acceptodds
Under review as a conference paper at ICLR 2027

UNISON: Joint Privacy and Backdoor Forgetting for Full-Information Deletion of Poisoned Training Data

Abstract

Machine unlearning aims to remove the influence of selected training examples without retraining from scratch. We study deletion when the examples to be forgotten also taught the model malicious behavior. Backdoor poisoned samples can leave a sample-level privacy trace and a trigger-dependent rule that generalizes beyond the samples themselves. Our experiments show that existing methods do not remove these effects together. Privacy-focused unlearning can reduce membership leakage while the backdoor remains active. Backdoor repair can reduce attack success while membership information remains detectable. Increasing either intervention eventually damages clean accuracy without reliably solving the other objective. We trace this separation to the forgetting gradients. Their overall alignment is low, but a small set of directions repeatedly supports both objectives across batches. We introduce UNISON, which identifies this stable shared structure, preserves the remaining objective-specific directions, suppresses updates that conflict with retained knowledge, and shifts effort toward the objective with the larger residual gap. Across CIFAR-10, CIFAR-100, Tiny ImageNet, and ImageNet-1K with four backdoor attacks and four image classifiers, UNISON reduces mean attack success rate to approximately 5.2% and membership-inference deviation to approximately 0.029, while remaining within 0.9 percentage points of retraining accuracy. These results show that poisoned data deletion requires balancing distinct forms of forgetting rather than optimizing privacy or behavior alone.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Machine Unlearning, Privacy attacks, Backdoor Attacks

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

This paper introduces UNISON, a privacy-oriented machine unlearning technique that targets models that have been trained on poisoned data and removes the global influence of the poisoning attack while keeping performance on the retain set. The authors study this problem in the contained setting of full-information deletion of identified poisoned examples with a known-trigger and restored labels. Their method, UNISON, approaches this problem through balancing objectives of forgetting the trigger and preserving retain set performance. It does this through identifying objective-based salient weights in each layer of the network based on gradient directionality under different loss functions. UNISON is evaluated on a suite of image classification tasks using models from the ResNet family and shows a good balance of performance across retention of performance on the clean data and on privacy metrics.

This paper should be rejected because 1) the setting is not well motivated or realistic, 2) the method has not been well situated within the literature and lacks substantial reference to the broader study of unlearning from poisoned models, 3) I have substantial concerns about the evaluation of this method and do not support a significant contribution.

Critical strengths
  • The UNISON mechanism used to balance the objectives is intuitive and well motivated
  • The authors have provided experimental code for their implementation, enabling open verification of their results.
  • Each claim was well scoped within the text and sentences are easy to read.
Critical weaknesses
  • The work is not well situated within the larger literature. This problem greatly coincides with that of Corrective Unlearning (Goel et al. 2024) and broader discussions of poisoning attacks in machine unlearning (Pawelczyk et al. "Machine unlearning fails to remove data poisoning attacks.").
  • The work is missing key, and more recent, baselines and discussions of unlearning methods in comparison to UNISON which look at subsets of weights when tackling unlearning (eg. SSD - Foster et al. 2023, RELOAD - Newatia et al. 2026). A discussion setting up UNISON in comparison to other methods which use salient weights would be very helpful to identify the strengths of UNISON.
  • The experimental setup covers multiple datasets and does good ablations of the components of the algorithm - but all experiments are performed on ResNet family models which leads me to ask if this method generalises beyond these models
  • On reading through the released code, and running some experiments, I observed that in the main UNISON code the model is left in train mode - which leads me to question whether the authors performed their metric evaluations in eval mode or not as this can cause BN statistics to drift.
  • It is not clear to me which evaluation procedure tests/measures how much impact the poisoned data has on the final model. ASR is undefined, but seems to be the main metric for this scenario and it is not clear why MIA is the right way to evaluate model poisoning.
  • A relatively milder criticism is that I do not see the justification/motivation for this setting. Corrective unlearning defines a substantially harder problem, whereas the setting outlined by the authors in this work assumes access to full identification of poisoned samples (not necessarily the case), a subset of the retain data, knows the mechanism through which the poisoning attack was conducted, and knows the corrected labels for the poisoned data. Furthermore, access to further held-out sets is assumed. Could the authors please provide a setting in which this is realistic?
Questions for authors
  • Could the authors please define the ASR metric? It is not clear whether this comes from prior work or not and how it is formulated
  • Could the authors report results with BN statistics frozen during unlearning or confirm whether this was done, to identify whether the main effect on the ASR is from the UNISON updates or from BN drift?
8 comments
AuthorsOpenReview ↗

