Research On Key Technologies Of Data Privacy Protection For Edge Computing

Privacy computing provides privacy protection technical support for the free flow of data between terminal devices and untrusted edge servers,but it still faces challenges in how to integrate into all aspects of data mining in edge computing to achieve a balance among privacy,utility,and computing efficiency.This paper conducts research from three aspects: data collection,aggregation,and inference.In terms of data collection,data is generated and stored in users’ terminal devices.Due to concerns about privacy disclosure and computing overhead,users are unwilling to contribute high-quality data free of charge.As the core factor of production in the digital economy era,data shortage or low quality means that it cannot effectively supply energy for the digital industry.In terms of data aggregation,due to the unbalanced technical levels of users or perception differences of devices,the quality of data contributed by users is uneven.Average aggregation will lead to inefficient data quality and increase the risks of privacy disclosure.In terms of data inference,machine learning is an important tool for big data mining,but its powerful learning ability enables it to infer private information from insensitive data,bringing indirect privacy threats.In view of the balance of privacy,utility and efficiency in the three stages of data collection,aggregation,and inference in edge computing,this paper proposes the privacy-preserving incentive mechanism,truth discovery,and streaming inference,respectively.The main innovations and contributions are summarized as follows:(1)For the data collection scenario of federated learning,a dual-privacy preserving and quality-aware incentive mechanism is proposed.The exponential mechanism and the Gaussian mechanism are used to respectively protect the bidding and local models to defend against untrusted central servers.Considering that there are differences among users in three aspects: the amount of training data,the degree of non-IID distribution,and the degree of privacy protection of local models,a multidimensional scoring function is designed to evaluate user quality.Based on the perturbed bids and estimated qualities,a multi-dimensional reverse auction-based incentive mechanism is proposed to maximize social welfare.It is theoretically proved that the proposed mechanism guarantees γ-truthfulness,individual rationality,and computational efficiency.Experimental results show that,compared with the state-of-the-art,this mechanism can effectively protect bid privacy and model privacy,and improve social welfare and model accuracy by at least 21% and 6%.(2)For the streaming data aggregation scenario of edge computing,a privacy-preserving streaming truth discovery algorithm toward edge computing is proposed.The algorithm introduces edge servers between the untrusted cloud server and users,and uses a homomorphic encryption algorithm to securely calculate the local truths and users’ reliabilities.According to the changes of local truths and user’s reliabilities,the disturbance timestamp and disturbance magnitude of differential privacy are dynamically adjusted so as to reduce the introduced noise and improve the aggregation accuracy.The expected utility and computational overhead of the proposed algorithm are theoretically analyzed,and the w-event differential privacy is proved to be satisfied.Experimental results show that compared with the state-of-the-arts at the same level of privacy protection,the proposed algorithm can achieve at least 47% improvement in aggregation accuracy on real-world datasets.(3)For the streaming data inference scenario of edge computing,a privacy-preserving inference framework for streaming data is proposed.Adaptive sampling,sensitive feature filtering,differential privacy perturbation,and data reconstruction are performed on streaming data to defend against attackers’ sensitive inferences while ensuring the accuracy of target task inferences.Considering the variable length of streaming data,the multi-layer perceptron(MLP)and recurrent neural network(RNN)are respectively used to construct the sensitive feature filtering model,and a distributed stream processing platform is used to deploy the proposed inference framework to realize real-time transmission of streaming data.Experimental results show that compared with the inference accuracy of the original data,the accuracy of the proposed framework on the target inference task is only reduced by 9% at most,but the accuracy of sensitive inference is greatly reduced,which is close to the probability of random guessing.The delay,computation and storage overhead of this framework are within a reasonable range.This paper effectively improves the balance among privacy,utility,and efficiency in the process of data collection,aggregation,and inference under edge computing,and provides a reference for the implementation of privacy computing in edge computing applications and the release of data value.It meets the data security requirements of the country and society,and has important theoretical significance and practical value.