Differentially Private Asynchronous Federated Learning with Buffered Aggregation for Face Recognition
Tongkun Xing, Ming Zhang · 2024
The rise of privacy concerns in face recognition applications necessitates the development of secure machine learning frameworks. This paper introduces Differentially Private Asynchronous Federated Learning with Buffered Aggregation (DP-AFLBA), a novel approach that combines differential privacy, asynchronous federated learning, and buffered aggregation to protect sensitive data while improving model accuracy. In DP-AFLBA, local model updates from participants are perturbed with carefully calibrated noise to ensure privacy, adhering to (ϵ, δ)-differential privacy standards. Asynchronous communication allows participants to send updates independently, reducing delays associated with traditional synchronous methods. To address the challenges of stale updates, a buffering mechanism aggregates updates over a defined time window, enabling the server to apply a more stable and accurate global model update. Experimental results demonstrate that DP-AFLBA achieves superior performance in terms of model accuracy and privacy protection compared to existing federated learning frameworks, making it particularly suitable for large-scale face recognition tasks.