Additional Baselines, BatchNorm Controls, and Corrective Unlearning Evaluation

Thank you for taking the time to review our paper and inspect the released code. We appreciate the specific concerns you raised. We have revised the paper and added several experiments to address them. We have provided the PDF with the changes and a diff file too.

  1. Relation to corrective unlearning and existing methods

We agree that the connection to earlier corrective unlearning work needed to be explained more clearly.

We have expanded Section 6 (Related Work) to discuss Corrective Machine Unlearning (Goel et al., 2024), the work of Pawelczyk et al. on residual poisoning effects after unlearning, and parameter-selective approaches such as SSD and RELOAD.

Our setting differs from partial-information corrective unlearning because we focus on identified poisoned examples with a known trigger and corrected labels. This allows us to study a specific question: whether membership leakage and learned backdoor behavior can be removed together, rather than treating them as the same forgetting objective.

  1. Additional baselines

Following your suggestion, we have added SSD and RELOAD to the comparison (Section 4.2 and Appendix G).

SSD and RELOAD achieve ASRs of 48.0% and 32.0%, respectively, compared with 3.2% for UNISON. Their diagnostic membership deviations are 0.050 and 0.038, compared with 0.017 for UNISON. Clean accuracy remains above 91% for all three methods.

These results provide further support for jointly addressing privacy and behavioral forgetting. We have also clarified the different information assumptions of the compared methods.

  1. BatchNorm concern

Thank you for pointing out the possible effect of BatchNorm running statistics. We have added two controls to examine this directly, as you suggested.

With BatchNorm statistics frozen during unlearning, ASR changes only slightly from 3.2% to 3.9%, while clean accuracy changes from 91.9% to 91.7%. The LR and BB membership deviations also remain close, changing from 0.021/0.024 to 0.023/0.026.

In comparison, the BN-only control retains 91.3% ASR. This suggests that the main reduction in attack success comes from UNISON's parameter updates rather than BatchNorm statistic changes alone.

We have included these results in Section 4.4 and Appendix G and clarified the evaluation-mode protocol in Appendix F.

  1. Generalization beyond ResNet

We have added a ViT-Tiny experiment on CIFAR-10 with BadNets (Section 4.4 and Appendix G). UNISON achieves 83.9% clean accuracy and 6.2% ASR, with LR and BB membership deviations of 0.032 and 0.038.

Although retraining performs better, these results show that the joint forgetting approach also works with a transformer architecture.

  1. ASR and membership-inference evaluation

We have clarified the ASR definition and evaluation procedure in Section 4.1, following standard backdoor evaluation practice.

We have also explained why both ASR and MIA are needed. ASR measures whether the learned malicious behavior remains active, while MIA measures whether the forgotten training examples still leave detectable membership information. Since these are different effects of poisoned data, measuring only one would not provide a complete evaluation.

  1. Motivation for the full-information setting

We have expanded Sections 2 and 5 to explain the intended use case more clearly. UNISON targets post-incident model maintenance, where poisoned samples have already been identified and the goal is to remove their privacy and behavioral influence without retraining the model from scratch.

We agree that this does not solve the separate problem of discovering unknown poisoned samples or triggers, and have made this limitation explicit.

