GRIN+: Towards Fast Yet Effective Machine Unlearning for Imbalanced Medical Data
cs.AI
Quick verdict
Gradient-ratio one-shot unlearning for imbalanced medical data: strong on MRI, but BUSI (RetDev 94.2, R-Acc 58.3%) contradicts the claimed PEU balance.
Summary
The paper tackles machine unlearning — removing a patient's data influence from a trained model — in medical imaging, where severe class imbalance breaks methods that implicitly assume balanced data. The authors frame a privacy-efficiency-utility (PEU) trilemma and propose GRIN+, a one-shot approximate unlearning method for ResNet18 classifiers. GRIN+ scores each parameter by the ratio of forget-set to retain-set gradient magnitudes, reweights scores with inverse-frequency class weights, selects the top p% parameters, and applies a single direction-constrained perturbation along the gradient difference; a first-order Lagrangian argument motivates the update direction. On ISIC skin lesions (2,357 images, 9 classes), Kaggle brain MRI (7,023, 4 classes), and BUSI ultrasound (780, 3 classes), GRIN+ reports MIA near chance (49.90% MRI, 52.92% ISIC, 50.45% BUSI), retains utility close to retraining on MRI and is competitive on ISIC, and is fastest on MRI (RTE 17.24 vs retrain = 1). Ablations on MRI credit class weighting (+16.73pp R-Acc) and the direction constraint (+2.19pp). Main limitations: on BUSI the method over-forgets (RetDev 94.23; R-Acc 58.34 vs retrain 86.11), where the BiO baseline dominates on utility and speed; the direction-constraint factor as specified is mathematically degenerate (sign-based cosine yields a constant gate); statistics are SEM-only with unstated run counts and no significance tests; and several definitions (epsilon reference set, Indisc metric, per-class weight application) are inconsistent or unspecified.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Class-balance assumptions in approximate unlearning fail on long-tailed medical data, causing majority-class gradient dominance and misdirected parameter updates that can erase rare-disease features (Sec 1, Fig 1).
- GRIN+ decouples forget-specific parameters via gradient-ratio scoring with inverse-frequency class weights, then applies a single direction-constrained perturbation to the top-p% parameters (Sec 4, Algorithm 1).
- GRIN+ drives MIA to near chance (49.90% MRI, 52.92% ISIC, 50.45% BUSI) while retaining utility close to retraining on MRI/ISIC and being the fastest method on MRI (RTE 17.24) (Table 2).
- The ablation (Table 3, MRI) reports R-Acc 65.21 with gradient-ratio selection alone, 81.94 after adding class weighting, and 84.13 with the full method, implying each component is necessary.
- The code and benchmark setup are open-sourced to support reproducibility.
Strengths
Clinically grounded problem framing
The PEU trilemma is tied to a concrete failure mode — majority-class gradient dominance erasing rare-disease features — and Figure 1 concretely contrasts CIFAR-10's uniform 10% per class with ISIC's 3.4-20.6% long-tailed distribution.
Genuinely cheap one-shot update
The pipeline (Algorithm 1) needs two full-dataset gradient passes, one sort and one masked perturbation — no iterative optimization — yielding the best RTE of all 11 methods on MRI (17.24 vs retrain = 1, Table 2).
Broad baseline sweep
Component ablation and code release
Weaknesses
BUSI results contradict the headline claim
Direction-constraint factor is degenerate as published
Efficiency superiority is MRI-only but claimed generically
Statistics too weak for the comparisons drawn
Reproducibility gaps despite released code
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
GRIN+ (arXiv:2609.15571, formatted for Knowledge-Based Systems) is a method paper on approximate machine unlearning (MU) for medical image classifiers trained on class-imbalanced data. The authors claim three contributions: (1) identification of a 'privacy-efficiency-utility' (PEU) trilemma specific to imbalanced medical data; (2) a framework that scores parameters by forget/retain gradient ratios, corrects class imbalance via inverse-frequency weights, and applies a one-shot, direction-constrained perturbation to a sparse (top-p%) parameter subset; (3) benchmarking on three public medical datasets showing MIA near chance and the best efficiency on MRI. The motivation is the GDPR/HIPAA 'Right to be Forgotten', and code is released (github.com/gzhu-hcai/Med-Unlearn).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is GRIN+ and what problem does it solve?
GRIN+ is a machine unlearning method: it removes the influence of specific patient images from an already-trained medical image classifier without retraining from scratch. It targets imbalanced medical datasets, where standard unlearning can accidentally degrade recognition of rare diseases. It works by scoring parameters using the ratio of gradients from the data to forget versus the data to keep, then nudging only the top ~10-20% of parameters once, in a constrained direction. The authors say this makes deletion fast while preserving accuracy and privacy.
Does GRIN+ actually preserve model accuracy?
Partially. On brain MRI it stays close to the retrained gold standard (84.13% retain accuracy vs 83.57% for retraining) and is the fastest method there (17.24x faster than retraining). But on the BUSI breast-ultrasound dataset it over-forgets: retain accuracy drops to 58.34% versus 86.11% for retraining, and the BiO baseline is both more accurate (83.27%) and faster in the paper's own Table 2. Results on ISIC are mixed but competitive.
Is the GRIN+ code and data available?
The authors state the code is open-sourced at github.com/gzhu-hcai/Med-Unlearn, and all three datasets are public: ISIC skin-lesion images (2,357 images, 9 classes), a Kaggle brain-tumor MRI set (7,023 images, 4 classes), and BUSI breast ultrasound (780 images, 3 classes). Experiments used ResNet18 on a single RTX 4090 GPU. This review did not independently verify the repository or the results.
What is the PEU trilemma in machine unlearning?
It is the authors' term for the three-way tension every unlearning method faces: privacy (the forgotten data must no longer be recoverable, e.g., against membership inference attacks), efficiency (removal must be much faster than retraining from scratch), and utility (the model must keep its accuracy on the remaining data). The paper argues all three must hold simultaneously on imbalanced medical data, and that most existing methods satisfy at most two of the three.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-2 are only partially conveyed by the text extraction; the RTE/R-Acc bubble comparison in Figure 1 could not be inspected.
- Mathematical symbols are partially garbled in extraction (dataset symbols, script letters); equations were reconstructed from context, while Tables 2-3 extracted cleanly.
- Only the 10-page main text is present; any supplementary material (e.g., hyperparameter sensitivity studies) may exist and was not reviewed.
- The claimed open-source repository (github.com/gzhu-hcai/Med-Unlearn) was not accessed or verified.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.