acceptodds
Withdrawn from ICLR 2027

Relearning Resistance Under Adam: A Gradient-Entropy Certificate for Machine Unlearning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Relearning provides a direct test of whether machine unlearning has removed information or merely suppressed its immediate behavior. Relearning Convergence Delay () quantifies this resistance by accumulating excess forget-set error along the relearning trajectory. Existing theory focuses on gradient descent, despite the widespread use of adaptive optimizers such as Adam for training modern models. We develop a framework for analyzing under Adam and derive an upper bound involving three quantities at the unlearning initialization: the forget-set loss gap, the local Hessian condition number, and Adam's second-moment condition number. Because large models are often fine-tuned by updating only a restricted subset of parameters, we further show that the expected certificate can be bounded in terms of the entropy and scale of the forget gradient. This analysis motivates Gradient-Entropy Maximization (GEM), a baseline-agnostic regularizer that promotes both gradient dispersion and magnitude. We establish an convergence rate to an -stationary point for the GEM objective and evaluate it empirically. Across classification, diffusion, and language-model unlearning, GEM can improve resistance to Adam relearning while largely preserving retain performance and forgetting quality.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
machine unlearning, relearning attack, optimization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_zaBoOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies relearning resistance in machine unlearning when forgotten data are relearned using Adam rather than standard gradient descent. It derives an upper bound on Relearning Convergence Delay (RCD) under a frozen Adam preconditioner and uses this result to motivate Gradient-Entropy Maximization (GEM), a regularizer designed to spread the forget-set gradient across model parameters and make relearning more difficult. Experiments on CIFAR-100 classification, Stable Diffusion concept removal, and TOFU language-model unlearning show that GEM can improve RCD for several methods, although the gains vary across baselines, tasks, and relearning learning rates.

Critical strengths
  • This paper studies a meaningful and relatively underexplored question: whether unlearned knowledge remains difficult to recover when relearning is performed with Adam.

  • GEM can be easily combined with existing unlearning methods, and the experiments cover image classification, image generation, and language-model unlearning with several representative baselines.

Critical weaknesses
  • The claimed relearning resistance is too closely tied to Adam. The evaluation assumes that relearning follows an Adam optimization trajectory, but an attacker can choose SGD, AdamW, related benign data, or a different recovery procedure. Prior work has shown that a small amount of related but non-identical data can reactivate supposedly forgotten knowledge [a]. It is therefore unclear whether GEM makes the target knowledge genuinely harder to recover or only makes a particular Adam trajectory converge more slowly.

The theoretical analysis further narrows this setting by assuming β1=0\beta_1=0, β2=1\beta_2=1, and a preconditioner frozen at the beginning of relearning. It also requires a locally positive-definite Hessian, a monotonic condition-number assumption, and a step size determined by the largest Hessian eigenvalue at each iteration. These assumptions remove the momentum and evolving second-moment estimate used by practical Adam and AdamW, leaving a substantial gap between the theory and the optimizers used in the experiments.

  • The gradient-entropy score and RCD do not provide a certificate that the target knowledge has been reliably removed. Theorems 3.1 and 3.4 only establish upper bounds on RCD and the relevant preconditioning term. Increasing such an upper bound does not imply that the actual RCD must increase. This gap is also visible in Appendix D.4, where the entropy score and measured RCD move in the same direction in only 5 of 10 comparisons for the 30% forgetting settings, with similarly partial agreement under Random 50%.

More fundamentally, a large RCD may result from poor conditioning or generally slower optimization rather than the absence of the target information from the model. Prior work has recovered apparently forgotten knowledge through benign relearning [a], broader behavioral and privacy tests [b], and probes of internal representations [c]. The paper therefore needs independent recovery attacks and residual-knowledge evaluations before interpreting a larger Adam RCD as stronger unlearning; otherwise, the term “certificate” overstates the guarantee.

[a] Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning. ICLR 2025.

[b] MUSE: Machine Unlearning Six-Way Evaluation for Language Models. ICLR 2025.

[c] An Illusion of Unlearning? Assessing Machine Unlearning Through Internal Representations. AISTATS 2026.

Questions for authors
  1. Does GEM improve relearning resistance when the attacker uses SGD with momentum, AdamW, or a different recovery procedure?

  2. Since Theorems 3.1 and 3.4 provide only upper bounds, what formal argument connects maximizing the gradient-entropy score to an increase in the actual RCD? How should the proposed “certificate” be interpreted given the partial disagreement between the score and RCD in Appendix D.4?

  3. Can the analysis be extended to practical Adam with nonzero momentum and an evolving second-moment estimate, or can the authors provide evidence that the frozen-preconditioner approximation accurately predicts the behavior observed in the experiments?

Reviewer_KPDSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies resistance to relearning after machine unlearning under Adam. It derives a relearning convergence delay (RCD) upper bound for a frozen-preconditioner specialization of Adam and relates the expected preconditioner condition number to gradient entropy and magnitude under random coordinate masks. This motivates Gradient-Entropy Maximization (GEM), an additive regularizer for existing unlearning objectives. Experiments cover image classification, diffusion models, and language models.

