acceptodds
Under review as a conference paper at ICLR 2027

Decentralized Data-Free Unlearning in Continual Learning: Task-Conditioned Hypernetwork Approach with Theoretical Guarantee

Abstract

Unlearning in decentralized continual learning requires agents to unlearn previously acquired knowledge even when the corresponding historical data are unavailable. Performing unlearning without reconstructing historical inputs presents two challenges: (1) *Lack of Direct Unlearning Supervision*, as agents lack direct supervision for unlearning the target class while preserving the remaining classes of the same task; and (2) *Cross-Task Interference during Unlearning*, as local updates and neighbor aggregation may disrupt retained tasks through shared parameters, while broadly restricting parameter changes can hinder the requested unlearning. To address these challenges, we propose Dec-UnHNet, in which a hypernetwork generates task-specific classifier heads from descriptors built with diffusion-derived class prototypes. For a data-free unlearning request, surrogate features provide supervision for unlearning the target class, while hypernetwork-output regularization anchors generated classifiers for retained tasks without historical-sample replay. Theoretically, we rigorously prove that *Dec-UnHNet* achieves an convergence rate for the decentralized joint global empirical risk across all tasks. Extensive experiments on two benchmark datasets demonstrate improved continual-learning accuracy and zero target-class accuracy at all evaluated unlearning steps.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Decentralized learning, Hypernetwork, Continual learning, Unlearning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_GQqVOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies targeted class unlearning in decentralized continual learning when historical training data are unavailable. The authors propose Dec-UnHNet, which uses a task-descriptor-conditioned hypernetwork to generate task-specific classifiers, allowing selective modification of the target task with limited interference. The method combines diffusion-model-derived class prototypes with synthetic-feature supervision and regularizes generated classifiers to preserve unaffected tasks; a convergence analysis and results on two benchmarks are reported.

Critical strengths

(1) The approach is well motivated. Generating separate classifiers from task descriptors and regularizing outputs for unaffected tasks directly constrains the mappings that should be preserved, which is a sensible way to let the shared hypernetwork handle unlearning while limiting collateral changes.

(2) The problem setting is relevant, and the initial results are promising. Across two datasets and two heterogeneous partitions, the method retains relatively strong accuracy on retained classes while driving target-class accuracy to zero after the evaluated unlearning requests.

Critical weaknesses

(1) The convergence guarantee is not established by the current proof. In Appendix B.2, replacing the stochastic local update direction with the exact gradient stack and expanding T2T_2 are explicitly left unjustified, and the final bound does not show that the full stationarity measure tends to zero; this conflicts with Corollary 1 (Eq. 17). The theoretical claim needs a complete justification or substantial qualification.

(2) The evaluation mainly shows that classifier predictions change, not that unlearning is effective beyond suppression. Tables 1–2 use zero target-class accuracy as the main evidence, but this cannot distinguish the method from simpler target-class suppression; Section 4.3 also states that the feature extractor is unchanged and only the deployed predictive mapping is modified. More evidence is needed within that scope.

(3) The contributions of the proposed unlearning components are not isolated. Dec-UnHNet already outperforms several baselines before unlearning (e.g., 69.68% vs. 55.95% for UG-CLU under Dir(λ = 0.5) in Table 1), so retained-accuracy gains cannot be cleanly attributed to the unlearning mechanism. Figure 2 only varies regularization coefficients and does not isolate diffusion-based prototypes, task descriptor modification, or surrogate feature optimization.

(4) The experiments do not directly validate the motivating decentralized scenario. The introduction stresses agents that learn about the target class through communication without local samples, but Tables 1–2 report aggregate results under two Dirichlet partitions without separating agents that originally observed the target class from those that only acquired its knowledge through communication; Appendix C.2 (Table 3) varies topologies but does not test this distinction.

Questions for authors

(1) Provide a mathematically valid justification for replacing the stochastic local update direction with the exact gradient stack and for the T2T_2 expansion in Appendix B.2. With nonvanishing terms in the final bound, explain exactly how Corollary 1 (Eq. 17) follows; otherwise the convergence claim must be withdrawn or substantially weakened.

