Bit Stability and Hash-Length Trade-Offs in Binary Face Templates

Binary face templates are an appealing alternative to floating-point embeddings for face verification in resource-constrained environments because they enable constant-time Hamming matching with minimal storage and input/output (I/O). This paper studies the bit-level behavior of hashes obtained by principal component analysis followed by iterative quantization (PCA–ITQ) at L∈{32, 64, 128} derived from a frozen lightweight face encoder. Using subject-disjoint splits on the MORPH longitudinal dataset and an eyeglasses stress protocol on CelebA, the analysis quantifies 1) bit balance and entropy, 2) within-identity bit stability via per-bit flip rates, and 3) verification performance at low false-accept rates in Hamming space. On MORPH, 64-bit PCA–ITQ codes achieve an area under the receiver operating characteristic curve (AUC) of 0.9978 and a true positive rate (TPR) of 96.5% at a false positive rate (FPR) of 1%, compared to 99.1% at 128 bits, while halving the template length; 32-bit codes remain feasible but drop to 85.7% at the same operating point and are more sensitive to nuisance variation. Across both datasets, codes are near-balanced and mostly stable, yet a small minority of bit positions accounts for most flips under the eyeglasses attribute. In this regime, 64-bit hashes offer a favorable size–accuracy trade-off, whereas 128-bit hashes approximate float-embedding behavior and 32-bit hashes require redundancy or additional robustness mechanisms. All evaluations use fixed seeds and subject-disjoint splits; thresholds are selected on validation and held fixed on test to reflect deployment conditions.

@article{Ganmati2025,
  title     = {Bit Stability and Hash-Length Trade-Offs in Binary Face Templates},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {16},
  number    = {11},
  year      = {2025},
  publisher = {The Science and Information Organization},
  author    = {Abdelilah Ganmati and Karim Afdel and Lahcen Koutti},
  doi       = {10.14569/IJACSA.2025.0161192},
  url       = {https://doi.org/10.14569/IJACSA.2025.0161192}
}