When a person asks for their data to be removed from a trained AI model, the ideal approach is to retrain the model from scratch without that data. This is too expensive for most real-world systems, so approximate methods are used instead. However, these methods only check how well the model performs on average across a dataset, which can hide failures for individual data points.
The authors introduce SALA, a tool that works with any existing approximate unlearning method. It steers each data sample toward a specific target: for samples that should stay in the model, it moves them toward what a member sample looks like; for samples that should be forgotten, it moves them toward a non-member sample. To make this practical, they also develop SMSE, a cheap way to estimate what those member and non-member targets should be for each sample.
When tested on classification and generation tasks, SALA reduced the per-sample gap compared to full retraining by up to 86.67%, while also making the unlearning process more robust to poor settings of hyperparameters. This means data erasure can be both efficient and reliably enforced at the level of individual records.