可穿戴设备流数据的隐私保护发布

Privacy-preserving publishing of streaming data for wearable devices

针对可穿戴设备流数据可能泄露个人隐私的问题,提出了一种基于自编码器和时频变换的隐私保护数据发布方法。通过分块离散余弦变换将滑动窗口数据变换为频谱数据,再通过自编码器实现脱敏变换,最后由重构的频谱数据逆变换回滑动窗口数据。利用预训练的活动识别与身份识别分类器评估自编码器输出结果的效用性和隐私性,通过多目标损失函数与反向传播更新自编码器权重。在Motion-Sense数据集上的实验结果表明,在重构数据上活动识别的F_(1)-score由0.944降低至0.940,而身份识别的F_(1)-score由0.908降低至0.673,重构加速度数据与原数据之间的均方误差为0.27。与同类算法相比,该算法能够更好地保留数据的效用性以及提高数据的安全性。

This paper proposed a privacy-preserving data publishing method based on autoencoder and time-frequency transformation to avoid privacy disclosure from the stream data of wearable devices.Firstly,this method transformed slide window data into frequency spectrum by the block discrete cosine transformation.Secondly,it desensitized frequency spectrum by autoencoder.Finally,it obtained slide window data from reconstructed frequency spectrum by the inverse block discrete cosine transformation.This paper applied pre-trained activity and identity classifiers to evaluate the privacy and utility of autoenco-der’s output,then updated the weights of network by multi-objective loss function and back propagation.The experimental results on the Motion-Sense dataset show that the F_(1)-score of activity recognition on the reconstructed data is reduced from 0.944 to 0.940,the F_(1)-score of identity recognition is reduced from 0.908 to 0.673,and the mean square error between reconstructed acceleration data and original data is 0.27.Compared with similar algorithms,this proposed algorithm can better retain the utility of data and improve data security.