Skip to content

AFRBench v1.0.0

Latest
Latest

Choose a tag to compare

View all tags
@Michael-wzl Michael-wzl released this 02 Oct 12:41
· 6 commits to main since this release
v1.0.0
fa50166

AFRBench v1.0.0

The first public release of AFRLib, a library of anti-facial-recognition (AFR) methods, and
AFRBench, a benchmark that evaluates them under one protocol.

AFRLib

  • 26 published AFR methods (2016–2026) behind one interface: train, fit, and protect on
    original photos, with the inputs of each stage declared by the method.
  • Attack surrogates are injected, so any method runs against any face-recognition model: the
    45 FRBench models, FaceNet, or your own.
  • Methods and FR models load from outside the library by import path ("package.module:Class").
  • Method weights download on first use (~/.afrlib); python -m afrlib lists, fetches, and
    registers them.

AFRBench

  • The AFRBench-basic protocol:
    • 50 protectees with 100 photos each, and fixed album, query, and gallery roles;
    • 5 high-quality impersonation targets and a pool of 2,000 target identities;
    • 22.8M MillionCelebs distractor faces;
    • scenarios 1A, 1B, 2A, and 2B;
    • intruder pipelines with JPEG, resizing, and restoration defenses;
    • a 17-job evaluation suite.
  • python -m afrbench.benchmark runs the protocol on one method, built-in or your own, and can
    also score it with another FR model. Every step is a YAML job with a lock file, so runs resume
    and re-run only what changed.
  • Data:
    • the repository ships the dataset metadata and the five target images;
    • afrbench.prepare assembles the dataset from your copies of the protectee photos;
    • afrbench.millioncelebs downloads and packs the distractors.

Documentation

docs/ covers library use, the benchmark from a fresh machine to results, the protocol, the
data, per-method inputs, and extending AFRLib. examples/custom_method is a minimal method to
start from.

Known limitations

  • Jobs run one at a time on one GPU; there is no multi-GPU runner yet.
  • Trained checkpoints and precomputed distractor embeddings are not released yet.
  • The result viewer shown in the README is a preview and is not included.
  • Methods and vendored components keep their own licenses, some for research use only; see
    THIRD_PARTY.md.

Citation

AFRBench grows out of Protego (CVPR 2026); see the README for the BibTeX entry.

Assets 2
Loading