Edge Computing Network Privacy Protection Method Based on Federated Learning

Abstract

As a new paradigm of sensing data collection, mobile crowdsensing is widely used in various life scenarios, but it also faces serious data privacy leakage problems. Most of the existing methods are based on the privacy protection of the original data, and do not consider the problem of internal theft and external attackers intercepting data. Based on these, an edge computing network privacy protection method based on federated learning is proposed. Firstly, participants use federated learning to train sensing data locally to obtain the local models, avoiding the interaction of the original data with edge computing nodes and the sensing platform. Secondly, the trained model parameter values are perturbed by local differential privacy and process-encrypted by homomorphic encryption, and then uploaded to edge computing nodes. Finally, edge computing nodes upload the encryption model parameters obtained by edge aggregation to the sensing platform for global aggregation operation in the form of ciphertext. Through security analysis, the proposed method can effectively prevent the leakage of participants' private information. The simulation results show that the method has a higher privacy protection level and accuracy, and less time cost, which can achieve a balance between privacy protection level and model performance.

Keywords: Mobile crowdsensing, edge computing, Federated Learning, Local differential privacy