(2) The current evaluation does not rule out simple target-class suppression. Add a class-suppression baseline and/or a reference model retrained without target data, and show how Section 6.2 distinguishes genuine unlearning from suppression.

(3) Given the initial-performance gap in Tables 1–2, you cannot attribute retained-accuracy improvements to the unlearning mechanism without controls. Provide controlled experiments or component ablations that isolate diffusion-based prototypes, task descriptor modification, and surrogate feature optimization.

(4) Directly evaluate agents that never observed the target class locally but acquired its knowledge through communication. Report separate before/after unlearning results for these agents and for agents that originally observed the target class, including the Appendix C.2 topology settings.

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

This paper studies continual learning and continual unlearning in a decentralized setting, where a network of agents must learn new tasks while unlearning previously learned tasks due to privacy and data-governance requirements. In this setting, historical samples may no longer be available.

The paper identifies two key challenges in this setting: (1) lack of direct unlearning supervision, where unlearning requests specify the classes to be forgotten without providing direct supervision; and (2) cross-task interference during unlearning, where local unlearning operations may affect knowledge that should be retained across agents.

To address these challenges, the paper proposes Dec-UnHNet, whose central technical contribution is a hypernetwork that represents historical knowledge through a learned mapping rather than relying on replaying historical samples. The hypernetwork is conditioned on diffusion-based prototypes to update the classifier for learning and unlearning operations, respectively. Experiments demonstrate that the proposed method significantly outperforms existing continual learning methods on the CIFAR-100 and Tiny-ImageNet datasets.

Critical strengths

(1) This paper is clearly organized and generally well-written. Figure 1 clearly illustrates the overall design of the proposed network.

(2) The hypernetwork-based design avoids relying on historical samples for unlearning, which is useful.

(3) This paper provides a convergence analysis to establish theoretical guarantees for the proposed algorithm.

(4) The performance is significantly better compared with existing algorithms.

Critical weaknesses

Limitations of My Evaluation

I have not actively worked in this research area for several years; therefore, I would encourage the AC and SACs to place less weight on my evaluation and rely more heavily on reviewers with relevant expertise.

Weaknesses

(1) The motivation relies heavily on hypernetworks, but the paper lacks a brief introduction to or discussion of their use in continual learning. Although hypernetworks are well-established machine learning models, they are not as widely recognized as CNNs or Transformers, which can generally be introduced without extensive discussion. The authors are encouraged to provide more intuition behind the use of hypernetworks to make the paper more accessible to readers without prior knowledge of them.

(2) This paper uses a frozen pretrained diffusion model to obtain prototypes, but it does not specify which pretrained model is used, making it difficult for future researchers to reproduce the proposed method.

(3) Although I have not actively worked on continual learning for at least five years, the experiments appear to still rely on relatively small-scale datasets, such as CIFAR-100 and Tiny-ImageNet, and older architectures, such as ResNet-18 and VGG-16. The compared baselines also include EWC and LwF, which are relatively early continual learning methods. I would encourage the authors to consider larger-scale experiments, with additional transformer architecture experiments and comparisons with more recent continual learning algorithms.

Questions for authors

(1) Why is a pretrained diffusion model necessary for generating prototypes? Could a pretrained encoder, e.g., DINO, be used instead? I do not fully understand the intuition behind or the necessity of using a diffusion model here.

(2) I do not fully understand Equation (5). Why can it be approximated in this way? Could the authors provide further clarification?

(3) In Lines 205–208, how are the prototypes iteratively mixed with those of neighboring agents? The process is currently described only as "according to the communication matrices," but the exact operation is unclear to me. Regarding the task descriptor defined in Equation (3), it appears to be a weighted average of prototypes. Could the authors explain why this can be interpreted as a task descriptor?

