Privacy-preserving Active Learning on Sensitive Data for User Intent Classification
This addresses privacy concerns for users in active learning systems, but appears incremental as it applies existing privacy techniques to a specific domain.
The paper tackles the privacy risk in active learning when sensitive user data is sent to annotators, proposing an approach with quantifiable privacy guarantees and evaluating the tradeoff between privacy, utility, and annotation budget on a binary classification task.
Active learning holds promise of significantly reducing data annotation costs while maintaining reasonable model performance. However, it requires sending data to annotators for labeling. This presents a possible privacy leak when the training set includes sensitive user data. In this paper, we describe an approach for carrying out privacy preserving active learning with quantifiable guarantees. We evaluate our approach by showing the tradeoff between privacy, utility and annotation budget on a binary classification task in a active learning setting.