#Data Attribution

We consider the adversarial attack on training data attribution methods.

We developed a efficient library for data attribution, aiming to streamline the development of data attribution algorithms

We consider the most influential subset selection (MISS) problem and identify various failure modes of the greedy-based algorithm used by the community, and prove that being "adaptive" helps in a non-obvious way.

We design the first game-theoretic efficient machine unlearning evaluation metric with provable properties.