Corpus graph

Disagreements (16)

Every critique edge in full, with whether we checked the claim. The graph shows where they are; this says what they say.

• verified kurmanji2023scrub criticises golatkar2020eternal

  Relies on assumptions deep networks violate -- SGD stability and small forget sets -- and its forgetting step scales quadratically with training set size.

• plausible kurmanji2023scrub criticises thudi2022unrolling

  The regulariser may hurt model performance, and single gradient unlearning removes only the effect the forget set had in the first training iteration.

• plausible bourtoule2021sisa criticises cao2015forget

  Not general and does not scale: deep networks use adaptive statistical queries, where the approach diverges in an unbounded fashion.

• verified bourtoule2021sisa criticises guo2020certified

  Effective for simple models such as linear regression, so falls short of being model-agnostic.

• plausible bourtoule2021sisa criticises ginart2019deletion

  Unclear whether Quantized k-Means is general enough and low-overhead enough to meet the stated goals.

• verified xu2025infotheoretic criticises bourtoule2021sisa

  Anchor-based definitions depend on an unknown retrained model and are therefore not auditable from the released predictor alone.

• verified neel2021descent criticises bourtoule2021sisa

  SISA is more ambitious in aiming at non-convex models, but uses a weaker, possibilistic deletion criterion -- that the post-deletion model could merely have arisen from retraining -- and gives no error guarantees; it also retrains affected partitions in full, uses no randomisation, and omits the reservoir-sampling step their indistinguishability guarantee depends on.

• verified thudi2022auditable criticises bourtoule2021sisa

  Its definition permits plausible deniability via a training algorithm as evidence, and forging constructs exactly that evidence without doing the work.

• unchecked goel2022adversarial criticises chundawat2023badteacher

  Methods that do not use retain-set information will have limitations on arbitrary deletions and will not perform well on the Interclass Confusion test.

• verified goel2022adversarial criticises golatkar2020eternal

  Could not be run at CIFAR scale due to large memory requirements.

• plausible foster2024ssd criticises golatkar2020eternal

  Fisher Forgetting is computationally very expensive and updates the whole model, which degrades retain-set accuracy; SSD cites Tarun et al.'s benchmarks as the evidence rather than its own.

• verified foster2024ssd criticises graves2021amnesiac

  Storing a list of samples and parameter updates for every training batch adds significant overhead and cannot be performed post hoc.

• verified foster2024ssd criticises ginart2019deletion

  SSD characterises Ginart et al.'s definition as probabilistic, requiring the unlearnt output distribution to be SIMILAR to a never-trained one; Ginart's Definition 3.1 is equality in distribution, i.e. exact. Recorded as a further instance of the exact-to-approximate drift documented in ginart2019deletion.md.

• verified fan2024salun criticises jia2023sparsity

  l1-sparse appears strongest when judged on unlearning accuracy alone, but that strength costs retain and test accuracy.

• verified hayes2024inexact criticises kurmanji2023scrub

  SCRUB's basic MIA is a population U-MIA using no shadow models, the weakest category, so it substantially overestimates the privacy protection achieved.

• verified hayes2024inexact criticises jia2023sparsity

  Retain-set privacy worsens after unlearning on over half the datapoints not selected for forgetting.