acceptodds
Under review as a conference paper at ICLR 2027

Theory of Certified Unlearning for a Family of Regularization-Based Continual Learning

Abstract

Integrating certified unlearning into continual learning, an emerging paradigm termed continual learning-unlearning (CLU), enables models to learn sequentially while retaining past knowledge and reliably removing requested information to protect privacy. Despite its importance, CLU remains largely unexplored. We provide the first unified -certified CLU framework for a broad family of regularization-based CL algorithms, with the objective of jointly minimizing CL excess risk and unlearning loss. We derive an algorithm-dependent CL excess-risk bound for nonconvex losses, showing that regularizers that better accommodate task drift, such as Hessian-based regularization, improve learning performance. We then develop two complementary certified-unlearning mechanisms. Our first-order mechanism exploits natural forgetting induced by CL updates with minimal additional cost. Our second-order method enables precise request-specific unlearning for arbitrary requests, with unlearning loss strongly depending on request order through prior-correction propagation and regularizer mismatch. Both methods show that CL algorithms with stronger regularization, such as MAS, achieve lower unlearning loss, revealing a fundamental trade-off between retaining undeleted-task knowledge and removing deleted-task influence. Empirical results reveal distinct Pareto frontiers between certified-unlearning loss and CL excess risk, showing that no single method minimizes both. Thus, method selection should prioritize lower unlearning loss for stringent privacy requirements and lower CL excess risk for better predictive performance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Certified unlearning, continual learning, continual unlearning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_2U54OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies certified continual learning–unlearning for regularization-based continual learning. It derives an algorithm-dependent excess-risk bound, develops a natural-forgetting procedure and a Hessian-based correction procedure, and analyzes the effects of regularization and deletion order. Experiments with a frozen ResNet-18 feature extractor and a trainable head on task-partitioned CIFAR-100 compare approximation error, accuracy, runtime, and storage.

Critical strengths
  1. The paper addresses the important tension between maintaining previously learned tasks and removing the influence of deleted tasks, rather than treating unlearning as an isolated post-training operation.
  2. The unified treatment of task-additive regularizers and the comparison between natural forgetting and second-order corrections provide a useful conceptual account of how regularization and deletion order interact.
  3. The evaluation considers several regularizers, early versus late deletions, and multiple requests, and reports both predictive quality and resource usage.
Critical weaknesses
  1. Limited effectiveness of natural forgetting for targeted unlearning. The natural forgetting method in Eq. (9) performs no explicit correction to remove the requested task's influence. Instead, it relies on subsequent continual-learning updates and calibrated Gaussian noise. Its practical effectiveness remains unclear, particularly for recently learned tasks or deletion requests arriving at the final learning stage, where few or no subsequent updates are available. The noise required to achieve certified unlearning may also significantly affect retained-task utility.
  2. Insufficient experimental validation. The experiments mainly focus on CIFAR-100 with a frozen feature extractor and a trainable classification head. More diverse datasets, model architectures, and continual-learning settings are needed. More importantly, the paper lacks direct empirical evaluations of unlearning effectiveness, even basic membership inference attack (MIA) experiments. Existing works with theoretical unlearning guarantees still conduct MIA evaluations to validate forgetting effectiveness. More recent methods, such as Unlearning Evaluation through Subset Statistical Independence (ICLR 2026), also provide alternative ways to assess unlearning. The current evaluation, which primarily relies on parameter approximation errors and predictive accuracy, is insufficient to fully assess practical forgetting effectiveness.
Questions for authors
  1. How effective is natural forgetting for targeted task-level unlearning, particularly when a deletion request concerns a recently learned task or arrives at the final learning stage? Since the method performs no explicit correction, how is the required Gaussian noise calibrated, and what is its impact on retained-task utility?
  2. Can the authors provide more comprehensive evaluations across different continual-learning benchmarks, model architectures, and deletion schedules? In particular, can they include MIA-based evaluations or other suitable empirical unlearning metrics, alongside retained-task accuracy, to provide more direct evidence of forgetting effectiveness?
6 comments
AuthorsOpenReview ↗

Response to Reviwer 2U54 (1/3)

Thank you for your valuable suggestions and comments. We address your questions and provide clarifications point by point below.

1. Response on W1 and Q1.

First, we clarify that our main contribution is not natural forgetting but Hessian-based unlearning in Section 5. We use natural forgetting as a simple benchmark for comparison with Hessian-based unlearning and to enhance it, yielding forgetting-enhanced Hessian unlearning in Appendix F.

Natural forgetting indeed has limitations for targeted unlearning and can perform poorly when deleting recent tasks. However, it requires no additional request-time model update, providing a particularly simple unlearning option. Also, since forgetting is an inherent property of continual learning, it is natural and well motivated to investigate how this effect can be exploited to remove deleted-task influence. Moreover, natural forgetting mirrors gradient-based methods in offline unlearning that continue training on retained tasks, supporting the motivation for studying natural forgetting in CL (Koloskova et al., 2025).

Also, we further introduce natural forgetting as a complementary unlearning option, broadening the framework’s applicability across different computational budgets and deletion scenarios. Our theoretical analysis in Theorem 4.1 explicitly characterizes its unlearning performance and provides a basis for choosing between natural forgetting and Hessian-based unlearning by comparing their theoretical guarantees in Theorems 4.1 and 5.2 and their operational costs. Motivated by the simplicity of natural forgetting, we also propose a forgetting-enhanced Hessian unlearning method in Appendix F for choice to balance operational cost and utility.

Our experiments also explicitly illustrate the limitations and practical effectiveness of natural forgetting. Table 1 shows that natural forgetting performs well when deleting early tasks, identifying a regime in which it can be useful. In other regimes, we primarily use Hessian-based unlearning or forgetting-enhanced Hessian unlearning. In Table 2 of the revised manuscript, we also show its benefit in terms of low storage cost. Figures 3–6 on page 10 of the revised manuscript also show that natural forgetting reduces membership leakage while retaining accuracy, although it performs worse than Hessian-based unlearning.

