Differentially Private Auctions for Private Data Crowdsourcing

The success of machine learning and deep learning depends on the large amount of for training. Data acquisition is an indispensable event for data science. In this paper, we study a scenario where a data broker wantes to buy private data from a population to estimate some statistics and resell the results to the data users in order to obtain profits. It's necessary to compensate owners of private data for their loss of privacy. In our setting, we assume that sellers are selfish agents who would like to maximize his or her utilities, thus we need to design truthful mechanisms. In our work, we will study a more generalized optimization goal for budget free crowdsoucer, profit maximization. The utility of the buyer is a function of the accuracy of the expected statistic. We will use a Profit-Extract algorithm for the auction, and a Laplace mechanism will be used for privacy protection. Our proposed designed mechanism has properties of individual rationality, truthfulness, computational efficiency and also has a performance lowerbound. Furthermore, we will design an online auction for our setting.