acceptodds
Under review as a conference paper at ICLR 2027

CoForget: Composing Heterogeneous Unlearning Requests to Activate Dormant Backdoors

Abstract

Unlearning-activated backdoor attacks keep a backdoor dormant after training and activate it later through an unlearning request. Requests composed entirely of camouflage samples can induce this transition but may also reduce benign accuracy. Since clean samples can influence backdoor behavior, incorporating them alongside camouflage samples provides an additional way to improve this balance. To this end, we introduce CoForget, a request-composition attack that combines a small fixed camouflage component with deliberately selected clean samples. CoForget first calibrates a universal trigger with a logit gap objective to promote dormancy without excessively suppressing the target response. It then greedily selects a clean subset whose cumulative gradient aligns with an activation direction estimated on a local shadow model. Neither trigger nor request construction requires access to the victim model or knowledge of the provider's unlearning algorithm. Experiments across three image classification datasets and three approximate unlearning algorithms show low attack success before unlearning and high success afterward. In the evaluated clean component ablations, selected samples improve mean benign accuracy over camouflage-only requests and retain high activation. Comparisons at matched total request sizes further show strong activation while using substantially fewer camouflage samples than the evaluated baselines. These findings motivate considering request composition when assessing unlearning security.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Backdoor attacks, machine unlearning, heterogeneous unlearning requests, clean sample selection

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

This paper attacks machine unlearning itself. A clean-label backdoor is installed together with camouflage samples that keep it dormant, and the backdoor activates once a legitimate unlearning request removes the camouflage. CoForget turns the request into a heterogeneous one, pairing a small camouflage component with deliberately selected clean samples, using logit gap calibration for the trigger and cumulative gradient alignment for sample selection.

Critical strengths
  1. Using a legitimate unlearning request as the trigger is a clear and security-relevant threat concept.
  2. The paper is written in clear and accessible language, and is generally easy to follow.
Critical weaknesses
  1. The heterogeneous request contains substantially more samples than the camouflage-only request. The two settings therefore do not have the same request size, so the reported gain in benign accuracy cannot be attributed solely to the combination of camouflage and selected clean samples. The claimed activation-utility balance is neither isolated nor mechanistically supported.
  2. The paper should provide a rule for selecting the logit-gap interval on new datasets.
  3. The comparison with prior attacks does not isolate request composition. CoForget combines a calibrated trigger with heterogeneous request construction, whereas the baselines use BadNets triggers. The paper does not separate gains from trigger calibration and gains from heterogeneous composition.
  4. The framework relies on a source model trained on a related-domain auxiliary dataset and on dataset-specific trigger calibration and selector settings. The paper does not use a single configuration across all datasets, and the perturbation budget, logit-gap interval, and class cap vary by dataset.
Questions for authors

As above.

2 comments
AuthorsOpenReview ↗

Response to Reviewer YjtK (Part 1/2): W1–W2

Thank you for your careful review and constructive feedback. We have revised the manuscript and highlighted the changes in blue in the updated PDF for ease of reference.

W1. Request size and attribution of the activation–utility gains

The heterogeneous request contains substantially more samples than the camouflage-only request. The two settings therefore do not have the same request size, so the reported gain in benign accuracy cannot be attributed solely to the combination of camouflage and selected clean samples. The claimed activation–utility balance is neither isolated nor mechanistically supported.

Response:

To isolate the effect of request composition, we compare heterogeneous and camouflage-only requests at the same total request size. We added a comparison within the CoForget framework under NegGrad+. The heterogeneous configurations reduce mean BA drops by 1.69, 1.92, and 4.65 percentage points on CIFAR-10, CIFAR-100, and Tiny-ImageNet, respectively, while maintaining high post-unlearning ASR.

