GRIN+: Towards Fast Yet Effective Machine Unlearning for Imbalanced Medical Data
Minghui Huang, Junxiao Wang
Timely, practical applied contribution with open code and broad baselines, but incremental novelty, an over-claimed central result contradicted by its own BUSI results, and narrow evaluation cap its expected impact.
Abstract
As deep learning models become fundamental to modern healthcare, the "Right to be Forgotten" mandated by privacy regulations like GDPR and HIPAA necessitates effective machine unlearning (MU) to remove sensitive patient data from trained models. However, existing MU techniques often struggle with a fundamental "privacy-efficiency-utility" (PEU) trilemma, particularly in medical scenarios where data is frequently characterized by severe class imbalance and long-tailed distributions. In such cases, standard unlearning methods can fail to protect key clinical knowledge or mistakenly delete features essential for diagnosing rare conditions due to the gradient dominance of majority classes. To address these challenges, we propose GRIN+, a novel machine unlearning framework designed for fast and precise data erasure in imbalanced medical scenarios. GRIN+ decouples unlearning-specific knowledge from generalized representations at the parameter level by analyzing the gradient contributions of both "forget" and "retain" sets. It introduces a class-adaptive influence scoring mechanism to rectify gradient dominance and employs a direction-constrained update strategy to prevent the unintended erosion of vital clinical knowledge. Comprehensive benchmarking across multiple medical datasets, including skin cancer (ISIC), brain tumor (MRI), and breast ultrasound (BUSI), demonstrates that GRIN+ achieves an optimal balance of the PEU trilemma. Experimental results show that GRIN+ maintains high diagnostic accuracy and robust privacy while significantly enhancing runtime efficiency compared to existing baselines. We open-source the GRIN+ code and benchmarks to support further research.
AI Impact Assessments
(1 model)Scientific Impact Assessment
Core Contribution. The paper addresses machine unlearning (MU) in medical imaging under severe class imbalance. Its stated contributions are three-fold: (1) framing a "privacy-efficiency-utility" (PEU) trilemma for imbalanced medical unlearning; (2) proposing GRIN+, a gradient-guided selective forgetting framework combining (a) a gradient-ratio influence score (large gradient on forget set / small on retain set), (b) a class-adaptive weight to counter majority-class gradient dominance, and (c) a direction-constrained perturbation that suppresses updates conflicting with retain-set gradients; and (3) a benchmark across three medical datasets (ISIC, MRI, BUSI) with open-sourced code. The central insight — that standard influence-function and Fisher-based importance estimates are biased by class imbalance, harming rare-disease knowledge — is a reasonable and clinically relevant observation.
Methodological Rigor. The design is competent but not exemplary. The theoretical "motivation" (Section 4.4) is a standard first-order Taylor/Lagrangian derivation showing the update direction should follow G_f − G_r; it is sound but elementary and does not constitute a novel theoretical result. There is a notable internal inconsistency: Eq. (6) defines the direction-constraint factor using `cos(·)` between gradient components, while Algorithm 1 line 17 uses `sign(·)` — these are not equivalent, and the discrepancy signals insufficient editorial care that undermines confidence in the exact implementation. The experimental setup is otherwise reasonable: 10 baselines spanning classical, SOTA general-purpose, and medical-specific methods; seven metrics covering the three PEU axes; a proper retrain gold standard; fixed seeds; and an ablation isolating each of the three components. The ablation is the most convincing part, cleanly showing CW recovers R-Acc (+16.7pp) and DCF restores MIA toward 0.5.
Evidence Strength. The evidence is mixed and the "optimal PEU balance" claim is overstated. On MRI, GRIN+ genuinely looks strong (R-Acc 84.13, MIA 49.90, highest RTE 17.24). On ISIC it is competitive. But on BUSI, GRIN+ posts R-Acc 58.34 vs. Retrain 86.11 and a RetDev of 94.23 — worse than several baselines (e.g., BiO at 5.85 RetDev with 83.27 R-Acc). This directly contradicts the narrative that GRIN+ dominates the trilemma; on the smallest dataset it substantially degrades utility. Error bars (SEM) are reported, which is good, but no significance testing is provided, and given large variances (e.g., ±11–14 on several MRI baselines) many differences may not be statistically meaningful. The privacy metric (MIA near 0.5) is favorable, but the utility trade-off is dataset-dependent in a way the discussion glosses over.
Potential Impact. The problem — regulatory-driven data deletion (GDPR/HIPAA "right to be forgotten") in clinical ML — is real and growing, and imbalance is a genuine and underexplored wrinkle. The method is lightweight (sparse parameter updates, no Hessian, no teacher model), which is an attractive practical property. However, the contribution is an incremental combination of well-established ideas: gradient-magnitude importance scoring, sparse/masked updates, inverse-frequency class weighting, and gradient-conflict-based direction constraints all exist in prior unlearning and long-tailed learning literature. The novelty lies mainly in assembling these for the medical-imbalance niche rather than in a new primitive. The released code and three-dataset benchmark add modest reusable value to a fragmented subfield (the related work itself notes medical unlearning research is "limited and somewhat fragmented").
Timeliness & Relevance. Highly timely — MU under privacy law is an active area, and long-tailed medical distributions are a recognized bottleneck (CXR-LT benchmarks cited). The paper sits squarely in a current concern, which raises its citation prospects even if the technical depth is modest.
Strengths.
Limitations.
Overall. This is a solid, practically-oriented applied-ML paper addressing a timely niche with a reasonable method and useful open resources, but with incremental novelty, an over-claimed central result contradicted by its own BUSI numbers, and limited evaluation breadth. It is likely to be cited within the medical-unlearning subfield as a useful baseline/benchmark contribution rather than to reshape practice.
Generated Sep 15, 2026
Comparison History (0)
Timely, practical applied contribution with open code and broad baselines, but incremental novelty, an over-claimed central result contradicted by its own BUSI results, and narrow evaluation cap its expected impact.