Algorithmic Bias in De-Identification Tools
Paul M. Heider · 2023
We present a series of experiments designed to quantify the algorithmic bias inherent in six off-the-shelf de-identification systems. We used false negative rate (FNR) and true positive rate (TPR) parity measures in addition to F1-score to evaluate the systems across two environments. In the first condition, we inferred gender and race/ethnicity labels for a pre-existing de-identification corpus and analyzed performance on disaggregated subgroups. In the second condition, we controlled the dominant race/ethnicity bias for names used as realistic surrogates for resynthesizing a different pre-existing de-identification corpus. In both conditions, we found cases of strong bias and near-zero bias.