DatasetRequestASR before (%)ASR after (%)ΔBA (pp)
CIFAR-10Heterogeneous12.23 ± 3.3298.98 ± 1.751.26 ± 0.31
CIFAR-10Camouflage-only2.48 ± 0.5865.68 ± 56.882.95 ± 0.22
CIFAR-100Heterogeneous19.34 ± 3.6499.23 ± 0.702.43 ± 0.96
CIFAR-100Camouflage-only14.12 ± 1.9299.38 ± 1.074.35 ± 0.92
Tiny-ImageNetHeterogeneous7.15 ± 2.4297.87 ± 1.743.36 ± 0.44
Tiny-ImageNetCamouflage-only3.55 ± 2.2695.41 ± 5.538.01 ± 6.68

Selected vs random comparisons identify an additional benefit from clean sample selection with matched camouflage and clean budgets. Selected clean samples achieve higher mean BA and ASR across the three unlearning algorithms on CIFAR-10 (Table 4) and across all three datasets under NegGrad+ (Appendix Table 17).

Our expanded Appendix B provides a theoretical foundation for the activation–utility gains. When clean selection increases the combined request's activation efficiency, the same first-order target-label loss reduction requires a smaller linearized displacement. Under Proposition 1's margin condition, the corresponding upper bound on benign accuracy loss does not increase. This explains how activation-oriented selection can support benign utility without explicitly optimizing BA.


W2. Choosing the logit-gap interval on new datasets

The paper should provide a rule for selecting the logit-gap interval on new datasets.

Response:

We determined the dataset-specific logit-gap intervals through empirical evaluation of multiple candidate intervals, jointly considering pre-unlearning ASR, post-unlearning ASR, and benign accuracy. For a new dataset, we recommend using the positive intervals in Appendix A.2 as initial candidates, evaluating them under the intended unlearning setting, and selecting an interval that balances initial dormancy, subsequent activation, and benign utility.

The default intervals provide a favorable balance between pre-unlearning dormancy, post-unlearning activation, and benign utility. On CIFAR-100, the default maintains high post-unlearning ASR, whereas [3,4] improves initial suppression but reduces subsequent ASR to 5.38%. On Tiny-ImageNet, [0.1,0.5] also improves suppression, but the default preserves higher post-unlearning BA.

DatasetIntervalASR before (%)BA after (%)ASR after (%)
CIFAR-10[3.0, 4.0]10.4692.6094.06
CIFAR-10[2.0, 3.0]13.6991.0499.99
CIFAR-10[0.2, 1.0]15.0492.5099.98
CIFAR-10[0.2, 0.6]16.6892.05100.00
CIFAR-10[0.1, 0.5]20.4792.22100.00
CIFAR-100[3.0, 4.0]8.3866.015.38
CIFAR-100[2.0, 3.0]18.1264.7598.29
CIFAR-100[1.0, 1.5]16.1267.9899.84
CIFAR-100[0.5, 1.0]21.5160.35100.00
CIFAR-100[0.2, 0.7]31.0467.86100.00
Tiny-ImageNet[3.0, 4.0]12.3854.0994.69
Tiny-ImageNet[2.0, 3.0]8.3052.2298.04
Tiny-ImageNet[0.2, 1.0]9.8554.0198.55
Tiny-ImageNet[0.2, 0.6]7.8252.7596.99
Tiny-ImageNet[0.1, 0.5]6.4651.4598.31
AuthorsOpenReview ↗

Response to Reviewer YjtK (Part 2/2): W3–W4

W3. Separating trigger calibration from request composition

The comparison with prior attacks does not isolate request composition. CoForget combines a calibrated trigger with heterogeneous request construction, whereas the baselines use BadNets triggers. The paper does not separate gains from trigger calibration and gains from heterogeneous composition.

Response:

To disentangle the contributions of trigger calibration and heterogeneous request composition, we conduct two complementary sets of controlled comparisons.

Separating request composition. We fix the CoForget trigger and compare heterogeneous and camouflage-only requests at the same total request budget under NegGrad+. Across three attacker seeds per dataset, heterogeneous configurations achieve smaller mean BA drops while maintaining high post-unlearning ASR across all three datasets (see the table in W1 above). Selected clean samples also achieve higher mean BA and ASR than random clean samples under the same trigger, victim checkpoint, camouflage component, and clean budget (Table 4 and Appendix Table 17).

Separating trigger calibration. Table 3 removes calibration while retaining the heterogeneous request construction and sample budgets. Removing calibration weakens the dormancy-to-activation transition, particularly on the CIFAR datasets. Appendix Table 16 further compares the calibrated trigger with five alternative trigger designs under the same heterogeneous request protocol and NegGrad+. The results show that alternative triggers exhibit different dormancy–activation trade-offs, whereas the calibrated trigger maintains low pre-unlearning ASR and high post-unlearning ASR across all three datasets.