Thank you again for the detailed feedback. We believe the additional baselines, BatchNorm controls, transformer evaluation, and clearer positioning have strengthened the paper and addressed the main questions raised by you. We are happy to work on any further suggestions you may have.

Thank you for the updated experiments, implementation investigation, and further details on literature and motivation. This raises a few more questions from me.

Firstly, and I apologise if this is already present in the paper and I have missed it, but have you studied multiple different attacks in conjunction?

Our setting differs from partial-information corrective unlearning because we focus on identified poisoned examples with a known trigger and corrected labels. This allows us to study a specific question: whether membership leakage and learned backdoor behavior can be removed together, rather than treating them as the same forgetting objective.

Could you further elaborate on "whether membership leakage and learned backdoor behavior can be removed together, rather than treating them as the same forgetting objective."?

This ties in to my question about the strong assumptions in the setting you have chosen to study, as I am not convinced by the modifications to Section 2 and 5 about the practicality of the setting. Your updates to Section 5 have stated model maintenance pipelines "in which suspicious training records are identified through data auditing or incident investigation, corrected labels are obtained through verification, and a small trusted dataset is retained for calibration". To me this is unrealistic and does not explain how/when a trigger transformation can be reliably identified.

I think it is important to further evaluate UNISON under partial-information settings closer to corrective unlearning, with stronger ablations to each component of UNISON's setting. In Table 15 the spatial mismatch is not characterised, explained, or quantified. Furthermore an ablation on not identifying all of the poisoned/corrupted data.

Attack Success Rate (ASR)

Thank you for the clarification on this metric. ASR as I understand it is reported as the success rate of attacks from the poisoned data. This raises a few questions, 1) How many attacks were conducted to calculate the percentage? 2) When is an attack considered a success?

Poisoning attacks are not typically a binary measure, and I would encourage the authors to quantify this in a more appropriate manner, unless I have misunderstood how ASR is calculated.

AuthorsOpenReview ↗

Further clarifications

Thank you for your follow-up questions and the opportunity to improve the work. We have revised the manuscript to clarify our evaluation, assumptions, and ASR calculation.

1. Evaluation with multiple poisoning attacks

We evaluate UNISON against four attacks: BadNets, WaNet, ISSBA, and Narcissus, tested separately across four datasets (Table 2, Table 5). We evaluated them separately to isolate their effects and understand whether UNISON can jointly reduce membership leakage and learned backdoor behaviour across different poisoning attacks. As mentioned in the manuscript, our contribution is to address this limitation of existing methods. For example, AMUN can reduce membership leakage while keeping the backdoor behaviour, and RNP can suppress attack success without removing membership information. Although we have not tested simultaneous attacks, results across four families support UNISON's effectiveness in different settings (Section 4.1). We will include this direction in future work.

2. Why privacy forgetting and backdoor forgetting are different

This distinction is a main motivation for UNISON. Privacy forgetting aims to remove information that could reveal whether a sample was used during training, while backdoor forgetting aims to remove the malicious behaviour learned from poisoned samples.

Our experiments show that improving one does not necessarily improve the other. For example, AMUN achieves a diagnostic membership-inference deviation of 0.025 but retains 74.6% ASR. In contrast, RNP reduces ASR to 4.8% but retains a diagnostic membership deviation of 0.163 (Section 4.2, Table 1). We also use independent likelihood-ratio and learned black-box membership attacks to evaluate privacy forgetting (Section 4.1, Appendix C), rather than relying only on the diagnostic loss-threshold metric.

3. Practicality of the setting and trigger information

Our work focuses on post-incident corrective deletion, where an investigation or data audit has already identified suspicious samples, verified their labels, and provided trigger information. The remaining challenge is removing their influence without retraining the entire model. We understand that obtaining this information may be difficult in practice, especially for unknown triggers. Our goal is not to solve poison detection or trigger reconstruction, but to study joint forgetting once the required information is available (Section 2). We have clarified this setting in Sections 2 and 5. We also evaluate imperfect identification, noisy labels, limited retain data, and inaccurate triggers in Section 4.4 and Appendices B and G.

