|
Abstract:
|
With the widespread applications of big data, privacy protection becomes critical for users. Differential privacy is a particular data privacy development where the data distribution is no longer sensitive to changes of data points from original data. This also brings challenges to preserve original statistical information and analysis results using release data. Under the framework of differential privacy protection, we propose a differential private data release algorithm based on latent factor model in that the proposed algorithm is able to add less noise to the data under the same level of privacy protection. Our algorithm can be applied for both continuous data and categorical data. Specifically, we build a latent factor model with a data matrix as an input matrix where categorical data can also be transformed to continuous data. Based on the information rate and Laplace mechanism, we obtain a privacy-preserving coefficient matrix by adding weighted noises. Consequently, we can privately release a set of data records. Theoretically, we show that the proposed algorithm achieves the differential privacy requirement. Our numerical studies demonstrate the superb performance on uti
|
