Exploiting Multi-Model Collaborative Inference for Privacy Enhancement in Text Classification

Text classification is a foundational task in natural language processing that involves categorizing raw text into pre-defined classes. This task holds significant importance in various applications, including but not limited to sentiment analysis and intent detection. With collaborative inference of multiple models, text classification may achieve an improved performance compared to the single model. However, if multiple models have access to the input text directly, it may create challenges on the privacy of sensitive data or model information. It is not easy to realize collaborative inference while preserving the privacy. This paper presents PPJP, a privacy-preserving joint system that helps achieve private collaborative inference in text classification with machine learning. Our method to instantiate it, is based on secure multiparty computation (MPC) and differential privacy (DP). We fulfill the privacy and scalability of text classification under multiple models inference. Secret-sharing-based MPC is used to protect the input and model parameters, while DP is used to protect against membership inference attack. We implement and evaluate prototype of our PPJP system based on the Twitter dataset. Experimental results show that text classification can guarantee privacy for model owners and clients with 54% inference accuracy. It achieves a balance between privacy and accuracy in case of collaborative inference.