Meta-learning-based adversarial training for deep 3D face recognition on point clouds

Recently, deep face recognition using 2D face images has made great advances mainly due to the readily available large-scale face data. However, deep face recognition using 3D face scans, especially on point clouds, has been far from fully explored. In this paper, we propose a novel meta-learning-based adversarial training (MLAT) algorithm for deep 3D face recognition (3DFR) on point clouds. It consists of two alternate modules: adversarial sample generating for 3D face data augmentation and meta-learning-based deep network training. In the first module, adversarial samples of given 3D face scans are dynamically generated based on current deep 3DFR model. In the second module, a meta-learning framework is designed to avoid the performance decrease caused by the generated adversarial samples. Overall, MLAT algorithm combines the adversarial sample generating and meta-learning-based network training in a uniform framework, in which adversarial samples and network parameters are optimized alternately. Thus, it can continuously generate diverse and suitable adversarial samples, and then the meta-learning framework can further improve the accuracy of 3DFR model. Comprehensive experimental results show that the proposed approach consistently achieves competitive rank-one recognition accuracies on the BU-3DFE (10 0%), Bosphorus (99.78%), BU-4DFE (98.02%) and FRGC v2 (98.01%) database, and thereby substantiate its superiority.

Automated summary

Recently, deep face recognition using 2D face images has advanced due to the availability of large-scale face data. However, deep face recognition using 3D face scans on point clouds still needs further exploration. This paper proposes a meta-learning-based adversarial training algorithm for deep 3D face recognition on point clouds. The algorithm combines adversarial sample generation and meta-learning-based network training to continuously generate diverse adversarial samples and improve the accuracy of the 3DFR model.