Latent Dirichlet Allocation Model Training With Differential Privacy

Latent Dirichlet Allocation (LDA) is a popular topic modeling technique for hidden semantic discovery of text data and serves as fundamental tool analysis in various applications. However, the LDA model well training process may expose information data, thus bringing significant privacy concerns. To address issue LDA, we systematically investigate protection main-stream algorithm based on Collapsed Gibbs Sampling (CGS) propose several differentially private algorithms typical scenarios. In particular, present first theoretical inherent differential guarantee CGS further centralized privacy-preserving (HDP-LDA) that can prevent inference from intermediate statistics training. Also, locally (LP-LDA) crowdsourced to provide local individual contributors. Furthermore, extend LP-LDA an online version OLP-LDA achieve mini-batches streaming setting. Extensive experiment results validate both effectiveness efficiency our proposed algorithms.