To calibrate noise for natural forgetting, Theorem 4.1 explicitly establishes an upper bound on the approximation error ∣wt−S1:t−wt−S≤t∣\\|w_t^{-\mathbf{S}{1:t}}-w_t^{-S{\le t}}\\| between the natural-forgetting model wt−S1:tw_t^{-\mathbf{S}{1:t}} and the retrained model. The noise standard deviation is proportional to this proven upper bound, following the standard certified-unlearning procedure introduced in Section 3.2.

2. Response on W2 and Q2.

Following your suggestion, we have significantly expanded our experiments to include additional benchmarks, datasets, and model architectures, as well as MIA-based evaluations. Specifically, we add experiments on CIFAR-10 using a convolutional neural network trained from scratch, comprising three convolutional blocks with 6464, 128128, and 256256 output channels, respectively, and 2,473,6102{,}473{,}610 trainable parameters in total. CIFAR-10 is a commonly used benchmark in continual learning experiments (Xie et al., 2025, Structural-Entropy-Based Sample Selection for Efficient and Effective Learning, ICLR).

New MIA evaluation. we have added MIA-based evaluations in the new Figures 3–6. Specifically, we evaluate membership leakage using a standard modified-entropy MIA and LiRA (Carlini et al., 2022), a state-of-the-art attack that uses shadow models to calibrate a likelihood-ratio test for membership inference. We also tested the evaluation methods proposed in Unlearning Evaluation through Subset Statistical Independence (ICLR 2026), but they did not yield informative results in our setting and are therefore not included. We compare four model states: the checkpoint immediately after learning the task to be deleted, natural forgetting (NF), Hessian unlearning (HU), and complete retraining with the deleted task excluded. Figure 3 shows that HU substantially reduces membership leakage for the deleted task, achieving modified-entropy AUCs of 0.510 and 0.498 on CIFAR-10 and CIFAR-100, respectively, close to retraining (0.509 and 0.501). Figures 4 and 5 show consistent LiRA results: AUC decreases from 0.791 and 0.690 immediately after learning to 0.511 and 0.494 after HU, versus 0.536 and 0.563 for NF. With smaller empirical approximation errors, HU requires less perturbation and achieves higher test accuracies than NF (56.53%/77.00% versus 49.56%/76.06%; Figure 6), approaching retraining performance on CIFAR-100. Although NF provides weaker protection than HU, it still reduces AUC relative to the just-learned checkpoint, demonstrating its effectiveness. Overall, HU provides stronger protection against the evaluated MIAs while preserving higher test accuracy.

AuthorsOpenReview ↗

Response to Reviwer 2U54 (2/3)

New Unlearning sequence evaluation. We evaluated unlearning performance under various request sequences in Appendix G.1 of the original manuscript. As suggested, we have moved these results to the main text, as shown in Figure 2 on page 10 of the revised manuscript.

On CIFAR-100, we evaluate HU and NF under three request sequences over a 20-task stream, all of which eventually delete tasks {4,5,8,10,11,13,14}\{4,5,8,10,11,13,14\}. Here, Q1=(7:{4,5}; 12:{8,10,11}; 17:{13,14})Q_1=(7:\{4,5\};\,12:\{8,10,11\};\,17:\{13,14\}) is forward-synchronous and satisfies Eq. (16), whereas the other two sequences are asynchronous. HU achieves a lower final approximation error than NF across all three sequences, demonstrating its effectiveness and robustness under complex multi-request schedules. The forward-synchronous sequence achieves the lowest final approximation error, consistent with our theoretical analysis in Theorems 5.2 and E.1.

We have also added corresponding experiments using the CNN on CIFAR-10, reported in Figure 8 of Appendix H.1 on page 56 in the revised manuscript. These experiments consider a 10-task stream with three schedules:

Q_1 = (4:\{2\};\,7:\{5,6\};\,9:\{8\}), Q_2 = (4:\{2\};\,7:\{5\};\,9:\{6,8\}), Q_3 = (7:\{5,6\};\,9:\{2,8\}). \end{aligned}$$ Here, $Q_1$ is forward-synchronous, while $Q_2$ and $Q_3$ are asynchronous. EWC on CIFAR-10 exhibits the same schedule ordering as EWC on CIFAR-100. At $t=9$, when all schedules have deleted $\{2,5,6,8\}$, HU achieves approximation errors of $0.31779$, $0.32247$, and $0.33198$ under $Q_1$, $Q_2$, and $Q_3$, respectively. After learning the final retained task at $t=10$, these errors decrease to $0.31048$, $0.31359$, and $0.31613$, preserving the ordering. Thus, the forward-synchronous schedule also achieves the lowest final error in this setting. **New evaluation of the unlearning–accuracy tradeoff on CIFAR-10.** Table 1(b) on page 9 of the revised manuscript presents our newly added CIFAR-10 evaluations, comparing NF and HU across EWC, MAS, Diag-Hessian, scalar $\ell_2$, and accumulated $\ell_2$ regularization. We use five tasks and issue a single deletion request at $t=5$, separately deleting Task 2 or Task 4. When predictive accuracy is prioritized, we should choose MAS for both deletion settings. Its parameter-dependent penalties constrain important directions while allowing adaptation elsewhere, consistent with Section 3. When approximation quality is prioritized, we should choose HU with Diag-Hessian for Task 2 and HU with accumulated $\ell_2$ for Task 4. This dependence on the deletion regime is consistent with Sections 4–5, where stronger effective regularization can attenuate the deleted task’s influence. All reported Pareto-optimal configurations use HU, indicating that explicit correction is beneficial in the evaluated five-task setting. **New evaluation on unlearning performance along CLU trajectory for CIFAR-10.** We added the experiment result of unlearning approximation error along CLU trajectory for CIFAR-10 in the new FIGURE 10 in page 57 of the revised manuscript. For CIFAR-10, we use a 10-task stream with requests deleting $\\{2\\}$, $\\{5\\}$, and $\\{6,8\\}$ at $t=4,7,9$, respectively. As shown in Figure 10 of the revised manuscript, both Hessian unlearning and forgetting-enhanced Hessian achieve lower final errors than natural forgetting under all three regularizers, with Hessian unlearning attaining the lowest error in each case. These results demonstrate the benefit of explicit Hessian correction for this multi-request setting.
AuthorsOpenReview ↗