4. Partial-information experiments and trigger mismatch

We have now clarified this in Section 4.4 and Appendix G.2, Table 15. For the 4 × 4 BadNets trigger on 32 × 32 images, shifts of 1, 2, and 4 pixels correspond to IoUs of 0.60, 0.33, and 0.00. ASR increases from 3.2% to 6.5%, 11.8%, and 22.4%, respectively, while clean accuracy remains between 91.2% and 91.9%. This shows that UNISON can still reduce backdoor success with small trigger inaccuracies.

We also evaluate incomplete poison identification (Appendix B.1, Table 6). At 75%, 50%, and 25% recall, ASR increases to 8.9%, 17.6%, and 39.2%, respectively. Additionally, combining 75% recall with a one-pixel trigger shift gives 91.1% clean accuracy and 13.6% ASR, with LR and BB deviations of 0.036 and 0.048 (Section 4.4, Appendix G.3, Table 16).

5. Clarification of ASR and attack success

We have added the mathematical definition and evaluation procedure in Section 4.1. For each targeted backdoor experiment, we apply the trigger to held-out test images. An attack is counted as successful when the model predicts the attacker's target class. Images already belonging to that class are excluded.

For CIFAR-10, we evaluate 9,000 eligible images per model across five independently trained models, giving 45,000 triggered predictions per attack configuration. We report the mean and standard deviation of ASR across the five models. Thus, ASR measures the percentage of successful triggered predictions, not whether an entire poisoning experiment succeeds.

We also include additional behavioural measures (Section 4.4, Appendix G.7, Table 20). UNISON achieves target-class confidence of 0.13 versus 0.08 for retraining and triggered correct-label accuracy of 87.2% versus 90.5%. Together, these results provide a broader understanding of the remaining backdoor behaviour beyond ASR alone.

We truly appreciate the opportunity to improve the work based on your expert and careful comments. These experiments and clarifications clarify our motivation, evaluation, and scope. Our main contribution is showing why privacy and backdoor forgetting, among others, should be considered together. We hope the revised manuscript addresses your key concerns. We welcome any further suggestions and commit to addressing them.

Reviewer_2yHcOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

UNISON addresses deletion of known poisoned training examples by jointly reducing membership leakage and backdoor behavior. It combines privacy and trigger-correction gradients, protects clean utility, and adjusts the balance using held-out signals. Experiments compare it with single-objective unlearning, backdoor repair, and full retraining.

Critical strengths

The paper convincingly distinguishes reducing membership leakage from removing a backdoor: optimizing either objective alone can leave the other failure mode.

Independent membership attacks, ablations, and imperfect poison-identification tests provide useful evidence about both effectiveness and operational limits.

Critical weaknesses

The full-information setup requires identified poisons, a known trigger transformation, restored labels, and clean calibration data. Robustness to an inaccurate or unknown trigger is not tested.

Low attack success and limited membership-inference deviation are useful diagnostics, but do not establish full deletion or equivalence to retraining across untested attacks and behaviors.

Performance deteriorates with poor poison identification or limited retain-data access; moreover, the selected shared coordinates change under function-preserving reparameterization, weakening a literal subspace interpretation.

Questions for authors

How sensitive is forgetting to an approximately specified trigger, not just missed or extra poisoned examples?

How close is the unlearned model to full retraining under additional membership attacks and downstream trigger-conditioned behaviors?

What practical pipeline supplies the poison identities, corrected labels, and clean calibration examples assumed by the method?

3 comments
AuthorsOpenReview ↗

Additional Experiments on Trigger Robustness, Privacy, and Behavioral Forgetting

Thank you for your thoughtful review and for recognizing the value of evaluating privacy and backdoor forgetting together. Your questions helped us identify some useful additions to the evaluation. These are summarized below and included in the updated manuscript attached. We have provided the PDF with the changes and a diff file too.

