Cognitive Tracing Data Trails: Auditing Data Provenance in Discriminative Language Models Using Accumulated Discrepancy Score

The burgeoning practice of unauthorized acquisition and utilization of personal textual data (e.g., social media comments and search histories) by certain entities has become a discernible trend. To uphold data protection regulations such as the Asia–Pacific Privacy Initiative (APPI) and to identify instances of unpermitted exploitation of personal data, we propose a novel and efficient audit framework that helps users conduct cognitive analysis to determine if their textual data was used for data augmentation. and training a discriminative model. In particular, we focus on auditing models that use BERT as the backbone for discriminating text and are at the core of popular online services. We first propose an accumulated discrepancy score, which involves not only the response of the target model to the auditing sample but also the responses between pre-trained and finetuned models, to identify membership. We implement two types of audit methods (i.e., sample-level and user-level) according to our framework and conduct comprehensive experiments on two downstream applications to evaluate the performance. The experimental results demonstrate that our sample-level auditing achieves an AUC of 89.7% and an accuracy of 83%, whereas the user-level method can audit membership with an AUC of 89.7% and an accuracy of 88%. Additionally, we undertake an analysis of how augmentation methods impact auditing performance and expound upon the underlying reasons for these observations.