Response to Reviwer 2U54 (3/3)

Comparison to non-certified methods We consider non-certified unlearning baselines, including GA, NegGrad+, and an adaptation of the UniCLUN unlearning objective, within our continual learning setting (Chatterjee et al., 2024; Kurmanji et al., 2024; Huang et al., 2025). Under the evaluated configurations, optimizing the objectives of GA and NegGrad+ can cause parameter explosion and numerical divergence, preventing a finite approximation-error comparison with retraining. Although the UniCLUN adaptation remains numerically stable, continued optimization increases its approximation error to approximately 10 while reducing its test accuracy to 8.62% on CIFAR-100. Calibrating additional noise to this empirical error would further perturb an already degraded model, limiting its usefulness as a competitive baseline for our method.

These baselines also rely on more permissive data-access assumptions. GA requires access to the forget data, while NegGrad+ and the UniCLUN adaptation additionally use retained examples through a replay buffer. In contrast, our regularization-based framework assumes that raw data from previously trained tasks, including the task to be deleted, are unavailable when an unlearning request arrives. Consequently, these adapted baselines are not directly comparable under our data-access constraints and, in the evaluated configurations, perform poorly even when granted additional data access.

Certified unlearning framework requires a provable indistinguishability guarantee between the parameter distributions of the unlearned and retrained models. Empirical evaluations, including MIA-based evaluations, provide weaker evidence and do not establish this guarantee. Consequently, strong empirical performance by a non-certified method does not necessarily translate into satisfactory performance when the method is incorporated into a certified unlearning framework. Consequently, the evaluated non-certified baselines generally fail to provide effective unlearning under our certified-unlearning criteria (Mu & Klabjan, 2026).

We have added this discussion in Appendix H.4.

Regarding certified unlearning benchmarks, to the best of our knowledge, this is the first work to study certified unlearning across this family of CL methods. Consequently, no existing certified unlearning method provides a directly comparable baseline in this setting, and we include natural forgetting as an approximate unlearning benchmark, as it serves as a CL counterpart to gradient-based unlearning in offline learning.

We thank you again for your careful review and hope that our responses have addressed your concerns. We would be happy to answer any further questions. If you find our responses satisfactory, we would appreciate your consideration of a higher score.

Reviewer_e9n1OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper provides a theoretical framework for a family of regularization-based CL methods. The framework seeks to unify such methods under a Continual Learning/Unlearning (CLU) paradigm and provide definitions as well as resulting bounds on the errors based on the algorithm choice.

Critical strengths

For the most part, the paper is well written and easy to follow despite the dense theory (with exceptions as mentioned in the weaknesses section).

The contribution is significant as it provides a rigorous unified framework for analysing a broad family of CL methods under continual learning/unlearning.

Critical weaknesses

There are a plethora of minor and major presentation issues, as detailed below. I believe these are the most significant sources of weakness for the paper, as it is a mostly theoretical work and should be fully understandable. I will detail the specific issues in the questions section.

Additionally, the appendix is unduly long, and verifying any of the proofs would require substantial effort. However, I understand that this is to be expected from a mostly theoretical work.

Questions for authors

If the authors can adequately address the points below, I would be inclined to raise my score:

Definition 2.1: As this sits at the crux of the contribution, I think the authors should provide ample intuition and explanation as to why Eq. 3 is useful/desirable, especially for readers unfamiliar with differential privacy.

Definition 2.2: I think it would help to first define εLU\varepsilon_{LU}, then show (easily) that it decomposes into εL\varepsilon_L and εU\varepsilon_U

Theorem 3.3: The constant B is defined but never used. P_j is said to be a “CL-algorithm-specific matrix”, but no examples are given, making it unclear what it actually is.

Equation 9 is extremely confusing, as it implies wt−S1:t=wt−S1:t−1w_t^{ -S_{1:t}} = w_t^{-S_{1:t-1}}. What does this mean?

Theorem 4.1: what is min StS_t over? StS_t is a set of sets; is it over their index, i.e., the first set?

Eq 13. The notation is not well defined; what do the subscript indices under ∇\nabla mean? e.g. ∇21\nabla_{21}

Eq 15. is not well defined. s∈Ss \in S are subsets of datasets per the paper’s own definition, but somehow an inequality is formed between s and t, an index. This is a serious presentation error; I don’t even understand the intended message. The whole notion of synchronous/asynchronous gets lost here.

5 comments
AuthorsOpenReview ↗

Response to Reviwer e9n1 (1/3)

Thank you for recognizing our unified certified unlearning framework's importance for a broad family of CL methods and for your valuable feedback; we clarify your questions point by point below.

1. Response about Weakness

We will improve the presentation and make the appendix more concise and readable by adding proof sketches for the key results in Theorems 3.3, 4.1, and 5.2.

1. Response about Definition 2.1

We have added the suggested intuition and explanation immediately after Definition 2.1 at the top of page 4 in the revised manuscript. Intuitively, this definition provides a rigorous criterion for removing deleted-task influence by requiring the published model to be statistically indistinguishable from one retrained on retained tasks. It controls the entire released-model distribution, providing a stronger guarantee than indirect empirical forgetting metrics, such as reduced accuracy on deleted tasks. This guarantee must hold at every time for every admissible request history, which is particularly important in CL, where deleted-task influence can propagate through subsequent updates and interact with later requests.

