Privacy-Preserving Social Recommendation System Based On Wearable Devices

The rapid development of the Internet has enriched users' social networks,but that makes users' address book management cumbersome.At the same time,online social networks mainly rely on users' social information in cyberspace but rarely use information about social activities in the physical world.The development of wearable devices provides us with an opportunity to mine users' social information from physical interactions.However,we are still facing many challenges.Firstly,the wearable sensor data has the problems of heterogeneity and imbalance.Differences in the characteristics of hardware devices and operating systems can cause noise in the collected data,resulting in differences in data quality.In the meanwhile,different users have different postures and behavior habits.There are differences in the frequency of different activities,which leads to an imbalance in the wearable sensor data.Secondly,unlike single-person activity recognition,users' social activity recognition contains more than one behavioral activity,which is more complicated.How to design an activity recognition model with high accuracy and low resource consumption on resource-constrained mobile devices is a big challenge.Finally,there may be multiple social groups in the same social area.How to accurately identify the groups of users'social interactions(especially asynchronous social activities)from the wearable sensor data is difficult.On this basis,we also need to ensure that users' private data is safe from being leaked.In order to solve the above problems,we proposed a privacy-preserving social rec-ommendation system based on wearable devices to help users achieve social recom-mendation functions such as friend recommendation and automatic labeling.This main work of this paper:(1)We collected 36 volunteers synchronous and asynchronous ac-tivity wearable sensor data.In order to ensure that the features of the data are preserved as much as possible while filtering data noise,we used residual analysis to find the best cut-off frequency for filtering.(2)We designed a lightweight neural network architec-ture based on voting for activity recognition.For a three-minute data segment,social activity recognition can achieve 100%accuracy.The model size is only 416KB,which is suitable for resource-constrained mobile devices.(3)By analyzing the sensor data features of users' social activities,we proposed a social interaction feature generation algorithm to generate interactive features and used gradient boosting tree classifiers to detect whether there are social activities among users.For example,the recognition ac-curacy for whether two people walking together reaches 92.2%,and whether two people playing ping-pong together reaches 92.6%accuracy.(4)We designed and evaluated a distributed privacy-preserving social mining framework based on garbled circuits and a multi-granularity data reconstruction model based on convolutional autoencoders to ensure the safety of users' social activity information.