W4. Auxiliary data and dataset-specific configurations

The framework relies on a source model trained on a related-domain auxiliary dataset and on dataset-specific trigger calibration and selector settings. The paper does not use a single configuration across all datasets, and the perturbation budget, logit-gap interval, and class cap vary by dataset.

Response:

Auxiliary data. CoForget uses auxiliary public data from a related domain to construct the shadow model required for trigger calibration and clean sample selection (Appendix A.2). This use of auxiliary data is consistent with prior unlearning attacks. UBA-Inf (Huang et al., 2024) also uses auxiliary training data and a publicly available ImageNet-pretrained surrogate model.

Dataset-specific configurations. CoForget uses the same loss formulations and cumulative greedy selection procedure across datasets, while allowing numerical settings to vary with dataset characteristics. Differences in dataset size, class structure, and image characteristics motivate dataset-specific configurations. We determine these settings through empirical evaluation of multiple configurations, assessing model performance before and after unlearning.

References

Zirui Huang, Yunlong Mao, and Sheng Zhong. UBA-Inf: Unlearning activated backdoor attack with Influence-Driven camouflage. In 33rd USENIX Security Symposium (USENIX Security 24), pp. 4211–4228, 2024.

Reviewer_12HrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies backdoor attacks in Machine Unlearning. Some existing methods activate dormant backdoors by unlearning Camouflage Samples, but this may reduce the model's accuracy on clean data. The authors propose CoForget, which adds selected Clean Samples to the unlearning request instead of only removing Camouflage Samples. The goal is to activate the backdoor while reducing the loss of normal model accuracy.

The method has two main parts. First, the authors use Logit Gap Calibration to adjust the Trigger, so the backdoor has a low attack success rate before unlearning but can still be activated later. Second, they use a local Shadow Model to estimate the gradient direction for backdoor activation. Then, a Greedy Selection method is used to select Clean Samples whose combined gradients are close to this direction. Finally, the selected Clean Samples and Camouflage Samples are combined into one Unlearning Request.

The experiments are conducted on CIFAR-10, CIFAR-100, and Tiny-ImageNet, using three unlearning algorithms: NegGrad+, SAM, and SCRUB. The results show that CoForget keeps a relatively low ASR before unlearning, while the ASR increases significantly after unlearning. The ablation studies also show that the proposed method achieves higher benign accuracy while maintaining a high ASR, compared with only unlearning Camouflage Samples or adding randomly selected Clean Samples.

The main focus of this paper is not to introduce a completely new backdoor attack, but to study how to select and combine samples in an unlearning request. The authors use Gradient Alignment to select Clean Samples and make use of their influence on model updates to help activate the backdoor while reducing the drop in benign accuracy.

Critical strengths
  1. The method introduces a novel approach. I think one strength of this paper is that it does not only rely on Camouflage Samples to activate the backdoor. Instead, it adds selected Clean Samples to the unlearning request. The authors use Gradient Alignment to select these samples rather than choosing them randomly. This provides a different way to study backdoor attacks in Machine Unlearning. Also, the attack does not require access to the Victim Model or prior knowledge of the Unlearning Algorithm.
  2. The experiments are relatively complete, and the results are convincing. The authors evaluate CoForget on three datasets and three Unlearning Algorithms. They also test the attack across different model architectures. The results show that CoForget achieves a high ASR after unlearning while maintaining relatively high BA. In addition, the ablation studies compare different Trigger designs, Clean Sample selection methods, and unlearning request compositions. These experiments provide evidence that the main components of the method are effective.
Critical weaknesses
  1. The relationship between Clean Sample Selection and benign accuracy is not very clear. The authors use Gradient Alignment to select Clean Samples, but the selection objective mainly focuses on backdoor activation and does not directly consider Benign Accuracy. Although the experiments show that this method can reduce the drop in accuracy, the authors do not fully explain why. Also, adding randomly selected Clean Samples can improve BA as well. Therefore, it is still unclear how much of the improvement comes from the proposed selection strategy.
  2. The experimental scope is relatively limited. The experiments mainly focus on image classification datasets such as CIFAR-10, CIFAR-100, and Tiny-ImageNet. Although the authors test different model architectures and Unlearning Algorithms, it is still unclear whether the method would work as well on more complex models or in real-world Machine Unlearning scenarios. In addition, the attacker needs to contribute a certain portion of the training data, which may limit the practical applicability of the method.
Questions for authors
  1. About Clean Sample Selection and Benign Accuracy. The authors use Gradient Alignment to select Clean Samples, but the selection objective does not directly consider Benign Accuracy. Although the experiments show that this method can reduce the drop in BA, could the authors further explain why this happens?
  2. Why does Gradient Alignment work across different unlearning algorithms? The authors use a first-order approximation based on gradient ascent to select Clean Samples. However, methods like SCRUB use different update rules. Why does this selection strategy still work for these methods?
  3. Limitations in more complex settings. The experiments mainly focus on image classification tasks. Could the authors discuss whether CoForget would work on more complex models or real-world Machine Unlearning services? What challenges might arise in these settings?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 12Hr (Part 1/2): Q1

