WARP: Improving the Attack Resistance of Machine Unlearning Protocols with Weight Teleportation

The researchers propose a plug-and-play defense method called WARP, which uses neural network symmetry to reduce the gradient energy of the forget set and increase parameter dispersion, thereby mitigating the privacy leakage caused by approximate machine unlearning while maintaining the accuracy on retained data. The paper first points out that the leakage stems from two facts: the large gradient norms of forget-set samples and the closeness of the parameters after unlearning to the original model, and proposes membership inference and data reconstruction attacks targeting unlearning, showing that existing methods such as NGP and SCRUB remain vulnerable to attack. Across six unlearning algorithms, the method reduces the adversarial advantage (AUC) by up to 64% in the black-box setting and by up to 92% in the white-box setting. The authors therefore consider weight teleportation a general means of reducing the attack success rate of approximate unlearning.