  1. Sensitivity to inaccurate triggers

We have added a new trigger-mismatch experiment in Section 4.4 and Appendix G. We kept the original poisoning trigger unchanged and shifted the trigger provided during unlearning.

ASR increases from 3.2% with the correct trigger to 6.5%, 11.8%, and 22.4% for shifts of 1, 2, and 4 pixels, respectively. Clean accuracy remains between 91.2% and 91.9%.

These results suggest that UNISON can tolerate small inaccuracies in trigger information, while also showing where its performance begins to decline.

  1. Additional privacy and behavioral evaluation

We have extended the evaluation to better understand how closely UNISON approaches full retraining.

Using the additional RULI membership-inference attack, UNISON achieves an AUC of 0.537, compared with 0.721 for the poisoned model and 0.513 for retraining. This provides further evidence that membership leakage is substantially reduced. Future evaluation could include a few more complementary attacks, such as RMIA (Zarifzadeh et al., 2024), Label-Only MIA (Choquette-Choo et al., 2021), and BASE (Lassila et al., 2025).

We also measured behavior on held-out triggered inputs. UNISON achieves 87.2% correct-label accuracy, compared with 90.5% for retraining. The mean target-class confidence is 0.13 for UNISON and 0.08 for retraining.

These results are included in Section 4.4 and Appendix G. They show substantial forgetting while also making the remaining gap to retraining clearer. We agree that low ASR or membership leakage alone cannot establish complete deletion, and we have added this to the manuscript too.

  1. Practical information requirements

We have expanded Section 5 to explain a possible deployment scenario. UNISON is intended for cases where a data audit or poisoning investigation has already identified the affected samples, corrected labels can be verified, and a small clean dataset is available for calibration.

We have also clarified the limitations related to incomplete poison identification, trigger mismatch, and parameterization.

Thank you again for the constructive questions. We believe these additions give a more complete picture of both the strengths and the limits of UNISON. We are happy to take in any other suggestion for better clarification of the work.

Thank you for the thoughtful rebuttal. The response has addressed my concerns, and I will keep my current score unchanged.

Thank you for your support. We highly appreciate it.

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

This paper studies how to remove the influence of poisoned training data through Machine Unlearning. The authors point out that existing methods usually focus on either privacy forgetting or backdoor removal, but solving one problem does not necessarily solve the other. To address this, they propose UNISON, which aims to remove both membership information and backdoor behavior while maintaining clean accuracy.

The method first calculates privacy and backdoor forgetting gradients and identifies stable shared coordinates across multiple batches. It then separates shared and objective-specific gradients, uses Retain Gradient Projection to protect clean accuracy, and adjusts the update weights based on the remaining gaps of the two objectives.

The experiments cover four datasets, four backdoor attacks, and four image classification models. UNISON achieves low attack success rates and membership leakage while keeping clean accuracy close to full retraining. The main contribution is a joint forgetting method that coordinates privacy and backdoor removal through stable gradient decomposition. However, the method assumes that poisoned samples and their triggers are already known.

