FRBench
FRBench is an easy-to-use, end-to-end differentiable face-recognition module developed at the HKU TASR Lab. It is designed to serve as a unified benchmark and plug-and-play backbone library for face recognition research and downstream applications (e.g., face privacy protection, adversarial robustness evaluation, and retrieval system testing).
See the GitHub repo for full documentation and usage.
Highlights
- 45+ pretrained weights — ready to use out of the box, no manual weight hunting.
- 25 backbones — from classic lightweight networks to modern large-capacity architectures.
- 9 loss functions — ArcFace, CosFace, SphereFace, AdaFace, and more.
- 3 training datasets — spanning a decade of face recognition benchmarks from 2015 to 2024.
- End-to-end differentiable — the entire pipeline (preprocessing → backbone → loss head) is differentiable, making FRBench natively compatible with gradient-based attacks, adversarial training, and privacy protection methods such as Protego.
Motivation
Most face recognition codebases scatter pretrained weights across different repositories with inconsistent interfaces, making fair cross-model comparison painful. FRBench centralises everything behind a single, consistent API so that:
- Researchers can swap backbones and loss functions with one line of code.
- Adversarial / privacy-protection methods can attack or evaluate any model without rewriting the forward pass.
- Results are reproducible — fixed preprocessing, fixed weight sources, fixed evaluation protocol.
Design
| Component | Details |
|---|---|
| Backbone zoo | 25 architectures including IResNet-{18,34,50,100,200}, MobileFaceNet, EfficientNet variants, ViT-based models, and more |
| Loss heads | ArcFace · CosFace · SphereFace · AdaFace · MagFace · ElasticFace · CurricularFace · AirFace · CircleLoss |
| Training sets | MS1MV2 (2019), WebFace4M (2021), and a 2024-generation large-scale dataset |
| Evaluation | LFW · CFP-FP · AgeDB-30 · IJB-B · IJB-C protocols built-in |
| Differentiability | Full gradient flow through image normalisation, backbone, and loss head — no NumPy detours |
Usage
import frbench
# Load a pretrained model in one line
model = frbench.load("iresnet50", loss="arcface", dataset="ms1mv2")
model.eval()
# Forward pass (fully differentiable)
embeddings = model(images) # images: (B, 3, 112, 112) tensor
Gradient-based usage (e.g., for adversarial perturbation):
images.requires_grad_(True)
loss = criterion(model(images), targets)
loss.backward() # gradients flow all the way back to pixel space