Thank you for your careful review and constructive feedback. We have revised the manuscript and highlighted the changes in blue in the updated PDF for ease of reference.

Q1. Benign accuracy gains and the contribution of clean selection

About Clean Sample Selection and Benign Accuracy. The authors use Gradient Alignment to select Clean Samples, but the selection objective does not directly consider Benign Accuracy. Although the experiments show that this method can reduce the drop in BA, could the authors further explain why this happens?

Also, adding randomly selected Clean Samples can improve BA as well. Therefore, it is still unclear how much of the improvement comes from the proposed selection strategy.

Response:

Why alignment can benefit BA. CoForget's cumulative gradient alignment targets the activation efficiency of the clean component: the first-order target-label loss reduction per unit parameter displacement. At a fixed shadow reference point, this efficiency is proportional to cosine alignment. Our expanded Appendix B analyzes the combined request direction hh, which includes both camouflage and clean gradients. Its activation efficiency and the corresponding linearized displacement for a first-order target-label loss reduction γ>0\gamma>0 are

\qquad \rho_h^{\mathrm{lin}}=\frac{\gamma}{\mathcal E(h)}.$$ **Proposition 1 (Benign accuracy loss bound).** For two request directions $g_1$ and $g_2$ at the same reference point, if $\mathcal E(g_1)>\mathcal E(g_2)>0$, then achieving the same first-order target-label loss reduction requires a smaller linearized displacement along $g_1$. Let $r>0$ denote the radius of the neighborhood where the margin-stability condition holds. For $0<\gamma\leq r\mathcal E(g_2)$, both linearized displacements remain within this neighborhood, and the corresponding BA-loss upper bounds satisfy $$U\left(\frac{\gamma}{\mathcal E(g_1)}\right) \leq U\left(\frac{\gamma}{\mathcal E(g_2)}\right).$$ The proof is provided in Appendix B. **Thus, when clean selection improves the combined request's activation efficiency, it also provides a theoretical basis for preserving benign prediction stability without explicitly optimizing BA.** **Contribution beyond random clean samples.** On CIFAR-10 under NegGrad+, adding random clean samples improves mean BA by 4.74 percentage points over camouflage-only requests, but reduces mean ASR from 99.98% to 73.81%. **Deliberate selection further improves mean BA by 1.10 percentage points and restores ASR from 73.81% to 98.98%** (Table 4). This additional benefit is consistent across all datasets, with selected clean samples yielding the following mean gains over random selection (Appendix Table 17): | Dataset | Mean BA gain (pp) | Mean ASR gain (pp) | | ------------- | ----------------- | ------------------ | | CIFAR-10 | +1.10 | +25.16 | | CIFAR-100 | +0.80 | +11.15 | | Tiny-ImageNet | +0.88 | +15.55 | Table 4 further confirms higher mean BA and ASR with selected samples under NegGrad+, SAM, and SCRUB on CIFAR-10.
AuthorsOpenReview ↗

Response to Reviewer 12Hr (Part 2/2): Q2–Q3 & W2

Q2. Transfer of the selection proxy across unlearning algorithms

Why does Gradient Alignment remain useful when different unlearning algorithms, such as SCRUB, use different update rules?