2. Response about Definition 2.2

We have revised the manuscript as suggested in Definition 2.2. In current Definition 2.2, we first define the post-unlearning excess risk of a CLU algorithm ELU\mathcal{E}_{\rm LU}, where

ELU=E[1Nt!!!∑τ=1τ∉S≤ttRτ(w~t−S1:t)]!−!min⁡w∈W1Nt!!!∑τ=1τ∉S≤tt!!!Rτ(w).;\displaystyle \mathcal{E}_{\rm {LU}}= \mathbb{E}\bigg[ \frac{1}{N_t}\\!\\!\\! \sum_{\substack{\tau=1\\\\ \tau\notin S_{\le t}}}^t R_\tau(\tilde w_t^{-\mathbf{S}_{1:t}})\bigg] \\!-\\! \min_{w\in\mathcal W} \frac{1}{N_t}\\! \\!\\!\sum_{\substack{\tau=1\\\\ \tau\notin S_{\le t}}}^t\\!\\!\\! R_\tau(w) .\\;

It decomposes as ELU=EU+EL\mathcal E_{\rm LU}=\mathcal E_{\rm U}+\mathcal E_{\rm L}, where the unlearning loss EU\mathcal E_{\rm U} is

EU:=E[1Nt∑τ=1τ∉S≤tt(Rτ(w~t−S1:t)−Rτ(wt−S≤t))],\displaystyle \mathcal E_{\rm U} :={} \mathbb E\bigg[ \frac{1}{N_t} \sum_{\substack{\tau=1\\\\ \tau\notin S_{\le t}}}^{t} \left( R_\tau(\tilde w_t^{-\mathbf S_{1:t}}) - R_\tau(w_t^{-S_{\le t}}) \right) \bigg],

and the continual learning excess risk EL\mathcal E_{\rm L} is

EL:=E[1Nt∑τ=1τ∉S≤ttRτ(wt−S≤t)]−min⁡w∈W1Nt∑τ=1τ∉S≤ttRτ(w).\displaystyle \mathcal E_{\rm L} :={} \mathbb E\bigg[ \frac{1}{N_t} \sum_{\substack{\tau=1\\\\ \tau\notin S_{\le t}}}^{t} R_\tau(w_t^{-S_{\le t}}) \bigg] - \min_{w\in\mathcal W} \frac{1}{N_t} \sum_{\substack{\tau=1\\\\ \tau\notin S_{\le t}}}^{t} R_\tau(w).

**3. Response about Theorem 3.3 **

The constant BB is used in Theorem C.1 in Appendix C of the revised manuscript, complete version of Theorem 3.3, which provides the complete form of the bound in Theorem 3.3. To avoid confusion, we have removed the BB from Theorem 3.3 and introduced it only in Theorem C.1 in the revised manuscript.

The algorithm-specific matrix PiP_i is also defined in Theorem C.1 using the quantities introduced in Eqs. (21) and (22). Specifically, Pi:=[H~τi(wτi−S≤t,w^τi)+∇222Ω~D1:τi−1−S≤t(wτi−1−S≤t,wτi−1−S≤t,wτi−S≤t,wτi−1−S≤t)]−1∇222Ω~D1:τi−1−S≤t(wτi−1−S≤t,wτi−1−S≤t,wτi−S≤t,wτi−1−S≤t).\displaystyle P_i := {} \left[ \widetilde H_{\tau_i} \left(w_{\tau_i}^{-S_{\le t}},\hat w_{\tau_i}\right) + \nabla_{22}^2\widetilde\Omega_{D_{1:\tau_{i-1}}^{-S_{\le t}}} \left( w_{\tau_{i-1}}^{-S_{\le t}}, w_{\tau_{i-1}}^{-S_{\le t}}, w_{\tau_i}^{-S_{\le t}}, w_{\tau_{i-1}}^{-S_{\le t}} \right) \right]^{-1} \nabla_{22}^2\widetilde\Omega_{D_{1:\tau_{i-1}}^{-S_{\le t}}} \left( w_{\tau_{i-1}}^{-S_{\le t}}, w_{\tau_{i-1}}^{-S_{\le t}}, w_{\tau_i}^{-S_{\le t}}, w_{\tau_{i-1}}^{-S_{\le t}} \right).

Here, the line-averaged quantities H~n\widetilde H_n and ∇ab2Ω~\nabla_{ab}^2\widetilde\Omega are defined in Eqs. (21) and (22) as H~n(x,y):=∫01∇2L(x+u(y−x),Dn),du\displaystyle \widetilde H_n(x,y) := \int_0^1 \nabla^2L\bigl(x+u(y-x),D_n\bigr),du and ∇ab2Ω~D1:n−S≤t(x1,y1,x2,y2):=∫01∇ab2ΩD1:n−S≤t(x1+u(y1−x1),x2+u(y2−x2)),du.\displaystyle \nabla_{ab}^2\widetilde\Omega_{D_{1:n}^{-S_{\le t}}} (x_1,y_1,x_2,y_2) := \int_0^1 \nabla_{ab}^2\Omega_{D_{1:n}^{-S_{\le t}}} \bigl( x_1+u(y_1-x_1), x_2+u(y_2-x_2) \bigr),du.