Critical strengths
  1. The paper clearly distinguishes two different forgetting objectives. One strength of this paper is that it shows that forgetting poisoned training samples does not necessarily remove the backdoor. In Table 1, AMUN reduces the membership-inference deviation to 0.025, but the ASR remains at 74.6%. RNP reduces the ASR to 4.8%, but the membership-inference deviation is still 0.163. These results show why Privacy Forgetting and Backdoor Forgetting need to be considered together.
  2. The gradient decomposition method is supported by the experiments. Instead of simply adding the two losses together, the authors identify parameter coordinates where the two gradients consistently agree across multiple batches. They also keep the objective-specific update directions. This design follows from their gradient analysis. In the ablation study, removing the shared gradient component increases the ASR from 3.2% to 12.6%. Removing Retain Protection also reduces clean accuracy by 4.1 percentage points. In addition, the authors use independent Membership Inference Attacks to evaluate privacy forgetting rather than relying only on loss-based metrics related to the optimization objective.
Critical weaknesses
  1. The paper is of low quality in presentation. The caption lacks details and the figure occupies too much space.
  2. Limitations of the Full-information Setting. UNISON assumes that the defender already knows which training samples are poisoned, what the Trigger is, and the correct labels. I think these assumptions may be difficult to meet in real-world situations. Although the authors test cases with incomplete information, the ASR increases to 17.6% when Poison Recall drops to 50%. Also, the paper does not test unknown Triggers. This makes me unsure how well the method would work when the defender has less information.
  3. Stable Shared Coordinates depend on model parameterization. The authors select shared coordinates based on gradient magnitude and sign consistency. However, Table 8 shows that even when the model's predictions remain almost unchanged, Channel Permutation and Positive Rescaling reduce the mask overlap to 0.41 and 0.36. Although UNISON still works under these changes, I am not sure how reliable these selected coordinates are across different parameter settings. More experiments or explanations about this would be helpful.
Questions for authors
  1. Can the proposed method generalize to real-world datasets and how can that be applied?
  2. Can the proposed method be applied to structures other than CNN-based ones and how that performance will be like?
2 comments
AuthorsOpenReview ↗

Clarifications on Practical Applicability, Shared Coordinates, and Generalization

Thank you for the detailed review and for recognizing the motivation behind jointly addressing privacy and backdoor forgetting. We have revised the paper to address your concerns and added new experiments to prove the practicality of UNISON. We have provided the PDF with the changes and a diff file too.

  1. Full-information setting and practical applicability

We agree that knowing the exact poisoned samples and trigger is a strong assumption. Our goal is to study what happens after a poisoning incident has been investigated and the affected samples have been identified. We have explained this practical setting more clearly in Sections 2 and 5.

We have also added a trigger-mismatch experiment in Section 4.4 and Appendix G. When the trigger used during unlearning is shifted by 1, 2, and 4 pixels, ASR increases from 3.2% to 6.5%, 11.8%, and 22.4%, respectively. These results show that UNISON can handle small trigger inaccuracies, although larger differences affect performance. At this stage, we have not considered fully unknown triggers and we plan to keep this as future work.

  1. Generalization beyond CNNs

We have added experiments using ViT-Tiny on CIFAR-10 with BadNets (Section 4.4 and Appendix G). UNISON achieves 83.9% clean accuracy and 6.2% ASR, compared with 84.6% and 1.8% for retraining. This shows that the approach also works with transformer architectures.

Our existing ImageNet-1K experiments also provide evidence at a larger dataset scale, although we do not claim that these results establish performance in unrestricted real-world deployments.

  1. Stability of shared coordinates

We agree that the selected coordinates depend on the model's parameterization. Our experiments show that, even when mask overlap decreases to 0.41 and 0.36 under channel permutation and positive rescaling, UNISON still achieves ASRs of 5.1% and 6.4%. We have clarified in Sections 3.2 and 5 that stability refers to agreement across batches under a fixed parameterization, rather than an invariant shared subspace.

  1. Presentation

We have improved the caption and reduced the size of Figure 1 to make the method easier to follow.

We hope these revisions make the scope of UNISON and the supporting evidence clearer. We are happy to take on any further feedback to help improve the clarity of the work. Thank you again for the suggestions.

Thank you again for your review. We wanted to share the changes made after all three reviews.

The other reviewers recognised the difference between privacy and backdoor forgetting, our gradient-based method and ablations, independent membership tests, and released code. We added experiments and improved the writing, as shown in the revised paper and diff. The following is a short summary of the changes made:

1. Why both types of forgetting matter

Backdoor removal alone does not meet the privacy need to forget compromised training records. For example, an attacker may modify private patient images used to train a hospital model. If these images are later deleted, the hospital may need to remove information revealing whether they were used in training and stop the trigger from affecting new images. The patients are not the attackers.

