DAFL: Domain adaptation-based federated learning for privacy-preserving biometric recognition

Abstract

With the introduction of data protection regulation in various countries, traditional centralized learning for the exploitation of sensitive biological information will gradually become unsustainable. We take face and speaker recognition systems as examples, to consider the possibility of implementing them using federated learning. It is known that federated learning methods still have some issues to be solved, e.g., parameter updates delivered can still reveal user privacy. Besides, existing privacy-preserving methods such as encryption and perturbation could pose other problems, including low accuracy and high training overhead, which can be more prominent in biometric recognition scenarios. We adopt the idea of domain adaptation and propose a new federated learning framework, DAFL, focusing on the improvement of the training efficiency and model performance, to achieve a practical privacy-preserving biometric recognition scheme for the first time. We also design a special dropout method for it to address the possible imbalance of biometric samples across users. Experiments on real-world datasets show that our approach can achieve up to 95.6% and 97.9% accuracy in face recognition and speaker recognition tasks, respectively, under medium privacy settings. Compared with existing work, it can save 56.0% of training time with a 3.1% accuracy loss in the best case.