Response:

Why gradient alignment transfers across unlearning algorithms. CoForget uses gradient alignment as a request-construction criterion on a shadow model, independently of the specific update rule used by the unlearning algorithm. It selects clean samples whose cumulative CE gradient aligns with the estimated activation direction. The resulting request is then processed using each unlearning method's own forget objective: NegGrad+ performs original-label CE ascent, SAM evaluates its CE-based forget gradient at perturbed parameters, and SCRUB maximizes teacher–student KL divergence on forget samples.

For SCRUB, our implementation uses temperature T=1T=1. With the teacher fixed, maximizing DKL(q∥p)D_{\mathrm{KL}}(q\|p) is equivalent to maximizing CE with the teacher's soft targets. Here, qq and pp are the teacher and student predictive distributions. Let JJ be the Jacobian of the student logits with respect to its parameters and eye_y the one-hot original label. At the same student state, the two forget gradients satisfy

g_{\mathrm{CE}} &= J^\top(p-e_y),\\ g_{\mathrm{KL}} &= J^\top(p-q)\\ &=g_{\mathrm{CE}}+J^\top(e_y-q). \end{aligned}$$ This identity also holds after averaging over a fixed clean subset. It shows that SCRUB's KL-based forget gradient retains the original-label CE gradient as a component, with an additional teacher-dependent correction. **This provides a mathematical connection between the CE-based selection criterion and SCRUB's different forget objective.** **Evidence under the complete unlearning procedures.** Table 1 shows strong activation across three datasets when the same preconstructed trigger and request are processed by NegGrad+, SAM, and SCRUB. Table 4 further shows that, under SCRUB, selected clean samples improve mean BA by 1.07 percentage points and mean ASR by 8.49 percentage points over random selection on CIFAR-10, with the same camouflage component, clean budget, and unlearning protocol. **These results demonstrate that CE-based selection retains its benefit under SCRUB's different forget objective.** --- **Q3 & W2. Experimental scope, real-world applicability, and attacker data requirements** > How well does CoForget generalize beyond image classification, and what practical challenges arise from model complexity, service-level unlearning support, and the attacker's data contribution requirement? **Response:** **Our current evaluation focuses on image classification.** Generalization to other vision tasks and non-vision modalities remains future work. **To examine architectural generalization beyond CNN victims, we evaluate CoForget on ViT-Lite-7/4 across all datasets.** We reuse the trigger and unlearning request constructed with a PreActResNet-18 shadow model. Post-unlearning ASR ranges from 89.31% to 98.25%, with BA decreases of 2.86 to 5.32 percentage points. | Dataset | BA before (%) | ASR before (%) | BA after (%) | ASR after (%) | | ------------- | ------------- | -------------- | ------------ | ------------- | | CIFAR-10 | 87.18 | 7.64 | 84.32 | 92.98 | | CIFAR-100 | 59.30 | 20.12 | 55.90 | 98.25 | | Tiny-ImageNet | 48.28 | 10.12 | 42.96 | 89.31 | **A key deployment challenge is that not all unlearning services support sample-level selective removal.** CoForget requires the service to accept attacker-contributed samples for training and later unlearn selected subsets while retaining the contributor's remaining data. Services that only support contributor-level deletion or complete data removal do not directly meet this requirement. Section 6 now identifies these conditions as practical deployment constraints. **CoForget also requires the attacker to contribute a portion of the training data, which may constrain its applicability in services with restrictive data admission policies.** However, our additional experiments show that effective activation remains achievable with less data than the default 10% budget. At 4% attacker-contributed data, CoForget achieves 92.45% BA and 100.00% ASR after unlearning on CIFAR-10 (Appendix E.5, Table 23).
Reviewer_hyZeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies backdoor attacks in the image domain, more specifically, backdoors that are unlearning-activated. The proposed approach is a request-composition attack that combines a small fixed camouflage component with selected clean samples. The attack first calibrates a universal trigger with a logit gap objective to promote dormancy without excessively suppressing the target response. It then greedily selects a clean subset whose cumulative gradient aligns with an activation direction estimated on a local shadow model. The paper provides experimental evaluation on three datasets, one victim architecture, and three cross-architectures. The paper evaluates the backdoor under three unlearning algorithms, and it also investigates how well the attack works when defenses are deployed. The paper also has a detailed appendix that includes numerous additional experiments.

