Recently, multi-task learning (MTL) has been extensively studied for various face processing tasks, including face detection, landmark localization, pose estimation, and gender recognition. This approach endeavors to train a better model by exploiting the synergy among the related tasks. However, the raw face dataset used for training often contains sensitive and private information, which can be maliciously recovered by carefully analyzing the model and outputs. To address this problem, we propose a novel privacy-preserving multi-task learning approach that utilizes the differential private stochastic gradient descent algorithm to optimize the end-to-end multi-task model and weighs the loss functions of multiple tasks to improve learning efficiency and prediction accuracy. Specifically, calibrated noise is added to the gradient of loss functions to preserve the privacy of the training data during model training. Furthermore, we exploit the homoscedastic uncertainty to balance different learning tasks. The experiments demonstrate that the proposed approach yields differential privacy guarantees without decreasing the accuracy of HyperFace under a desirable privacy budget.
Keywords: balance different learning tasks; differential privacy guarantees; differential private stochastic gradient descent; multi-task learning; privacy preserving.
Copyright © 2020 Zhang, Hu, Xie, Gong and Yu.