Our results show why this matters. AMUN reduces diagnostic membership deviation to 0.025 but retains 74.6% ASR. RNP lowers ASR to 4.8% but leaves 0.163 membership deviation. UNISON achieves 3.2% ASR and 0.017 deviation with 91.9% clean accuracy (Table 1).

The main contribution of UNISON is showing that privacy and backdoor forgetting need different updates, but some useful gradient directions are shared. Rather than simply combining two losses, UNISON identifies these shared directions and keeps the updates. Removing the shared component raises ASR from 3.2% to 12.6%. Removing retain protection reduces clean accuracy by 4.1 points. We clarify this in Sections 2, 3, and 6.

2. Practical assumptions and incomplete information

We agree that finding poisoned samples, checking labels, and recovering triggers is difficult. UNISON studies forgetting once this information is available, not poison detection or unknown trigger recovery.

With 75%, 50%, and 25% poison recall, ASR is 8.9%, 17.6%, and 39.2%. With 75% identification precision, ASR is 10.8%, and with 10% label noise, 5.6%.

Shifting the assumed trigger by one, two, and four pixels gives 6.5%, 11.8%, and 22.4% ASR. We added details on how much these shifts overlap the real trigger.

With 75% recall and a one-pixel shift together, UNISON achieves 91.1% clean accuracy, 13.6% ASR, and LR/BB deviations of 0.036/0.048. The poisoned model has 98.4% ASR. These results show what UNISON can handle and where it struggles when information is incomplete (Section 4.4, Appendices B and G).

3. Different models, datasets, and attacks

To test beyond ResNets, we added ViT-Tiny on CIFAR-10 with BadNets. UNISON achieves 83.9% clean accuracy and 6.2% ASR, versus 84.6% and 1.8% for retraining. LR/BB deviations are 0.032/0.038.

On ImageNet-1K, UNISON achieves 74.8% average clean accuracy and 5.6% ASR across attacks, versus 75.2% and 0.2% for retraining.

Together with CIFAR-10, CIFAR-100, Tiny ImageNet, and four attacks (BadNets, WaNet, ISSBA, Narcissus), these results show that UNISON is not limited to one dataset size or poisoning method (Section 4).

4. Stability of shared coordinates

We agree that selected coordinates depend on parameterization. Channel permutation and positive rescaling reduce mask overlap to 0.41 and 0.36, but UNISON still achieves 5.1% and 6.4% ASR. We now clarify that stability means agreement across batches with fixed parameterization, not identical coordinates across equivalent models (Sections 3.2 and 5).

5. Additional experiments after the reviews

We added SSD and RELOAD baselines. They achieve 48.0% and 32.0% ASR, versus 3.2% for UNISON. Their diagnostic membership deviations are 0.050 and 0.038, versus 0.017.

We checked BatchNorm statistics. Freezing them gives 3.9% ASR instead of 3.2%. Changing only these statistics leaves 91.3% ASR. We clarified that all evaluations use eval mode.

We added the RULI membership test. UNISON achieves AUC 0.537, versus 0.721 for the poisoned model and 0.513 for retraining.

On held-out triggered images, UNISON achieves 87.2% correct-label accuracy versus 90.5% for retraining. Mean confidence in the attacker's target class is 0.13 versus 0.08.

We also clarified ASR calculation using 9,000 eligible triggered CIFAR-10 images per model across five independently trained models (Section 4, Appendices F and G).

6. Writing improvements and limitations

We revised Section 2 to explain the need for both objectives, Section 3 to clarify the gradient steps, Section 4 to explain metrics and experiments, and Section 5 to discuss practical assumptions and limits. Section 6 now compares UNISON with corrective unlearning, SSD, and RELOAD. We improved Figure 1 and its caption and added appendix details. The diff shows these changes.

We believe joint privacy and backdoor forgetting is an important new research direction when compromised private records need deletion. Existing methods do not fully address both needs, and UNISON offers a method designed for them.

Thank you for your feedback. We hope these changes address your concerns and welcome further suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.