Critical strengths
  • The topic is interesting, and the paper is well written.

  • The experimental analysis is quite detailed, and the results are good.

Critical weaknesses
  • The novelty element needs more discussion.

  • Not all design elements are sufficiently defined and/or justified.

  • The paper considers simple neural network architectures, and the results require more discussion.

Questions for authors

Please discuss the novelty more. What is the difference from previous works? The paper states "CoForget constructs heterogeneous unlearning requests that combine camouflage samples with a selected clean subset." but it is difficult to properly assess the differences from this. Also, later it mentions it is inspired by the trigger design of Narcissus.

Does the approach work only for approximate unlearning algorithms? Could you please explain the limitation better?

While I consider the attack nice, I would like to understand better what the practical scenario for it is. In what deployment settings would this attack be preferred over standard backdoor attacks?

Equation (7): I do not see what the values for \ lambda are or how those values are selected.

Please discuss the per-class cap q more: how is it selected?

The experiments consider several architectures, but they are all rather simple. Would the approach work for ViT architectures also?

Each unlearning request comprises 0.05% camouflage samples and 0.5% selected clean samples. -> how are these values selected?

Looking at the results, it is not necessarily easy to understand what the advantage of the proposed approach is. It seems the baseline attacks are rather competitive and, in certain aspects, maybe even better? I think this should be better discussed and explained what are the relevance of the results. For instance, the paper states, "Across all datasets, CoForget achieves a post-unlearning ASR exceeding 98.5% while using 1/11 as many camouflage samples as the prior attacks," but is the fact that the proposed approach requires many fewer camouflage samples important (practically)?

"though the prior attacks maintain lower pre-unlearning ASR." I think this is a very important part and potentially a significant weakness of the proposed attack?

In Section D.1, three alternative designs are well-known but rather old. Why not try some state-of-the-art approach?

I did not fully follow Sections E. 1 and E. 2. I would say the budget discussion is important, but the authors do not really provide motivation for the choices there.

5 comments
AuthorsOpenReview ↗

Response to Reviewer hyZe (Part 1/4): Q1–Q3

Thank you for your careful review and constructive feedback. We have revised the manuscript and highlighted the changes in blue in the updated PDF for ease of reference.

Q1. Novelty and relationship to prior work

Response:

Goal and gap in prior designs. Our goal is to improve the activation–utility balance of unlearning-activated backdoors while reducing reliance on specially modified camouflage samples. BAU, UBA-Inf, and ReVeil rely on removable camouflage to activate backdoors, but the resulting updates may degrade benign accuracy. Their camouflage-only requests do not leverage unmodified clean samples to shape the aggregate update.

Our design rationale and novelty. CoForget optimizes the clean component of a heterogeneous request alongside a small fixed camouflage component. Unlike simply adding clean samples or ranking them individually, our greedy selector accounts for gradient interactions by maximizing cumulative alignment with a shadow-model activation direction (Sec. 4.2). Appendix B provides a theoretical foundation: when selection improves the combined request's activation efficiency, Proposition 1 establishes a smaller linearized displacement for the same first-order target label loss reduction, with no increase in the corresponding upper bound on benign accuracy loss under its stated conditions.

Evidence for the design. Comparisons with camouflage-only requests show that incorporating selected clean samples improves BA while maintaining high post-unlearning ASR (Table 4). At matched total request budgets, heterogeneous requests achieve smaller mean BA drops while maintaining high post-unlearning ASR across all three datasets (Table 5). Selected clean samples also outperform random selection (Appendix Table 17).

Relationship to Narcissus. Narcissus uses inward-trigger optimization to strengthen the target class response after training. CoForget addresses a fundamentally different objective: the backdoor remains dormant after training and activates only after unlearning. We draw on Narcissus's inward-trigger idea for trigger initialization, and formulate the subsequent calibration specifically for this different objective. Our calibration applies a soft penalty to the logit gap between the true and target classes, balancing initial dormancy and subsequent activation (Sec. 4.1).