Reviewer_G4NnOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper considers a problem that combines decentralized continual learning with class-level unlearning, where historical data is no longer available. The authors propose Dec-UnHNet, a task-conditioned hypernetwork that generates classifier heads when conditioned on diffusion-derived class prototypes. For a data-free unlearning request, the method uses surrogate features in order to provide supervision for unlearning the target class. The use of hypernetwork-output regularization limits cross-task interference without requiring replay samples. The paper then adds theoretical and empirical support to justify the proposed method. The core concern is limiting optimization during unlearning to only the classifier parameters. The paper needs significant rewriting to remove apparent contradictions and unresolved editorial notes scattered throughout. More importantly, the paper needs to justify the use of various mechanisms through a systematic ablation study. As things stand, the proposed problem formulation is promising but the paper is in need of substantial change to address the concerns laid out in the review.

Critical strengths

The problem setting is timely and well-motivated. Combining decentralized continual learning with data-free unlearning is not straightforward and presents a number of practical challenges. The paper’s reported retained classes accuracy numbers are strong. The proposed method outperforms the baselines in most scenarios while keeping the target class accuracy zero.

Critical weaknesses

The biggest weakness with this approach is that the feature extractor remains frozen during unlearning. Any unlearning is therefore a suppression of class mappings rather than genuine unlearning. Even after unlearning, the unlearned class may be recoverable via, say, the construction of class prototypes in a few-shot manner. The paper is, however, honest about this and states this plainly. The paper could benefit from Membership Inference Attack (MIA) evaluations. This would help establish whether the unlearned model retains information about the forgotten class samples. While the zero target-class accuracy is encouraging, it does not necessarily imply successful removal of the target data's influence. Relevant prior work is not discussed. UnCLe (Adhikari et al., arXiv 2509.17530, 2025) seems to be the most related work, considering that it too focuses on continual learning and unlearning, that too involving hypernetworks as a central mechanism. Particularly, in Appendix C3, the authors seem to do something along the lines of what UnCLe does, which is reverting the learned task-specific parameters to random weights while retaining other task mappings. Of course, the unlearning scope is different. UnCLe seems to be a task-level unlearning method whereas the proposed method is class-level unlearning. Regardless, I feel that the paper would benefit from appropriate context. The method introduces various components such as diffusion-derived prototypes, prototype aggregation, task descriptors, surrogate-feature synthesis, mapping-alignment and output regularization at once without accompanying it with a systematic ablation that shows the individual contributions of the various mechanisms involved. This makes it hard to identify which of these mechanisms are crucial. For instance, why do we need expensive diffusion-derived prototypes? Could regular learned task embedding suffice? The paper, therefore, opens up a lot of such questions that are not sufficiently addressed. Lines like the following scattered around the paper appear to be editorial notes that have not been properly resolved. I await the authors’ clarification on the following: “The first inequality above replaces the stochastic local update direction by the exact agent-gradient stack without displaying the required error bridge; we retain the line as part of the provisional source derivation rather than supplying an unproved step.” “The ensuing T2 bound is therefore provisional pending an author-supplied estimator identity.” “UniCLUN is the label in the experimental record, but the available sources do not establish its bibliographic or implementation mapping. The records separately name a local CLMUL implementation with a 200-sample buffer per agent and do not establish that it maps to tabled UniCLUN.” “Embedded legacy notation … ” in Figure 1’s caption.

Questions for authors

The paper does not state how the task is identified during inference. Is a task identifier provided, or is it inferred by some means? The abstract states that the paper rigorously proves a convergence rate. This is contradicted in Appendix B, wherein it is stated, “The full endpoint remains a residual-neighborhood bound … Consequently, this derivation does not establish that the full stationarity measure converges to zero.” How do the authors reconcile this statement with the claim in the abstract? Under what additional assumptions, if any, do the residual terms vanish? The convergence analysis defines a task-level supervised empirical risk but the actual unlearning procedure optimizes a different objective, which is classifier-parameter alignment with output regularization. Hence, it is unclear whether the theoretical results apply to the unlearning procedure itself or only to the continual-learning component.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.