A Workload Division Differential Privacy Algorithm to Improve the Accuracy for Linear Computations.

Differential privacy algorithm is an effective technology to protect data privacy, and there are many pieces of research about differential privacy and some practical applications from the Internet companies, such as Apple and Google, etc. By differential privacy technology, the data organizations can allow external data scientists to explore their sensitive datasets, and the data owners can be ensured provable privacy guarantees meanwhile. It is inevitable that the query results that will cause the error, as a consequence that the differential privacy algorithm would disturb the data, and some differential privacy algorithms are aimed to reduce the introduced noise. However, those algorithms just adopt to the simple or relative uniform data, when the data distribution is complex, some algorithms will lose efficiency. In this paper, we propose a new simple \(\varepsilon \)-differential privacy algorithm. Our approach includes two key points: Firstly, we used Laplace-based noise to disturb answer to reduce the error of the linear computation queries under intensive data items by workload-aware noise; Secondly, we propose an optimized workload division method. We divide the queries recursively to reduce the added noise, which can reduce computation error when there exists query hot spot in the workload. We conduct extensive evaluation over six real-world datasets to examine the performance of our approach. The experimental results show that our approach can reduce nearly 40% computation error for linear computation when compared with MWEM, DAWA, and Identity. Meanwhile, our approach can achieve better response time to answer the query cases compared with the start-of-the-art algorithms.