Q2. Applicability to approximate unlearning

Response:

Our work focuses on approximate unlearning. Exact unlearning typically relies on retraining the model or affected components on retained data, as in full retraining and SISA, and can incur substantial computational costs. In contrast, approximate unlearning updates an already trained model to approximate the effect of removing specified samples. CoForget is designed for the latter setting, using a local gradient-ascent proxy on a shadow model to guide clean sample selection.

The extension of CoForget to retraining-based exact unlearning remains future work. We have clarified this scope in the Limitations paragraph of the revised manuscript.


Q3. Deployment scenarios and delayed activation

Response:

CoForget targets user-contributed data services that allow users to request the removal of selected subsets of their training contributions after deployment. For example, an image classification service may train on user-submitted labeled images and support subsequent sample-level unlearning requests (Sec. 3.2).

Unlike conventional backdoors that become activated immediately after training, CoForget delays backdoor activation until a subsequent unlearning request is processed. This distinction is advantageous in services that perform backdoor evaluation or model repair before deployment, as the dormant backdoor may evade initial security checks. Our defense evaluation examines three categories of existing defenses: data detection, request screening, and model repair. The results show that these defenses do not reliably prevent CoForget: most malicious samples evade detection, ASR remains high after request screening, and model repair leaves substantial residual activation (Sec. 5.5 and Appendix F). This exposes a security gap between initial model evaluation and subsequent unlearning operations.

AuthorsOpenReview ↗

Response to Reviewer hyZe (Part 2/4): Q4–Q6

Q4. Calibration weights in Equation (7)

Response:

We use λcal=1.6\lambda_{\mathrm{cal}}=1.6 and λanc=0.3\lambda_{\mathrm{anc}}=0.3 across all three datasets, as reported in Appendix A.2. We determined these shared defaults through empirical evaluation of multiple parameter configurations, jointly considering pre-unlearning ASR, post-unlearning ASR, and benign accuracy.

We also assess sensitivity by halving or doubling λcal\lambda_{\mathrm{cal}} under NegGrad+, keeping the anchor weight fixed (Appendix Table 21). Doubling λcal\lambda_{\mathrm{cal}} on CIFAR-100 substantially increases pre-unlearning ASR, while halving it on Tiny-ImageNet sharply reduces post-unlearning activation. These findings support the default calibration weight as a practical choice for balancing dormancy, activation, and benign accuracy.

Datasetλcal\lambda_{\mathrm{cal}}ASR beforeBA afterASR after
CIFAR-100.815.3790.76100.00
CIFAR-101.615.0492.5099.98
CIFAR-103.220.2792.40100.00
CIFAR-1000.815.5164.5589.42
CIFAR-1001.616.1267.9899.84
CIFAR-1003.250.8268.68100.00
Tiny-ImageNet0.84.9152.645.95
Tiny-ImageNet1.69.8554.0198.55
Tiny-ImageNet3.210.7853.8498.48

Q5. Choice of the per-class cap

Response:

The per-class cap qq limits class concentration in the selected clean subset, requiring at least ⌈kcln/q⌉\lceil k_{\mathrm{cln}}/q\rceil represented classes. A tighter cap encourages broader class coverage but restricts the feasible choices for gradient-aligned selection. We use q=60q=60 for CIFAR-10 and CIFAR-100 and q=20q=20 for Tiny-ImageNet, as listed in the selector configuration table in Appendix A. We determined these defaults through empirical evaluation of multiple cap configurations, jointly considering post-unlearning BA and ASR.

We also compare the default cap with a tighter cap and no cap under NegGrad+, holding the trigger, victim checkpoint, camouflage component, and clean budget fixed. The default caps achieve the highest post-unlearning BA among the evaluated settings while maintaining strong activation across all datasets, supporting their use as empirical defaults.

