Typed decision models often do more than pick a category. They also output probabilities for each candidate and, in some cases, a separate confidence score. These extra fields are useful for downstream decisions, but they may also leak information about the underlying data or the model's training set.
The new paper introduces TUPLE-Bench, a benchmark designed to evaluate and mitigate the privacy-utility trade-offs that arise from these outputs. By focusing on typed decision models, the work addresses a common but under-examined source of potential exposure: the auxiliary numbers that accompany a prediction.
The benchmark appears to provide a systematic way to measure how much information these fields reveal, and to test strategies for reducing leakage while preserving utility. The abstract notes that the extra fields support decisions but may reveal information, which is the core tension the benchmark aims to quantify and address.