Privacy Preservation in User Behavior Analysis for Mobile-Edge Computing

In mobile edge computing, user behavior analysis can provide a more clear understanding of edge user behavior, which is crucial for providing more accurate recommendation services to mobile users. However, privacy leakage is inevitable in the process of edge user behavior data collection, storage and analysis. How to protect edge user privacy is very urgent for the development of mobile edge computing. Existing schemes dealing with these problems still suffer from 1) high computational overhead, 2) low model accuracy and poor convergence under privacy-preserving, and 3) poor model robustness when edge nodes fail. Therefore, we propose a novel privacy-preserving model in edge user behavioral analysis termed Safe-FL to protect the behavioral privacy of all edge participants by introducing secure data aggregation and encryption in the federated learning (FL) framework. To reduce the privacy leakage of data uploading in edge user behavior analysis, we first construct a federated learning-based edge user behavior analysis framework. Secondly, we utilize a hybrid encryption scheme to encrypt the local model hierarchically to protect the gradient privacy in edge user behavior analysis and reduce the encryption computation overhead. And finally, to improve the robustness of the model, we propose a decentralized secure gradient aggregation algorithm based on blockchain. Extensive experiments are conducted on two public datasets, whose results illustrate that our model has 1) strong privacy preservation and small computational overhead; 2) high accuracy and convergence speed; and 3) good robustness, compared to other models.