MC-GEN:Multi-level Clustering for Private Synthetic Data Generation

Nowadays, machine learning is one of the most common technology to turn raw data into useful information in scientific and industrial processes. The performance of the machine learning model often depends on the size of dataset. Companies and research institutes usually share or exchange their data to avoid data scarcity. However, sharing original datasets that contain private information can cause privacy leakage. Utilizing synthetic datasets which have similar characteristics as a substitute is one of the solutions to avoid the privacy issue. Differential privacy provides a strong privacy guarantee to protect the individual data records which contain sensitive information. We propose MC-GEN, a privacy-preserving synthetic data generation method under differential privacy guarantee for multiple classification tasks. MC-GEN builds differentially private generative models on the multi-level clustered data to generate synthetic datasets. Our method also reduced the noise introduced from differential privacy to improve the utility. In experimental evaluation, we evaluated the parameter effect of MC-GEN and compared MC-GEN with three existing methods. Our results showed that MC-GEN can achieve significant effectiveness under certain privacy guarantees on multiple classification tasks.