Research On Key Techniques Of Differentially Private Crowdsourcing Service

Crowdsourcing service is a new type of task execution mode,which can outsource tasks originally performed by specific professionals to nonspecific workers in a free and voluntary manner.It then pools swarm intelligence through efficient collaboration of large-scale swarm intelligence resources,thus provides strong support for problem solving.According to the life cycle of crowdsourcing service,it can be divided into three stages:crowdsourcing preparation stage,crowdsourcing ongoing stage and service output stage.Among them,the key of the preparation stage is worker recruitment,the core of the ongoing stage is task allocation,and the foundation of the service stage is truth value discovery.In crowdsourcing service,if participants worry about disclosing their personal sensitive information during participation,their enthusiasm for participating in crowdsourcing will be greatly reduced,which will seriously affect the service quality of crowdsourcing service.Privacy disclosure has become the main bottleneck restricting the further development of crowdsourcing.Differential Privacy(DP)technology proposed in recent years has become the primary means to protect Privacy.In particular,it includes Geo-indistinguishability(Geo-I)to protect location privacy and Local Differential Privacy(LDP)to protect numerical privacy.However,the existing researches on crowdsourcing scenarios or related technologies are still in its infancy.Specifically,in terms of crowdsourcing scenarios,the existing research work has just begun to explore the problem of worker recruitment satisfying Geo-I and the problem of truth discovery satisfying LDP.In terms of related technologies,existing research work often only meets the weak version of DP,resulting in insufficient privacy protection.To this end,this thesis focuses on privacy protection in the whole life cycle of crowdsourcing services,conducts indepth research on the basis of strictly meeting DP,and achieves the following research results:(1)For the problem of area coverage-based worker recruitment under geo-indistinguishability,this thesis proposes a Geo-indiStinguishable ArEa Coverage-based WorkeR rEcruitmenT,namely SECRET.Existing research work potentially assumes that the obfuscated locations of the participants are still in the area to be perceived,and they do not consider the impact of the participants’ perception radii on the area coverage.To this end,SECRET combines the perception range of participants to generate obfuscated locations and determines the collection of recruited workers according to their spatial relationships.In SECRET,to protect each participant’s location,this thesis develops an optimized geographical exponential mechanism with solid privacy and utility guarantees.To select the recruited workers based on the obfuscated locations while ensuring large coverage for the target region,this thesis designs a coverage-aware worker selection method.This thesis shows SECRET satisfies ε-geoindistinguishability.This thesis further gives its utility guarantee.Experiments are carried out on two real-world datasets and one synthetic data set.The experimental results show that SECRET can ensure the high spatial coverage of the recruited participants while guaranteeing privacy.(2)For the problem of task allocation under geo-indistinguishability,this thesis proposes a Task alloCAtioN approach via grOup-based noisE Addition,namely CANOE.Existing research work potentially assumes that workers can be assigned arbitrary tasks,and the noise injected for privacy protection is unbounded and random.To this end,in CANOE,each worker uploads the noisy distances between his true location and the obfuscated locations of his preferred tasks instead of uploading his obfuscated location.In particular,this thesis uses one location to represent a group of geographically close locations.This indicates only one piece of noise will be added to the above locations in the same group.In particular,to alleviate the total noise when conducting grouping,this thesis puts forward an optimized global grouping with adaptive local adjustment method with convergence guarantee.In addition,this thesis formulates a mixed-integer nonlinear program problem with a non-convex constraint.Due to the NP-hard characteristic,this thesis devises a solving solution based on benders decomposition and alternating direction method of multipliers.To collect the noisy distances that are required for task allocation,this thesis develops a utility-aware obfuscated distance collection method with solid privacy and utility guarantees.It further theoretically analyzes the privacy and utility guarantees of CANOE.Experiments are carried out on two real-world datasets.The experimental results show that CANOE can guarantee the average travel distance of the task assignment results while guaranteeing privacy.(3)For the problem of truth discovery while achieving the rigorous LDP for each worker with continuous inputs without the independence assumption,this thesis presents a Locally Differentially Private Truth Discovery approach via Sampling and Inference,namely PrivTDSI.Existing research work potentially assumes that the tasks are independent for each other.Therefore,it is necessary to allocate a separate privacy budget for each task,resulting in high overall noise injection.To this end,in PrivTDSI,the server samples a small portion of the raw values under local differential privacy and infers the unsampled values for the truth discovery scenario.In particular,to determine the sample proportion,this thesis formulates a constrained nonlinear programming problem and give a closed-form solution to this problem.Moreover,to avoid that some workers are not sampled or some tasks are not sampled during the sampling process,which may reduce the utility of the final truth,a two-stage sampling algorithm is designed.Furthermore,to infer the unsampled values accurately,this thesis designs a quality-aware inference method based on matrix factorization.This thesis further gives its privacy,utility and complexity guarantees.Experiments were carried out on two realworld datasets.The experimental results show that PrivTDSI can provide high effectiveness of the noisy truths while guaranteeing privacy.