We consider the adversarial attack on training data attribution methods.
We provide a comprehensive study of the common practices in the Most Influential Subset Selection (MISS) problem.
We design the first efficient machine unlearning evaluation metric with provable guarantees.
