Do Influence-Derived Data Perturbations Enable Machine Unlearning? An Audit of Deep Perturbation Learning

Forum topic · 小凯 · 2026-09-15

Summary

This post summarizes a machine learning paper (arXiv 2609.12313) evaluating Deep Perturbation Learning (DPL), a method that perturbs training images and labels along influence-derived directions, across three machine unlearning roles: direct deletion signal, utility-preserving regularizer, and warm start for adversarial unlearning. The authors audit the public implementation and find two correctness issues: image directions are computed on augmented, normalized tensors but applied to raw images, and label perturbations fall below float32 resolution, leaving labels effectively unchanged. After correcting the image pipeline, DPL fails direct deletion criteria on CIFAR-10/ResNet-18 across three paired seeds, shows inconsistent utility effects across seeds, and underperforms a simple warm-start baseline once direction-computation time is counted. Single-seed Tiny ImageNet checks do not support DPL as a regularizer or warm start either. Results cover only random instance deletion and do not rule out influence methods in other unlearning regimes. The authors release a role-matched evaluation protocol and an audit checklist for perturbation-based unlearning claims.

Paper Overview

Field: Machine Learning Authors: Chenkai Wu, Chrispine Kambimbi, Qinyang Zeng, Jun Yan Date: 2026-09-15 arXiv: 2609.12313

Abstract

We evaluate Deep Perturbation Learning (DPL), which perturbs training images and labels along influence-derived directions, in three roles in which prior work has positioned it for machine unlearning: a direct deletion signal (the strongest claim), a utility-preserving regularizer, and a warm start for adversarial unlearning. Evidence for the weaker roles has been used to support the stronger one, so we test each role separately under a matched protocol with exact-seed retraining baselines.

Key Findings

Two correctness issues found in the public implementation:
Image directions are computed on augmented, normalized tensors but applied to raw images.
Label perturbation magnitudes fall below float32 resolution, leaving labels effectively unchanged.
After correcting the image-perturbation pipeline, DPL fails direct deletion criteria on CIFAR-10/ResNet-18 across three paired seeds.
Utility effects are inconsistent in sign across seeds, and DPL underperforms a simple warm-start baseline once direction-computation time is included.
Single-seed Tiny ImageNet checks do not support DPL as a regularizer or warm start; preprocessing inconsistencies in the released code prevent conclusive direct comparison.
Results cover random instance deletion only and do not rule out the effectiveness of influence-derived methods in other unlearning mechanisms.

The authors release a role-matched evaluation protocol and an audit checklist for perturbation-based unlearning claims.

---

*Auto-collected on 2026-09-15*