Discovering Interpretable Feature Directions in the Embedding Space of Face Recognition Models

Modern face recognition (FR) models, particularly their convolutional neural network based implementations, often raise concerns regarding privacy and ethics due to their "black-box" nature. To enhance the explainability of FR models and the interpretability of their embedding space, we introduce in this paper three novel techniques for discovering semantically meaningful feature directions (or axes). The first technique uses a dedicated facial-region blending procedure together with principal component analysis to discover embedding space direction that correspond to spatially isolated semantic face areas, providing a new perspective on facial feature interpretation. The other two proposed techniques exploit attribute labels to discern feature directions that correspond to intra-identity variations, such as pose, illumination angle, and expression, but do so either through a cluster analysis or a dedicated regression procedure. To validate the capabilities of the developed techniques, we utilize a powerful template decoder that inverts the image embedding back into the pixel space. Using the decoder, we visualize linear movements along the discovered directions, enabling a clearer understanding of the internal representations within face recognition models. The source code will be made publicly available.