How visual embeddings leak identity and how to fix it
CVPR 2026 paper overview with research scientist Daniel George, a coauthor of “From Measurement to Mitigation: Quantifying and Reducing Identity Leakage in Image Representation Encoders with Linear Subspace Removal." He discusses some of Persona’s recent research efforts, embeddings, and the paper’s focus.
The paper was accepted to the Conference on Computer Vision and Pattern Recognition (CVPR) 2026, a premier conference in computer vision and machine learning.
It addresses a fundamental challenge in computer vision: how can we ensure that visual embeddings (numerical vector representations of images) don't inadvertently leak identity information while still preserving their utility for non-biometric tasks?
At scale, identity verification relies on processing millions of images, and advancing state-of-the-art approaches in computer vision and machine learning is essential to building identity infrastructure the internet can trust.
An interactive and simplified version of the paper: https://withpersona.com/research/from-measurement-to-mitigation-quantifying-and-reducing-identity-leakage