GDPR — Right to Erasure (Article 17)

Why

The legal driver behind machine unlearning. This note is only the parts of the regulation the
thesis needs for motivation, not a legal analysis.

Summary

Source. Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016
(General Data Protection Regulation), Official Journal of the European Union L 119, 4.5.2016.
88 pages in the PDF.

Article 4(1) — personal data

“Personal data” means any information relating to an identified or identifiable natural person
(“data subject”); an identifiable person is one who can be identified, directly or indirectly, in
particular by reference to an identifier such as a name, an identification number, location data,
an online identifier, or to one or more factors specific to the physical, physiological, genetic,
mental, economic, cultural or social identity of that person.

Article 17 — Right to erasure (“right to be forgotten”)

Article 17(1). The data subject has the right to obtain from the controller the erasure of
personal data concerning them without undue delay, and the controller has the obligation to erase
personal data without undue delay where one of the following grounds applies:

  • (a) the data are no longer necessary for the purposes for which they were collected or processed;
  • (b) the data subject withdraws consent on which the processing is based, and there is no other
    legal ground;
  • (c) the data subject objects to the processing under Article 21(1) and there are no overriding
    legitimate grounds, or objects under Article 21(2);
  • (d) the data have been unlawfully processed;
  • (e) erasure is required by a legal obligation in Union or Member State law;
  • (f) the data were collected in relation to the offer of information society services under
    Article 8(1).

Article 17(2). Where the controller has made the data public and must erase them, it must take
reasonable steps, considering available technology and cost, to inform other controllers processing
the data that the data subject has requested erasure of links to, copies of, or replications of
those data.

Article 17(3). Paragraphs 1 and 2 do not apply where processing is necessary for:

  • (a) exercising the right of freedom of expression and information;
  • (b) compliance with a legal obligation, or performance of a task in the public interest / official
    authority;
  • (c) public interest in public health under Article 9(2)(h)–(i) and Article 9(3);
  • (d) archiving purposes in the public interest, scientific or historical research purposes, or
    statistical purposes under Article 89(1), insofar as the right would likely render impossible or
    seriously impair the achievement of those purposes;
  • (e) the establishment, exercise or defence of legal claims.

Reading for the thesis

  • The regulation speaks of personal data, not arbitrary graphs. The thesis’s synthetic ER/BA/WS
    graphs are not personal data; the legal framing is motivational only. A deployed model trained on
    user graphs would fall under Article 17.
  • Article 17(3)(d) is the scientific-research exception: a research use may be exempt when
    erasure would seriously impair the research. This is relevant if the thesis discusses deployment,
    but not for the experiments.
  • The regulation talks about erasure of data, not erasure from model weights. Machine
    unlearning is the technical interpretation of what “erasure” should mean for a trained model —
    this gap is exactly the thesis’s motivation.

Thesis hook

  • One sentence in the introduction: GDPR Article 17 creates the right, retraining is the only
    exact way to honour it, retraining is too expensive, so approximate unlearning is needed.
  • The “right to be forgotten” also explains why verification matters: a controller must show
    that erasure happened, which is why MIA-based evaluation is central.
  • The scientific-research exception is a caveat for the ethics/legal section, not a result.

Open questions

  • Does the thesis need to address the “personal data” boundary, given that its graphs are
    synthetic? Probably only in the introduction and ethics section.
  • If a model is trained on a graph where a node represents a person, does deleting the node suffice,
    or does the edge structure also count as personal data? This is the graph-unlearning analogue of
    item vs feature removal.