A Survey of Machine Unlearning

Published venue: ACM Transactions on Intelligent Systems and Technology (TIST), 2025 (per dblp) Version actually read: arXiv:2209.02299 v6, 17 September 2024 — the preprint, in ACM format with a placeholder ISBN line ("978-x-xxxx-xxxx-x/YY/MM"). Affiliations: Griffith University; EPFL; University of Bremen; Hanoi University of Science and Technology; University of Queensland.

⚠️ Version discrepancy — recorded, not glossed

dblp records the published article as ACM TIST 2025. The PDF fetched from arXiv is v6, dated September 2024. These are not necessarily the same text: a journal version may add, cut or revise content during review.

Consequence: any page, section or figure reference taken from this file refers to the arXiv v6 preprint, and must not be cited as though it came from the TIST article. If the review ever needs to cite this survey with a locator, the published version must be obtained first. papers.yaml records both.

Read status: title page, author list, abstract and §1 opening read; §§2–13 skimmed only for structural orientation. Deliberately not read in depth — see below.

1. Why this file is short, on purpose

This is the one entry in papers.yaml categorised survey, and the project's standing rule for it is recorded in that entry:

"Use for orientation and for finding primary sources only. No claim in LITERATURE_REVIEW.md may cite this survey in place of the primary paper."

Reading a survey deeply and then citing it would import its authors' characterisations of ~200 papers into our review without any of them being checked. This project has already found, by checking:

• five mutually inconsistent attributions for one dataset (VERIFICATION_REPORT.md §5b),
• an (\varepsilon,\delta) definition widely attributed to Making AI Forget You: Data Deletion in Machine Learning whose main-body definition is exact (ginart2019deletion.md §4),
• a "naive retraining" baseline in Amnesiac Machine Learning that is actually finetuning,
• and a citation-key pair in Towards Unbounded Machine Unlearning that would misattribute the forging result if swapped.

Every one of those would have been silently inherited from a secondary source. So the survey is used as an index, not as evidence.

2. What it is, and what it is legitimately good for

The abstract's framing: models "often 'remember' the old data", contemporary attacks "have proven that we can learn whether an instance or an attribute belonged to the training data", and recent work "has not been able to completely solve the problem due to the lack of common frameworks and resources."

Legitimate uses for this project:

1. Coverage check. Once LITERATURE_REVIEW.md is drafted, comparing its reference list against this survey's is a cheap way to find whole sub-areas the review has missed. That is a use where being secondhand does not matter.
2. Locating primary sources in corners the reading list does not reach — e.g. its treatment of statistical query learning, which is the framework underpinning Towards Making Systems Forget with Machine Unlearning.
3. Confirming that a claim is field-wide rather than one paper's view, before then citing the primary sources for it.

3. What it must not be used for

• Any statement of what a specific paper says, proves, assumes or reports.
• Any number, bound, definition or equation.
• Any characterisation of a method's performance relative to another's.
• Establishing a paper's venue, year or authorship — papers.yaml uses dblp and fetched pages for that.

4. Links

• Indexes essentially every primary source already in papers.yaml.
• Its statistical-query-learning discussion is the framework behind Towards Making Systems Forget with Machine Unlearning.
• Not a substitute for Are we making progress in unlearning? Findings from the first NeurIPS unlearning competition, which is a findings report with its own methodology rather than a secondary summary.

@article{nguyen2025survey,
  title         = {A Survey of Machine Unlearning},
  author        = {Thanh Tam Nguyen and Thanh Trung Huynh and Zhao Ren and Phi Le Nguyen and Alan Wee-Chung Liew and Hongzhi Yin and Quoc Viet Hung Nguyen},
  year          = {2025},
  journal       = {ACM Transactions on Intelligent Systems and Technology (TIST)},
  eprint        = {2209.02299},
  archiveprefix = {arXiv},
  url           = {https://arxiv.org/abs/2209.02299}
}