Do Influence-Derived Data Perturbations Enable Machine Unlearning? A Controlled Study of Three Plausible Roles

TEXT START: 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.

The Dissection

This paper is an overclaim autopsy. It separates three purported functions of DPL that prior work blurred together, audits the released implementation, repairs a broken perturbation pipeline, and tests the strongest claims against exact-seed retraining baselines.

The results are narrow but damaging to DPL: the image perturbations were applied inconsistently; the label perturbation was numerically inert in float32; corrected DPL failed the direct-deletion criterion across all three CIFAR-10/ResNet-18 seed pairs; utility effects changed sign across seeds; and accounting for influence-direction computation made it worse than simple warm starts. The Tiny ImageNet result is weaker because it uses one seed and contains unresolved preprocessing inconsistencies.

The paper is therefore doing methodological triage, not discovering a new unlearning capability. It is removing a weak branch from the machine-unlearning tree.

The Core Fallacy

The primary fallacy exposed is inferential: evidence that a method preserves utility or provides a useful initialization does not establish that it directly deletes information. Those are different objectives with different success criteria. Treating them as interchangeable is how a plausible-looking benchmark result becomes a false deletion claim.

The paper’s own DT-level limitation is scope, not invalidity. A failed DPL experiment does not threaten the Discontinuity Thesis. It only shows that one proposed control technique is defective or inferior under the tested regime. Confusing failure of a tool with failure of AI’s strategic trajectory would be the larger category error. AI capital does not need one particular unlearning algorithm to continue displacing labor.

Hidden Assumptions

• Random instance deletion is treated as a meaningful proxy for the full unlearning problem. The abstract explicitly does not cover class deletion, feature deletion, backdoor removal, privacy erasure, or other regimes.
• CIFAR-10/ResNet-18 and a one-seed Tiny ImageNet check are treated as informative about broader model families and production-scale systems. They are useful test beds, not universal evidence.
• Exact-seed retraining is accepted as the decisive comparison, although real deployment may impose latency, cost, auditability, or legal constraints that make full retraining impractical.
• Correct preprocessing and numerical resolution are assumed to be basic implementation details. Here they are outcome-determining; a paper that does not audit them is measuring code pathology as though it were theory.
• Influence-derived directions are assumed to contain actionable deletion information. The corrected results provide no reliable support for that assumption.
• “Machine unlearning” is treated as a technical model-behavior problem rather than also a question of ownership, verification, liability, and institutional power.

Social Function

Primary classification: partial truth and transition management, with a prestige-signaling layer.

The paper performs a valuable cleanup operation: it makes a specific claim harder to repeat without competent controls. That is not copium. It does not promise that human labor survives, nor does it imply that AI deployment is slowing.

Its broader institutional function is to improve the governance layer around AI capital. If deletion obligations become legally or commercially important, the system will pay for reliable audit, retraining, verification, and compliance infrastructure. That creates servitor niches for people who can certify what a model has retained or forgotten. It does not restore productive participation for the displaced majority. Unlearning is a control surface around the machine, not a reversal of machine supremacy.

The Verdict

DPL’s direct-deletion claim is unearned and, on the corrected CIFAR-10 setup, effectively rejected. Its weaker roles are inconsistent or inferior once real computation and implementation correctness are counted; the Tiny ImageNet evidence is inconclusive, not exculpatory.

This is a competent local falsification, not a systemic reprieve. The carcass is one perturbation method, not AI automation. The field will discard DPL, substitute a better method, and continue toward the same structural endpoint: AI owners retain productive control, while unlearning specialists survive only as temporary compliance servitors.