Critical strengths
  1. Evaluating recovery under adaptive optimizers addresses an important limitation of studying unlearning only through immediate post-unlearning behavior or gradient-descent relearning.
  2. The paper evaluates GEM across three domains and multiple baselines, with paired comparisons, learning-rate sweeps, and utility measurements.
Critical weaknesses
  1. A larger RCD upper bound can simply be looser without increasing actual RCD. This leaves a substantial gap between the theoretical analysis and the proposed optimization objective.
  2. Setting (beta1beta_1=0) and freezing the second moment yields fixed diagonal preconditioned gradient descent.
  3. The regularizer combines gradient entropy and magnitude, but lacks entropy-only and magnitude-only ablations.
  4. The theoretical bound uses loss-based RCD, whereas classification and language-model experiments use accuracy- and probability-based errors. Matched initial forgetting, loss-based trajectories, and stronger adaptive relearning attacks would help establish whether GEM slows recovery beyond increasing initial error.
Questions for authors

see weakness

Reviewer_yCFvOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper develops two loosely coupled threads of ideas impacting machine unlearning: 1) an analysis of the recently proposed Relearning Convergence Delay (RCD) metric for the special case of Adam optimizer; 2) motivated by this analysis, proposing a new regularizer -- Gradient-Entropy Maximization (GEM) -- promoting both gradient dispersion and magnitude.

RCD was introduced by Tran & Woo [1], in a paper that established a theory for RCD (RCD and its finite-horizon approximation, bound with forget-loss gap, convergence toward an error floor via unlearning perturbations) and also proposes Influence Eliminating Unlearning (IEU) to remove the influence of the forgetting set by degrading its performance and incorporating weight decay and noise-injection into the model’s weights. The analysis of RCD in Tran & Woo is based on gradient descent, which motivates the study of RCD under Adam.

Therefore the authors extend the analysis of RCD for a Adam, following similar arguments presented in [1] to obtain new but similar bounds with an added factors or terms (e.g. pre-conditioner condition number). The proposed regularizer GEM target to bound the expected pre-conditioner condition number using the entropy and scale of the forget gradient. The experimental evaluation is broad, covering classification, diffusion and language-model (TOFU) based unlearning benchmark. For classification the empirical gains are clear, while for diffusion and TOFU are less consistent.

[1] Tran, K; Woo, SS. Efficient Unlearning through Maximizing Relearning Convergence Delay. Findings of CVPR 2026.

Critical strengths
  1. The idea of novel mechanisms to perturb the gradients dynamics to improve RCD seems intuitively interesting, and particularly using some dispersion metric like entropy or variance.

  2. The empirical evaluation is broad covering classification, diffusion and text generation. Matched baselines with and without the regularizer are evaluated in this broad benchmark.

Critical weaknesses

I present two groups of weakeness one conceptual/methodological, another mathematical/algorithmic.

A) Methodological

  1. The empirical protocol is insufficient to establish a reliable improvement. The paper explains that IEU targets the forget-loss gap and Hessian conditioning, while GEM targets the third factor. The obvious natural experiment is therefore IEU versus IEU+GEM under matched GD and Adam relearning. Furthermore the Adam theory is a natural extension of the GD theory of RCD from Tran and Woo, which should motivate a similar matched comparison with IEU as the first line of evidence.

  2. The optimized score in GEM is S=eHgmax⁡iGi2S=e^{H_g}\max_i G_i^2 which mixes dispersion/entropy with gradient scale, so even when GEM improves RCD is not easy to tell if entropy or gradient scale is responsible for that. This also touches another issue that is not clear that entropy or simply variance of the gradients is the key innovative factor (item B.1 points the mathematical reasons, given that we could have better bounds on moments of the gradients without multiplying the entropy term).

B) Mathematical

  1. The theoretical justification for GEM is insufficient. Theorem 3.4 provides an upper bound on the preconditioner condition number, subsequently used to upper-bound RCD. Maximizing this bound does not imply increasing actual relearning resistance. Moreover, the entropy-dependent bound is looser than a direct bound in terms of the gradient norm (which is already presented in Lemma A.6 but made loose by multiplying the entropy), making the role of entropy theoretically questionable.

  2. The Adam analysis does not cover standard Adam. Theorem 3.1 assumes β1=0\beta_1=0 and a frozen second-moment preconditioner β2=1\beta_2=1, effectively analyzing preconditioned gradient descent rather than Adam. Consequently, the theoretical guarantees do not directly apply to the adaptive optimizer used in the experiments.

  3. The stationarity complexity is incorrectly stated. Theorem 4.1 establishes an O(1/K)O(1/K) bound on the squared gradient norm, implying O(1/ε2)O(1/\varepsilon^2), not O(1/ε)O(1/\varepsilon), iterations to obtain ∥∇Jβ∥≤ε\|\nabla J_\beta\|\leq\varepsilon. Furthermore, the required smoothness assumptions are not established for the gradient-entropy regularizer.

Questions for authors
  • How does GEM compare with existing explicitly relearning-resistant approaches, such as IEU? Could you provide a comparison of IEU versus IEU+GEM, with entropy-only and gradient-scale-only controls?
  • How would the analysis change if we consider a typical adam setup (nonzero momentum, an evolving second-moment estimate)?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.