Decision unlearning performance varies with deletion and label changes
DecisionUnlearn-Bench: Unlearning Probabilistic Decisions under Changing Schemas
Cryptography and SecurityArtificial Intelligence
Summary
Machine unlearning tries to make a model act as if it never saw certain data. This paper introduces a way to compare how well different methods remove learned information based on what options the model sees when making decisions. The authors show that changing the labels or which choices are available during testing can greatly affect how well the unlearning works. They also find that different unlearning methods perform differently depending on many factors like model type and what data is removed. Their results highlight the importance of evaluating unlearning in ways that consider the specific decision setting.
What this means in practice
- •For machine learning engineers: Assess and compare unlearning methods under different decision scenarios for sensitive data removal in deployed models.
- •For data privacy teams: Improve protocols for reliable deletion of user data in AI systems by understanding how candidate options affect unlearning.
Authors
Tiantong Wu, Wei Yang Bryan Lim
Abstract
Machine unlearning aims to reproduce the behavior of a model trained without designated data. For probabilistic decision models, this target also depends on the candidate options supplied at inference. We introduce DecisionUnlearn-Bench, an empirical framework that compares full, unlearned, and retain-only models across deletion granularities, label renderings, and candidate sets. Across 20 eligible model configurations on the BANKING77 dataset, full-to-oracle displacement increases consistently from example to association to class deletion. Replacing semantic option labels with opaque codes nearly eliminates the retrained oracle's recognition of deleted classes, even with successful acquisition of the full model, whereas changing the candidate set alters both the oracle and the measured unlearning progress. SimNPO satisfies the screening criterion in 19 of 20 configurations, but its gains vary substantially across seeds and deletion targets. Development tuning on the CLINC150 dataset does not establish a positive pooled effect across 48 intents. At approximately matched suppression, methods can differ markedly in oracle fidelity. Recovery experiments further distinguish initial forgetting from subsequent adaptation and show conclusions that depend on both architecture and metric. These results show that evaluating decision-model unlearning relative to a schema-specific retraining target, with separate reporting of effect size, utility, robustness, and recoverability.