Assessing Unintended Memorization in Neural Discriminative Sequence Models

Despite their success in a multitude of tasks, neural models trained on natural language have been shown to memorize the intricacies of their training data, posing a potential privacy threat. In this work, we propose a metric to quantify unintended memorization in neural discriminative sequence models. The proposed metric, named d-exposure (discriminative exposure), utilizes language ambiguity and classification confidence to elicit the model’s propensity to memorization. Through experimental work on a named entity recognition task, we show the validity of d-exposure to measure memorization. In addition, we show that d-exposure is not a measure of overfitting as it does not increase when the model overfits.