The research evaluated Deep Perturbation Learning (DPL) across three proposed roles for machine unlearning: direct deletion, utility preservation, and warm start for adversarial unlearning. The study employed a matched protocol with exact-seed retraining baselines to assess each role’s effectiveness. Initial implementation issues, including incorrect image direction computation and label perturbation resolution, were identified and corrected. After these corrections, DPL failed to meet the direct-deletion criterion when tested on CIFAR-10 with a ResNet-18 model across three different seed values. The experiments utilized Tiny ImageNet for a one-seed check, but DPL did not demonstrate a clear advantage as a regularizer or warm start. The analysis considered direction-computation time, and DPL underperformed simple warm-start baselines once this factor was included. The research focused solely on random instance deletion and did not evaluate influence-based methods in other deletion regimes. A role-matched evaluation protocol and an audit checklist for perturbation-based deletion claims were released.
Specifically, the image direction computations were performed on augmented, normalized tensors but applied to raw images. The label perturbation resolution was below float32, resulting in unchanged labels. The experiments used CIFAR-10 and ResNet-18 models. The study did not provide quantitative benchmarks or performance metrics beyond the failure to meet the direct-deletion criterion. The research aimed to provide a rigorous assessment of DPL’s potential for machine unlearning based on influence-derived data perturbations.
This work highlights potential challenges in deploying influence-based methods for data deletion. The findings suggest that simple, direct methods may remain more reliable for achieving data removal in machine learning models. Further research is needed to explore alternative unlearning strategies and address the identified implementation issues.
Source: https://arxiv.org/abs/2609.12313
