AFRBench / AFRLib

AFRBench evaluates privacy protection methods against face-recognition-based image retrieval. AFRLib brings these methods behind shared interfaces so researchers can train, protect images, and evaluate them through a common workflow.

The two components serve different roles:

• AFRLib: 26 anti-facial-recognition method implementations with consistent interfaces.
• AFRBench: a common retrieval evaluation protocol for 24 methods, with support for the 45+ models supplied by FRBench.
• Evaluation data: 5,000 photos of 50 protectees, with 22.8 million distractor faces for large-scale retrieval evaluation.
• AFRBench-DF: a complementary dataset derived from DeeperForensics source videos for studying pose and expression variation.

The public release includes configuration-based experiment workflows, dataset preparation tools, and documentation for extending the library. FR models are supplied by FRBench.

The first public release, v1.0.0, was published on 2 October 2026. See the repository and documentation for the current scope and release roadmap.