For example, for the Reg-CL methods with task-additive regularizers introduced in Section 4.1, PiP_i takes the form Pi=[H~τi(wτi−S≤t,w^τi)+λI+∑k=1i−1F(wτk−S≤t,Dτk)]−1[λI+∑k=1i−1F(wτk−S≤t,Dτk)].\displaystyle P_i = {} \left[ \widetilde H_{\tau_i} \left(w_{\tau_i}^{-S_{\le t}},\hat w_{\tau_i}\right) + \lambda I + \sum_{k=1}^{i-1} F\left(w_{\tau_k}^{-S_{\le t}},D_{\tau_k}\right) \right]^{-1} \left[ \lambda I + \sum_{k=1}^{i-1} F\left(w_{\tau_k}^{-S_{\le t}},D_{\tau_k}\right) \right]. Here, τ1<⋯<τNt\tau_1<\cdots<\tau_{N_t} enumerate the retained task indices in chronological order, with τ1,…,τNt=[t]∖S≤t{\tau_1,\ldots,\tau_{N_t}}=[t]\setminus S_{\le t} and Nt=t−∣S≤t∣N_t=t-|S_{\le t}|. Different CL methods construct different task-specific statistics F(w,D)F(w,D), such as Fisher information for EWC. Consequently, PiP_i captures how the chosen CL regularizer affects the propagation factors ∣Pi∣\\|P_i\\| and ∣I−Pi∣\\|I-P_i\\| in Theorem 3.3.

We have added an explicit reference to the definition of PiP_i in the revised statement of Theorem 3.3.

AuthorsOpenReview ↗

Response to Reviwer e9n1 (2/3)

4. Response about Equation 9

Eq. (9) holds for natural forgetting because we perform no explicit unlearning update and only add noise to the model wt−S1:tw_t^{-\mathbf S_{1:t}} to output publish model w~t−S1:t\tilde w_t^{-\mathbf S_{1:t}}, calibrated to an upper bound on the approximation error ∣wt−S1:t−wt−S≤t∣\\|w_t^{-\mathbf S_{1:t}}-w_t^{-S_{\le t}}\\|.

Specifically, In Eq. (9), we aim to clarify how natural forgetting operates within the CLU framework. In Figure 1 and the second paragraph of Section 2.1, we define wt−S1:t−1w_t^{-\mathbf S_{1:t-1}} as the model obtained after training on task tt's dataset DtD_t. After receiving and processing the unlearning request StS_t, the model is updated to wt−S1:tw_t^{-\mathbf S_{1:t}}, the output of the unlearning operation RA\mathcal R_{\mathcal A} with wt−S1:t−1w_t^{-\mathbf S_{1:t-1}} as its input model:

wt−S1:t=RA(wt−S1:t−1,D1:t,S1:t).\displaystyle w_t^{-\mathbf S_{1:t}} = \mathcal R_{\mathcal A}\left(w_t^{-\mathbf S_{1:t-1}},\mathbf D_{1:t},\mathbf S_{1:t}\right).

Natural forgetting relies on the forgetting effect of continual learning to remove deleted-task influence. Thus, RA\mathcal R_{\mathcal A} leaves its input model unchanged:

RA(x,D1:t,S1:t)=x.\displaystyle \mathcal R_{\mathcal A}\left(x,\mathbf D_{1:t},\mathbf S_{1:t}\right) = x.

Substituting x=wt−S1:t−1x=w_t^{-\mathbf S_{1:t-1}} gives the identity stated in Eq. (9):

wt−S1:t=RA(wt−S1:t−1,D1:t,S1:t)=wt−S1:t−1.\displaystyle w_t^{-\mathbf S_{1:t}} = \mathcal R_{\mathcal A}\left(w_t^{-\mathbf S_{1:t-1}},\mathbf D_{1:t},\mathbf S_{1:t}\right) = w_t^{-\mathbf S_{1:t-1}}.

To avoid confusion, we have revised Eq. (10) (previously Eq. (9)) to explicitly describe natural forgetting as

RA(wt−S1:t−1,D1:t,S1:t)=wt−S1:t−1.\displaystyle \mathcal R_{\mathcal A}\left(w_t^{-\mathbf S_{1:t-1}},\mathbf D_{1:t},\mathbf S_{1:t}\right) = w_t^{-\mathbf S_{1:t-1}}.

5. Response about Theorem 4.1

We would like to clarify that StS_t is a set of task indices, rather than a set of sets. Each index in StS_t identifies a task whose dataset is requested to be unlearned.

Moreover, Theorem 4.1 uses min⁡S≤t\min S_{\le t}. Here, S≤tS_{\le t} is the cumulative set of task indices requested for unlearning up to and including time tt, i.e., S≤t=⋃i=1tSiS_{\le t}=\bigcup_{i=1}^t S_i, as defined in the second paragraph of Section 2.1.

For example, consider a continual learning process over tasks 1,…,51,\dots,5, with two unlearning requests: one arrives at time 33 to unlearn task 11, and the other arrives at time 44 to unlearn task 22. The request sets are S1=S2=S5=∅,S3=1,S4=2.\displaystyle S_1=S_2=S_5=\emptyset, S_3=\\{1\\}, S_4=\\{2\\}. The cumulative sets of requested task indices are S≤1=S≤2=∅,S≤3=1,S≤4=S≤5=1,2.\displaystyle S_{\le 1}=S_{\le 2}=\emptyset, S_{\le 3}=\\{1\\}, S_{\le 4}=S_{\le 5}=\\{1,2\\}. Thus, StS_t specifies the task indices requested for unlearning at time tt, while S≤tS_{\le t} contains all task indices requested for unlearning through time tt. In this example, min⁡S≤3=min⁡S≤4=min⁡S≤5=1\min S_{\le 3}=\min S_{\le 4}=\min S_{\le 5}=1, identifying the earliest task requested for unlearning.

We have added a reference to this example in the second paragraph of Section 2.1, directing readers to Appendix A on page 14 of the revised manuscript.

6. Response about Eq 13.

Following your suggestion, we have revised the two paragraphs preceding Eq. (14) on page 7 in the revised manuscript (Eq. (13) in the original manuscript) to formally define this notation. We use ∇i\nabla_i to denote the gradient with respect to the ii-th argument of a function and define ∇ij2:=∇j(∇i)\nabla_{ij}^2:=\nabla_j(\nabla_i) as the corresponding second-order derivative.

