Research On Incentive Mechanism And Data Analysis Method Based On Differential Privacy In Crowdsourcing Platform

Since the concept of crowdsourcing has been put forward,it has received increasing attention from academia and industry.With the vigorous development of Internet technology and smart mobile devices,the execution of crowdsourcing tasks has become more efficient.However,with the widespread application of the crowdsourcing model,privacy issues also follow,since most people are not willing to disclose too much personal information when performing crowdsourcing tasks.Generally,the privacy issues of the crowdsourcing system mainly exist in the following two aspects: On one hand,there is a risk of disclosing bidding information of crowdsourcing workers in the incentive mechanism based on reverse auctions;On the other hand,training stage of the prediction model using the crowdsourcing data leaks workers' personal privacy.Aiming at the problem of privacy leakage in crowdsourcing scenarios,the main contributions of this thesis are as follows:Considering the problem of leakage of bidding information for crowdsourcing workers in the reverse auction-based incentive mechanism,this thesis proposes an incentive mechanism based on differential privacy,denoted as DP-mp RA.After the worker uploads his own bidding information,the platform determines the set of paid prices.For each price in the set,the final total payment amount is optimized under the condition that the error threshold after data aggregation is met,and the winner set is selected.The differential privacy mechanism is used to select the final payment scheme with a certain probability.Finally,differentiated payments based on the level of each worker further reduce the total payment of the platform.Through rigorous theoretical proof,the incentive mechanism DP-mp RA satisfies differential privacy protection,authenticity and personal rationality.The simulation results demonstrate that the DP-mp RA mechanism is superior to the existing incentive mechanism,that is,under the same privacy protection level,the total platform payment obtained by the DP-mp RA mechanism is lower than the total platform payment obtained by the existing mechanism.At the stage of training the predictive model using the private data provided by the worker,if the predictive model is directly released without privacy protection,the attacker can infer the original data used to train the model,thus exposing the worker's privacy.This paper proposes a deep learning scheme DNN-DP based on differential privacy.This scheme is mainly divided into two steps: the determination of adaptive noise and the training part of the noisy data.Firstly,we use the random forest method to evaluate the importance of features and then determine the adaptive noise added to each feature according to the importance of each feature and the range of the feature value.Secondly,we add adaptive noise to the first layer of the neural network to form the differential privacy affine transformation layer.Experiments on the US census data set show that the DNN-DP network proposed in this paper has better adaptability to classified data sets and can produce higher accuracy.Considering the fact,the DNN-DP algorithm mainly affects the boundary data,that is,the data at the classification boundary are permutated,which,in a sense,has positive effect on the classification accuracy of DNN-DP,but this operation has a large negative impact on the prediction of continuous values.Then we propose an algorithm called DP-SGD,which applies differential privacy to the gradient inside the neural network.In other words,instead of the noise simply being added into the raw dataset,this scheme goes deep inside the neural network.Specifically,this scheme crops the gradient first,and then adds the Laplacian noise to the cropped gradient.Finally,experiments on a continuous data set verify the effectiveness of the proposed scheme,which has a higher accuracy rate than the DNN-DP algorithm.