Towards Differential Relational Privacy and its use in Question Answering

srikar appalaraju

Abstract: Memorization of the relation between entities in a dataset can lead to privacy issues when us- ing a trained model for question answering. We introduce Relational Memorization (RM) to un- derstand, quantify and control this phenomenon. While bounding general memorization can have detrimental effects on the performance of a trained model, bounding RM does not prevent effective learning. The difference is most pronounced when the data distribution is long-tailed, with many queries having only few training examples: Im- peding general memorization prevents effective learning, while impeding only relational memo- rization still allows learning general properties of the underlying concepts. We formalize the no- tion of Relational Privacy (RP) and, inspired by Differential Privacy (DP), we provide a possible definition of Differential Relational Privacy (DrP). These notions can be used to describe and com- pute bounds on the amount of RM in a trained model. We illustrate Relational Privacy concepts in experiments with large-scale models for Ques- tion Answering.