p21 / Preprint
Evaluating the Defense Potential of Machine Unlearning against Membership Inference Attacks
arXiv:2508.16150 · 2025
In plain language
The study asks whether targeted machine unlearning can reduce privacy leakage from trained models.
Research context
Removing data influence needs evaluation on retained users as well as the users requesting deletion.
Approach
Negative Gradient, SCRUB, and SFTC are evaluated on four datasets and three model complexity groups against membership inference attacks.
Findings
Protection depends on the algorithm, model capacity, and learning rate. SFTC can increase membership signals for retained data, while SCRUB gives a more balanced defense in the reported experiments.
Summary based on abstract. Summary and author order follow arXiv version 4, revised 22 January 2026. First submitted in 2025.