For example, for a regularizer Ω(x,y)\Omega(x,y), ∇212Ω(x,y)\nabla_{21}^2\Omega(x,y) is the Jacobian of ∇2Ω(x,y)\nabla_2\Omega(x,y) with respect to xx, whereas ∇222Ω(x,y)\nabla_{22}^2\Omega(x,y) is its Hessian with respect to yy.

For the loss L(w,D)L(w,D), all derivatives are taken with respect to the model parameter ww, with DD held fixed. We therefore write ∇L\nabla L and ∇2L\nabla^2L in place of ∇1L\nabla_1L and ∇112L\nabla_{11}^2L, respectively.

AuthorsOpenReview ↗

Response to Reviwer e9n1 (3/3)

7. Response about Eq 15.

Thank you for raising this concern. After carefully checking, we believe there is no notation error here and that the use of StS_t is consistent with its definition in the manuscript. The second paragraph of Section 2.1 states: “After learning task tt, the system may receive an unlearning request StS_t, a subset of previous task indices identifying the datasets Dii∈St\\{D_i\\}_{i\in S_t} whose influence must be removed from wt−S1:t−1w_t^{-\mathbf S_{1:t-1}}.” We also state that "require each admissible request to satisfy St⊆[t]∖S≤t−1S_t\subseteq[t]\setminus S_{\le t-1}". These statements explicitly define StS_t as a set of task indices, rather than a set of datasets. Underling this definition, As specified in Eq. (16) of the revised manuscript (Eq. (15) in the original manuscript), forward-synchronous unlearning requires that, for any two consecutive unlearning times tit_i and ti+1t_{i+1}, the request Sti+1S_{t_{i+1}} contain only tasks with indices greater than tit_i. For example, consider a continual learning process over tasks 1,…,51,\dots,5, with two unlearning requests: one arrives at time 22 to unlearn task 11, and the other arrives at time 44 to unlearn task 33. Since the task index in S4=3S_4=\\{3\\} is greater than the previous unlearning time 22, this sequence is forward-synchronous and satisfies Eq. (16).

Any unlearning sequence that violates this condition is called asynchronous. For example, consider a sequence in which one request arrives at time 33 to unlearn task 22, and another arrives at time 44 to unlearn task 11. Since the task index in S4=1S_4=\\{1\\} is less than the previous unlearning time 33, this sequence violates Eq. (16) and is therefore asynchronous. We have added a sentence on page 8, immediately after Eq. (16), directing readers to the new example in Appendix A in the revised manuscript.

We thank you again for your careful review and hope that our responses have addressed your concerns. We would be happy to answer any further questions. If you find our responses satisfactory, we would appreciate your consideration of a higher score.

Reviewer_FCEtOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces the Continual Learning-Unlearning (CLU) problem, combining regularization-based CL (EWC, MAS, Hessian methods) with (epsilon,delta)-certified machine unlearning. The core contribution is a decomposition of post-unlearning excess risk into unlearning error epsilon_U and CL excess risk epsilon_L, which motivates two algorithms: Natural Forgetting Unlearning (NFU, first-order, near-zero overhead) and Hessian-Based Unlearning (HBU, Newton-style correction, tighter epsilon_U). The paper shows these sit on a Pareto frontier and validates on CIFAR-100/ResNet-18.

Critical strengths
  1. Novel problem formulation. CLU is a genuine gap in the literature; the risk decomposition (Eq. 4-5) cleanly separates unlearning and CL components and drives the algorithm design.
  2. General framework. Results cover the full Reg-CL family via a single regularizer, with concrete instantiations for EWC and task-additive regularizers.
  3. Complementary algorithms with Pareto analysis. Providing two methods with provably distinct trade-off profiles is more useful than a single bound.
  4. Rigorous certification. The Pub-CLU Gaussian mechanism correctly instantiates standard DP machinery for certified unlearning.
Critical weaknesses
  1. Unrealistic assumptions for deep networks. Strong convexity (Assumption 2.3) and Lipschitz Hessian (Assumption 2.4) do not hold for ResNet-18. The theory-practice gap is large and unaddressed.
  2. Thin experiments. Only one dataset (CIFAR-100) and one architecture, with a fixed pretrained backbone. No unlearning baselines are provided; it is impossible to assess whether NFU/HBU improve on simpler approaches such as gradient ascent or fine-tuning on retained data.
  3. HBU compute/storage unaddressed. Exact Hessian storage is O(d^2). The paper does not discuss approximation strategy (diagonal, K-FAC, etc.) or the resulting error. The oracle estimator required by Theorem 5.2 is unavailable without retraining, yet no bound on the approximation gap is given.
  4. Pareto frontier is qualitative only. No quantitative condition characterizes when to prefer NFU vs. HBU as a function of task count, deletion set size, or regularization strength.
Questions for authors
  1. How sensitive are the bounds to the strong convexity constant mu as mu -> 0?
  2. How is the Hessian approximated in the experiments, and what approximation error does this introduce?
  3. What happens to HBU when unlearning requests arrive out of temporal order (violating the forward-synchronous structure)?
  4. Can the authors report numerical epsilon_U and epsilon_L values for NFU and HBU alongside a simple baseline on CIFAR-100?
3 comments
AuthorsOpenReview ↗

Response to Reviwer FCEt (1/3)

Thank you for your valuable suggestions and comments. We address your questions and provide clarifications point by point below.

1. Response to W1 and Q1 We would like to clarify that Assumption 2.3 covers both μ>0\mu>0 and μ≤0\mu\leq 0, rather than being restricted to μ>0\mu>0. Immediately after Assumption 2.3, on line 207 of the revised manuscript, we explicitly state that μ<0\mu<0 is allowed to accommodate nonconvex loss functions. All theoretical results throughout the paper hold for general values of μ\mu.

