AI News and Headlines

NDSS 2025 – Provably Unlearnable Data Examples

The NDSS 2025 paper tackles the lack of provable guarantees in machine‑unlearning by introducing a certification framework for (q, η)-learnability. It proposes Provably Unlearnable Examples (PUEs) that tighten certification bounds and resist simple weight‑recovery attacks. Experiments show PUEs cut certified learnability up to 18.9% on ImageNet and lower empirical test accuracy by 54.4% on CIFAR‑100, outperforming existing unlearnable examples. This work bridges the gap between heuristic defenses and measurable privacy protection for shared data.