition models makes it challenging to interpret their decision-making process, to understand the reasoning behind specific success and failure cases, or to predict their responses to unseen data characteristics. It is, therefore, critical to design mechanisms that explain the inner workings of contemporary FR models and offer insight into their behavior. To address this challenge, we present in this paper a novel template-inversion approach capable of reconstructing high-fidelity face images from the embeddings (templates, feature-space representations) produced by modern FR techniques. Our approach is based on a novel Deep Face Decoder (DFD) trained in a regression setting to visualize the information encoded in the embedding space with the goal of fostering explainability. We utilize the developed DFD model in comprehensive experiments on multiple unconstrained face datasets, namely Visual Geometry Group Face dataset 2 (VGGFace2), Labeled Faces in the Wild (LFW), and Celebrity Faces Attributes Dataset High Quality (CelebA-HQ). Our analysis focuses on the embedding spaces of two distinct face recognition models with backbones based on the Visual Geometry Group 16-layer model (VGG-16) and the 50-layer Residual Network (ResNet-50). The results reveal how information is encoded in the two considered models and how perturbations in image appearance due to rotations, translations, scaling, occlusion, or adversarial attacks, are propagated into the embedding space. Our study offers researchers a deeper comprehension of the underlying mechanisms of ConvNet-based FR models, ultimately promoting advancements in model design and explainability. [Display omitted] Above are example DFD reconstructions of selected sample images from a single subject in the LFW dataset. The rows display: original images (1st row), reconstructions from the VGG–VGG DFD model (2nd row), reconstructions from the VGG–ResNet DFD model (3rd row), reconstructions from the ResNet–VGG DFD model (4th row), and reconstructions from the ResNet–ResNet DFD model (5th row). • We present the Deep Face Decoder (DFD), a novel SOTA template inversion technique. • We use DFD to analyze the embedding space of ConvNet face recognition (FR) models. • Using DFD we explore the impact of image perturbations, occlusions, and attacks. • Through DFD we investigate various template construction strategies. [ABSTRACT FROM AUTHOR]