We would also like to clarify that Assumption 2.4 does not impose any Lipschitz continuity condition on the Hessian. Rather, it states that the model can attain the global minimum loss within a neighborhood of its initialization, a phenomenon widely observed in overparameterized neural networks (Jacot et al., 2018; Chizat et al., 2019; Li & Liang, 2018).

Therefore, Assumption 2.3 with general μ\mu and Assumption 2.4 are compatible with overparameterized ResNet-18 models. We believe that these assumptions do not introduce an obvious gap between our theoretical analysis and practical implementation.

2. Response to W2

We have significantly expanded our experiments to include additional benchmarks, datasets, and model architectures, as well as MIA-based evaluations. Specifically, we add experiments on CIFAR-10 using a convolutional neural network trained from scratch, comprising three convolutional blocks with 6464, 128128, and 256256 output channels, respectively, and 2,473,6102{,}473{,}610 trainable parameters in total.

New MIA evaluation. we have added MIA-based evaluations in the new Figures 3–6. We evaluate membership leakage using a standard modified-entropy MIA and LiRA (Carlini et al.,2022), a state-of-the-art attack that uses shadow models to calibrate a likelihood-ratio test for membership inference. Figure 3 shows that HU substantially reduces detectable membership leakage for the deleted task, achieving modified-entropy AUCs close to retraining. Figures 4 and 5 show consistent LiRA results: HU approaches retraining and leaves less detectable membership information than NF. Its smaller empirical approximation error also requires less perturbation under the same noise-calibration procedure, yielding higher test accuracy than NF and approaching retraining performance on CIFAR-100 (Figure 6). Although NF provides weaker protection than HU, it reduces attack AUC relative to the target-task checkpoint, demonstrating the contribution of natural forgetting. Overall, HU offers a more favorable empirical privacy–utility tradeoff under the evaluated attacks.

We also added the following experiments in the revised manuscript:

  • Unlearning request sequences: We moved the CIFAR-100 sequence experiments from the appendix to the main text (Figure 2) and added corresponding CNN experiments on CIFAR-10 (Appendix H.1, Figure 8 on page 56).
  • Unlearning–accuracy tradeoffs: We added CIFAR-10 experiments comparing approximation error and retrained-model accuracy across five regularization-based CL methods (Table 1(b), page 9).
  • Unlearning along the CLU trajectory: We added CIFAR-10 experiments comparing natural forgetting, Hessian-based unlearning, and forgetting-enhanced Hessian unlearning under multiple deletion requests (Figure 10 on page 57).

We evaluated GA as a non-certified unlearning baseline, as suggested. However, under the tested configurations, optimizing its objective can cause parameter explosion and numerical divergence, preventing a finite approximation-error comparison with retraining. Calibrating additional noise to a large empirical error would introduce substantial perturbations to an already degraded model, limiting its usefulness as a competitive baseline. Moreover, GA requires access to the forget data, whereas our setting assumes that raw data from previously trained tasks, including the task to be deleted, are unavailable when an unlearning request arrives. GA therefore does not provide a directly comparable baseline under our data-access constraints.

Regarding certified unlearning benchmarks, to the best of our knowledge, this is the first work to study certified unlearning across this family of CL methods. Consequently, no existing certified unlearning method provides a directly comparable baseline in this setting. We therefore include natural forgetting as an approximate unlearning benchmark that relies on subsequent learning updates to attenuate the influence of previously learned tasks, serving as a continual-learning counterpart to gradient-based unlearning through continued training on retained data in offline settings.

AuthorsOpenReview ↗

Response to Reviwer FCEt (2/3)

3. Response to W3

We would like to clarify that our framework explicitly allows approximate Hessians in place of exact Hessians, and our theoretical bounds for Hessian-based unlearning account for the resulting approximation error. On lines 356 of page 7, we introduce ∇2L^\widehat{\nabla^2 L} as an approximation to the exact loss Hessian and use it in the Hessian-based unlearning update in Eq. (14). Theorem 5.2 explicitly accounts for the increase in the unlearning-error bound due to Hessian approximation through algorithm-specific parameters.

For example, Theorem 5.2 involves ρˉqtk\bar\rho_q^{t_k}, which bounds the discrepancy between the approximate correction-propagation operator and its exact, path-averaged counterpart. Equation (62) in Appendix E.1 defines this quantity as ρˉjt:=sup⁡∥P~jt−P^jt∥,\bar\rho_j^t:=\sup\left\|\widetilde P_j^t-\widehat P_j^t\right\|, where

