Shrinking embeddings, not accuracy: Performance-preserving reduction of facial embeddings for complex face verification computations

Conventional embeddings employed in facial verification systems typically consist of hundreds of floating-point numbers, a widely accepted design paradigm that primarily stems from the swift computation of vector distance metrics for identification and authentication such as the L2 norm. However, the utility of such high-dimensional embeddings can become a potential concern when they are integrated into complex comparative strategies, for example multi-party computations. In this study, we challenge the presumption that larger embedding sizes are always superior and provide a comprehensive analysis of the effects and implications of substantially reducing the dimensions of these embeddings (by a factor of 29). We demonstrate that this dramatic size reduction incurs only a minimal compromise in the quality-performance trade-off. This discovery could lead to enhancements in computation efficiency without sacrificing system performance, potentially opening avenues for more sophisticated and decentral uses of facial verification technology. To enable other researchers to validate and build upon our findings, the Rust code used in this paper has been made publicly accessible and can be found at https://github.com/mobilesec/reduced-embeddings-analysis-icprs.