On the Preservation of Soft-Biometric Information in Biometric Face Embeddings

The development of deeply-learned features is the foundation for the success of contemporary face recognition systems. These embeddings are designed to encode a person’s identity so that they can be used for identification. Recent studies, however, have demonstrated that these embeddings also store information on demographics, image qualities, and social traits in addition to the user’s identity. This brings up issues with prejudice and privacy in facial recognition. We examine the predictive power of 73 various soft-biometric features on three well-liked face embeddings with various learning philosophies. The tests were run on two databases that were accessible to the general public. We developed a huge attribute classifier that can accurately express the confidence in its predictions as part of the evaluation process. As a result, we are able to construct more complex statements concerning the property predictability. The findings show that most of the attributes under investigation are encoded in face embeddings. For instance, a robust encoding for accessories, accessories, and accessories was discovered. We discovered that these characteristics are particularly easy to anticipate from face embeddings, despite the fact that face recognition embeddings are taught to be resilient against non-permanent elements. Our research is intended to inform future efforts to create better bias-reducing and privacy-preserving face recognition technology.