The Measure of Deception: An Analysis of Data Forging in Machine Unlearning
The paper analyzes data forging, where an adversary tries to create a sequence of training batches that mimics the gradient trajectory of a target point to falsely claim unlearning occurred. It rigorously proves that for linear regression and smooth loss functions, the set of such deceptive trajectories has a small Lebesgue measure, meaning they are geometrically rare.