This post on zhichai.net summarizes a machine learning paper (arXiv: 2609.12313) by Chenkai Wu, Chrispine Kambimbi, Qinyang Zeng, and Jun Yan, published 2026-09-15.
Key points
- What is evaluated: Deep Perturbation Learning (DPL), which perturbs training images and labels along influence-derived directions, in three roles prior work positioned it for machine unlearning: 1. A direct deletion signal (the strongest claim) 2. A utility-preserving regularizer 3. A warm start for adversarial unlearning
- Motivation: Evidence for the weaker roles had been used to support the stronger one, so the authors test each role separately under a matched protocol with exact-seed retraining baselines.
- Implementation audit findings (two correctness issues in the public code):
- Image directions are computed on augmented, normalized tensors but applied to raw images.
- Label perturbation magnitude falls below float32 resolution, so labels are effectively unchanged.
- Main results after correcting the image-perturbation pipeline:
- DPL fails the direct deletion criterion on CIFAR-10/ResNet-18 across three paired seeds.
- Its utility impact is inconsistent in sign across seeds.
- Once direction-computation time is counted, DPL underperforms a simple warm-start baseline.
- Single-seed Tiny ImageNet checks do not support DPL as a regularizer or warm start either; preprocessing inconsistencies in the released code prevent definitive direct comparison.
- Scope caveat: Results cover random instance deletion only and do not rule out the effectiveness of influence-based methods in other unlearning regimes.
- Artifacts: The authors release a role-matched evaluation protocol and an audit checklist for perturbation-based unlearning claims.