\widehat P_n^t := {}& -\left[ \widehat H_n+ \nabla_{22}^2\Omega_{D_{1:n-1}^{-S_{\le t}}} \left(w_{n-1}^{-\mathbf S_{1:n-1}},w_n^{-\mathbf S_{1:n-1}}\right) \right]^{-1} \nabla_{21}^2\Omega_{D_{1:n-1}^{-S_{\le t}}} \left(w_{n-1}^{-\mathbf S_{1:n-1}},w_n^{-\mathbf S_{1:n-1}}\right), \end{aligned}$$ and $\widehat H_n=\widehat{\nabla^2 L}(w_n^{-\mathbf S_{1:n-1}},D_n)$ is the approximate loss Hessian. Thus, $\bar\rho_j^t$ increase with Hessian approximation error, together with discrepancies arising from the different evaluation points and regularizer estimates. Larger values of $\bar\rho_j^t$ increase the unlearning-error bound in Theorem 5.2. Appendix G further analyzes the dependence on Hessian approximation error for specific CL methods. In particular, Appendix G introduces a uniform Hessian approximation-error bound $\nu$ satisfying $$\left\| \widehat{\nabla^2 L}(w,D)-\nabla^2 L(w,D) \right\|_2 \leq \nu$$ and explicitly incorporates $\nu$ into the bounds on the algorithm-specific parameters used in Theorem 5.2. Our main contribution of certified-unlearning framework for continual learning accommodates a broad range of Hessian approximations to reduce storage costs. For example, our experiments use a diagonal Hessian approximation requiring $O(d)$ storage. We have added a footnote on page 7 to highlight this flexibility. **We would also like to clarify that constructing the oracle regularizer estimator used in Theorem 5.2 does not require retraining.** In the Hessian-based unlearning update in Eq. (14), the quantities $\Omega_{\mathbf D_{1:n-1}^{-\mathbf S_{1:t}}}$ and $\Omega_{\mathbf D_{1:n-1}^{-\mathbf S_{1:n-1}}}$ for $i,j=1,2$ are regularizer estimators, rather than the exact retraining-induced regularizers. We use this formulation to accommodate different estimators of the retraining-induced regularizer. Each estimator must be constructed using only information available during the CLU process. For example, Section E.3 provides the following estimator: $$\frac{1}{2} \lVert y-x\rVert_{\lambda I+\sum_{i\in[n-1]\setminus S_{\le t}}F(w_i^{-\mathbf S_{1:i-1}},D_i)}^{2}.$$ Here, each task-specific statistic $F(w_i^{-\mathbf S_{1:i-1}},D_i)$ is computed and stored during CLU. Upon receiving an unlearning request, the estimator excludes the statistics associated with the deleted tasks. Its construction therefore requires neither retraining nor access to historical task datasets at unlearning time.
AuthorsOpenReview ↗

Response to Reviwer FCEt (3/3)

4. Response to W4

We can establish an explicit criterion for choosing between NFU and HBU in practice. We have derived unlearning-loss bounds for both methods in Theorems 4.1, 5.2, and E.1. These bounds depend on the task count, deletion set size, regularization strength, and algorithm-specific parameters. In practice, we can first numerically estimate the required parameters. Upon receiving an unlearning request, we can then evaluate the corresponding bounds using these estimates together with the task count, deletion set size, and regularization strength, and select the method with the smaller bound for that request.

We can also obtain quantitative guidance from the experimental results in Table 1. When predictive accuracy is prioritized, we should choose Diag-Hessian on CIFAR-100 and MAS on CIFAR-10, as shown in Table 1(a) and Table 1(b), respectively. Their parameter-dependent penalties constrain important directions while allowing adaptation elsewhere, consistent with Section 3.

When approximation quality is prioritized, Table 1(a) suggests choosing HU with MAS for Tasks 2 and 8, and HU with accumulated ℓ2\ell_2 for Task 5. On CIFAR-10, Table 1(b) favors HU with Diag-Hessian for Task 2 and accumulated ℓ2\ell_2 for Task 4. These results are consistent with Sections 4–5, where stronger effective regularization can attenuate the deleted task’s influence.

Deletion timing also guides the choice between NF and HU. Under Diag-Hessian on CIFAR-100, NF is preferable for Tasks 2 and 5, which benefit from more subsequent updates, whereas HU is preferable for the more recent Task 8. On CIFAR-10, HU is preferable in both evaluated settings, consistent with the shorter training horizon and the analysis in Sections 4–5.

We have added this quantitative discussion of the experimental results on page 9 of the revised manuscript.

5. Response to Q2

We use a simple diagonal Hessian approximation in all experiments. Over the sampled region, the numerically estimated maximum approximation error, $ \left\lVert \widehat{\nabla^2 L}(w,D)-\nabla^2 L(w,D) \right\rVert_2, $ is approximately 2. Our empirical results show that Hessian-based unlearning remains effective under this approximation. We can further reduce the Hessian approximation error by using more accurate Hessian approximations.

6. Response to Q3

We would like to clarify that HBU does not require a forward-synchronous structure and can handle requests that violate this condition. The HBU update formula in Eq. (15) applies to any unlearning sequence, and our theoretical results bound the unlearning approximation error for arbitrary request orders.

Specifically, Theorem 5.2 provides an approximation-error bound for forward-synchronous sequences. We define sequences that violate the forward-synchronous condition in Eq. (16) as asynchronous and discuss them in Section 5.2. Theorem E.1 in Appendix E provides the formal approximation-error bound and proof for asynchronous sequences, while Section 5.2 explains the intuition behind unlearning under such request orders.

We also evaluate unlearning under both forward-synchronous and asynchronous request sequences, as shown in Figure 2 of the revised manuscript.

7. Response to Q4

Calculating EU\mathcal{E}_{\mathrm{U}} and EL\mathcal{E}_{\mathrm{L}} requires access to the underlying data-generating distributions, which are unavailable in practice. Nevertheless, the approximation error ∥wt−S1:t−wt−S≤t∥\left\lVert w_t^{-\mathbf{S}{1:t}}-w_t^{-S{\le t}}\right\rVert and test accuracy reported in Table 1 serve as empirical proxies for EU\mathcal{E}_{\mathrm{U}} and EL\mathcal{E}_{\mathrm{L}}, respectively. As discussed in Section 3.2, the upper bound on EU\mathcal{E}_{\mathrm{U}} scales linearly with the approximation-error bound under fixed certification parameters. In continual learning, test accuracy is commonly used as an empirical proxy for EL\mathcal{E}_{\mathrm{L}} (Kirkpatrick et al., 2017; Aljundi et al., 2018; Yin et al., 2020; Zenke et al., 2017). Higher accuracy generally suggests lower EL\mathcal{E}_{\mathrm{L}} and better CL performance. Thus, Table 1 provides an empirical assessment of the two components in our theoretical decomposition.

We thank you again for your careful review and hope that our responses have addressed your concerns. We would be happy to answer any further questions. If you find our responses satisfactory, we would appreciate your consideration of a higher score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.