Federated Face Recognition using Curriculum Learning

This work proposes a federated FR framework that uses curriculum learning to sort the images from “easy” to “difficult” during training. Two distinct curricula are considered: Face image quality assessment (FIQA) scores and head rotation. The performance of these curriculum designs is assessed both for fully-supervised and semi-supervised federated face recognition setups.

Head Rotation-based Curriculum

We sort the images from "easy" to "difficult" based on head pose angles. The difficulty is estimated by computing the sum of the absolute values of the pitch, yaw, and roll angles using the OpenFace 2.0 toolkit.

• Image Sorting: For each client, images are ranked by the sum of absolute head pose angles.
• Subset Splitting: We tried different splitting criteria but the best results were obtained using the following as shown in our preliminary work:
  • 🟩 Subset 1: Top 50% (with low head pose angles).
  • 🟥 Subset 2: Bottom 50% (with high head pose angles).

See the Head Pose Sort folder for more details.

Face Image Quality-based Curriculum

We sort the images from "easy" to "difficult" based on face image quality scores which are estimated using CR-FIQA.

• Image Sorting: Each image is scored using CR-FIQA and sorted in descending order based on quality.
• Subset Splitting:
  • 🟩 Subset 1: Top 50% (with high CR-FIQA scores).
  • 🟥 Subset 2: Bottom 50% (with low CR-FIQA scores).

See the CR-FIQA Sort folder for more details.

Fully-Supervised Training

We use the FedFR algorithm for all the fully-supervised training approaches for each curriculum as follows:

• Stage 1: Train the model on Subset 1.
• Stage 2: Initialize with Stage 1 parameters and continue training on all samples (Subset 1 + Subset 2).

Semi-Supervised Training

We use the SS-FedFR algorithm for all the semi-supervised training approaches for each curriculum as follows:

• Stage 1: Train the model on Subset 1.
• Stage 2: Initialize with Stage 1 parameters and continue training on all samples (Subset 1 + Subset 2).

See the SS-FedFR folder for more details.

References

Baltrušaitis, T., Robinson, P., & Morency, L.P. (2016). OpenFace: An open source facial behavior analysis toolkit. In 2016 IEEE Winter Conference on Applications of Computer Vision (WACV) (pp. 1–10).

Boutros, F., Fang, M., Klemt, M., Fu, B., & Damer, N. (2023). CR-FIQA: Face Image Quality Assessment by Learning Sample Relative Classifiability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 5836–5845).

Liu, C. T., Wang, C. Y., Chien, S. Y., & Lai, S. H. (2022, June). FedFR: Joint optimization federated framework for generic and personalized face recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 36, No. 2, pp. 1656-1664).