Research On Privacy Protection For Wearable Device Data Based On Deep Learning

With the continuous innovation of artificial intelligence and microelectronics technology,applications based on wearable devices are developing rapidly,and the collection demand for users' data is growing day by day.However,acceleration and other data collected by wearable devices may expose users' privacy,from which attackers could infer some sensitive information.The key to reducing the risk of privacy leakage is to protect the users' identity information,that is,to reduce the identification risk of wearable device data.Although scholars have proposed some privacy protection methods based on deep learning,they have not been able to balance the utility and privacy risk of data well.Therefore,further research is still necessary.The main contributions of this thesis are as follows:(1)Aiming at the problem that existing methods distort data when protect users' identity,a privacy-preserving transformation for wearable device data based on the distinguish autoencoder(Dis AE)is proposed.By adding noise to the data to improve the difficulty of identification,and then through the autoencoder to reduce the data differences among different people,and further improve the difficulty of identification.The autoencoder is composed of convolutional neural network,and trained by distinguish loss.In the training process,firstly,constructing siamese autoencoder;then feeding sample pairs into network and calculating their latent variable distance,and multiplying different coefficients by distance depending on whether their activity labels or identity labels are the same;finally,training the autoencoder by loss function that sum of distance and reconstruction loss.(2)Aiming at the problem that existing methods cannot properly balance activity recognition and identity protection,a privacy-preserving transformation for wearable device data based on block discrete cosine transform and autoencoder(BDCT-AE)is proposed.The scheme first transforms raw data to the frequency domain,and then reconstruct to the time domain after autoencoder transforms.To constrain the utility and privacy risk of the reconstructed data explicitly,pre-trained activity classifiers and identity classifiers are connected to the output layer of encoder and decoder.The weighted sum of classification loss,privacy loss and reconstruction loss is used as the multi-objective loss for network training.Moreover,introducing DCT in the data preprocess can improve the performance of activity recognition and reduce reconstruction error.(3)To further improve the security of data collection,a data collection scheme for wearable devices based on local transformation and anonymous communication channels is proposed.In this scheme,data providers should process their data by Dis AE or BDCT-AE locally,and send transformed data to the data collector through random routing and RSA encryption algorithm,so that reduce the risk of identity leakage from the data level and communication level.Experimental results show that the Dis AE and BDCT-AE schemes designed in this thesis have better privacy protection effects than some existing schemes.In addition,performance tests on mobile devices show that the data collection is real-time.