Design And Implementation Of Wearable Devices Oriented Privacy Preservation System Based On Deep Learning

Due to the diversity of sensors on wearable devices and the ubiquity of wireless broadband connections,wearable devices can be used to collect a variety of information about users and their surroundings.Compared with traditional wireless sensors networks,mobile sensing applications have the advantages on mobility,scalability and cost-efficiency.In the wearable devices based sensing applications,users are required to report their sensed data,which result in user privacy leakage.Existing privacy preservation approaches often changes all the sensed data(e.g.,perturbation or transformation).It will reduce the data utility and impose negative impacts on data mining.A better choice is to keep non-private data unchanged and only handle private data.However,the main problem is how to accurately pick out the non-private data from the sensed data.We formulate this task as a one-class classification problem,and develop a novel discrimination model based on deep learning.More specifically,the model adopts an AutoEncoder to learn the discriminative representation of non-private data,and employs multiple generative adversarial networks(GANs)to stimulate the AutoEncoder to improve its performance.To improve the discrimination of our model,we propose to combine the individual discrimination decisions using majority voting.Our proposed approach has three advantages.First,it does not require collecting private data for model training,making it easy to deploy.Second,it is applicable to many MCS applications,since it does not require changing the non-private data.Last but not least,it is able to learn discriminative representation of non-private data and hence has a high decision accuracy.The experiments over three public datasets and one self-constructed dataset show that our approach outperforms four state-of-art related approaches with respect to accuracy and F1-score.To verify the feasibility of our approach,we designed and implemented a private activity recognition and processing system based on our proposed model.The systems collects users' sensed data and then eliminates the private part.The server of the system is based on microservices,which can be divided into three parts: general module,message system and algorithm deployment.In addition to the server,we develop an Android APP for smart phone and smart watches.Finally,the performance test verified that the proposed approach is feasible in practice.