DatasetPer-class cap qqBA after (%)ASR after (%)
CIFAR-104092.3599.99
CIFAR-106092.5099.98
CIFAR-10No cap92.11100.00
CIFAR-1001067.6899.91
CIFAR-1006067.9899.84
CIFAR-100No cap65.83100.00
Tiny-ImageNet1053.5396.86
Tiny-ImageNet2054.0198.55
Tiny-ImageNetNo cap53.1097.15

Q6. Evaluation on Vision Transformers

Response:

Our additional experiments show that CoForget transfers from a CNN shadow model to a ViT-Lite victim across all datasets. We evaluate ViT-Lite-7/4 [1] under NegGrad+, reusing the trigger and unlearning request constructed with a PreActResNet-18 shadow model. Post-unlearning ASR ranges from 89.31% to 98.25%, while BA decreases by 2.86 to 5.32 percentage points.

DatasetBA before (%)ASR before (%)BA after (%)ASR after (%)
CIFAR-1087.187.6484.3292.98
CIFAR-10059.3020.1255.9098.25
Tiny-ImageNet48.2810.1242.9689.31

Reference

[1] Ali Hassani et al. Escaping the Big Data Paradigm with Compact Transformers. arXiv:2104.05704, 2021.

AuthorsOpenReview ↗

Response to Reviewer hyZe (Part 3/4): Q7 & Q11, Q8

Q7 & Q11. Choice of unlearning request budgets and motivation for the sensitivity analyses

Q7: How were the default camouflage and selected clean sample budgets chosen?

Q11: What motivates the sensitivity analyses in Appendix E.1 and E.2?

Response:

We chose the default budgets of 0.05% camouflage and 0.5% selected clean samples empirically to balance pre-unlearning dormancy, post-unlearning activation, and benign utility. We have expanded the sensitivity analyses in Appendix E.1 and E.2 to all three datasets.

E.1 examines how a fixed request budget should be allocated between camouflage and clean samples. Since camouflage provides suppressive supervision during training while selected clean samples steer the unlearning update toward activation, we vary their proportions while keeping the total budget fixed to assess the effect of request composition (Table 18). The results show that the default compositions maintain a favorable activation–utility balance.

E.2 examines the effect of total request size independently of composition. We scale both components at a fixed 1:10 ratio (Table 19). The results show that larger requests do not consistently improve the activation–utility balance, supporting the default budget as an empirical choice.

We have clarified the motivations for these analyses in the revised Appendix E.


Q8. Advantages over baselines and the value of fewer camouflage samples

How does CoForget improve over competitive baseline attacks, and what is the practical significance of using substantially fewer camouflage samples?

Response:

CoForget offers a more favorable activation–utility balance while requiring substantially fewer modified camouflage samples. Table 2 compares the attacks at the same total request size. On both CIFAR datasets, CoForget better preserves benign accuracy while achieving activation comparable to the baselines. On Tiny-ImageNet, it combines high activation with less BA degradation than ReVeil and BAU; UBA-Inf has a smaller BA drop but substantially lower activation.

CoForget replaces most camouflage samples with selected clean samples, using 0.05% camouflage samples and 0.5% selected clean samples, compared with 0.55% camouflage samples for the baselines. At matched total request budgets, this heterogeneous composition achieves smaller mean BA drops while maintaining high post-unlearning ASR across all datasets (Table 5). The camouflage-only configuration also fails to activate in one of the three CIFAR-10 runs.

DatasetRequestASR before (%)ASR after (%)ΔBA (pp)
CIFAR-10Heterogeneous12.23 ± 3.3298.98 ± 1.751.26 ± 0.31
CIFAR-10Camouflage-only2.48 ± 0.5865.68 ± 56.882.95 ± 0.22
CIFAR-100Heterogeneous19.34 ± 3.6499.23 ± 0.702.43 ± 0.96
CIFAR-100Camouflage-only14.12 ± 1.9299.38 ± 1.074.35 ± 0.92
Tiny-ImageNetHeterogeneous7.15 ± 2.4297.87 ± 1.743.36 ± 0.44
Tiny-ImageNetCamouflage-only3.55 ± 2.2695.41 ± 5.538.01 ± 6.68

Appendix B provides a theoretical rationale for this activation–utility improvement. When clean selection increases the combined request's activation efficiency, the same first-order target-label loss reduction requires a smaller linearized displacement, with no increase in the BA-loss upper bound under Proposition 1's conditions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.