acceptodds
Under review as a conference paper at ICLR 2027

DIMU: Importance-Guided Feature Masking with Self-Distillation for Unlearning

Abstract

Machine unlearning seeks to suppress or remove the influence of designated training data while preserving utility on retained data. We propose Distillation- Regularized Importance-Masked Unlearning (DIMU), an approximate unlearn- ing framework that combines importance-guided feature-channel masking with self-knowledge distillation. DIMU suppresses forget-sensitive channels in the effective penultimate representation exposed to the classifier and adapts only the classifier head while keeping the feature extractor frozen. Self-knowledge distillation regularizes retained predictions using cached pre-unlearning logits. We evaluate class-level unlearning on CIFAR-10/100 across convolutional and transformer-based architectures, random sample-level unlearning on CIFAR-10, and large-scale class-level unlearning on ImageNet-1K using ResNet-50. DIMU achieves zero forgotten-class test accuracy in the primary single-class benchmarks while preserving strong retained utility, and shows favorable forgetting–retention trade-offs under sample-level deletion. It also exhibits competitive empirical membership-distinguishability behavior and a 24.89× speedup over full retraining on ImageNet-1K.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Machine Unlearning, Selective Forgetting, Importance-Guided Feature Masking, Self-Knowledge Distillation, Representation-Aware Unlearning, Approximate Unlearning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Tbz3OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper propose the novel approximate unlearning method DIMU (Distillation-Regularized Importance-Masked Unlearning). Pointing out the problems of existing approximate unlearning that risk leaving residual information, this study seeks to address these issues by utilizing self knowledge distillation. DIMU consists of CRM feature-channel masking and self knowledge distillation. For the efficiency, they freeze the feature extractor and mask the forget sensitive channel. With the various dataset and model architecture experiments, DIMU shows that performance and speed effectiveness.

Critical strengths
  • The motivation and approach is very intuitive and clear.
  • Experiments are conducted extensively.
  • In particular, the control experiments used to explain each component of the algorithm are well organized, making the presentation highly effective.
Critical weaknesses
  • Although authors emphasize the importance of removing the influence of forgotten data in the introduction seciton, DIMU freezes the feature extractor, leaving the underlying representations unchanged. This raises a fundatmental concern about whether DIMU actually removes the learned information or merely suppress its contribution to the predictions. I think this is very important issue in machine unlearning.
  • In sample-wise unlearning setting, authors interpret the discrepancy between DIMU and retrain model as targeted suppression rather than retraining-equivalent forgetting. However, this interpretation seems to deviate significantly from the fundamental objective of machine unleanring. Although such an evaluation perspective may be understandable from a practical prespective, suppression-based forgetting is arugably more acceptable in class/concept-wise unlearning. In sample-wise, preserving generalization while selectively removing the influence of samples is particularly important. That's why we use MIA and relearning to evaluate the unleanring method.
  • It is encouraging that the authors consider relearning attacks. Evaluating unlearning from the perspective of relearning, in addition to MIA, is important. However, rather than simply comparing DIMU with retrain, it would be more informative to investigate how resistant DIMU is to relearning compared with other approximate unlearning. In particular, I suspect that attack such as the Prototypical Relearning Attack proposed in [1] could be effective against DIMU, given its frozen feature extractor.
  • It remains unclear whether DIMU offers a significant computational advantage over other approximate unlearning methods. I wonder why Boundary Unlearning [2], which achieves efficient forgetting by modifying decision boundary, is not included as a baseline. A comparison with this approach would help better establish the computational efficiency and advantage of DIMU.

[1] Ha, S., Park, S. & Yoon, S.W.. (2026). Unlearning’s Blind Spots: Over-Unlearning and Prototypical Relearning Attack. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:39108-39126 [2] Min Chen, Weizhuo Gao, Gaoyang Liu, Kai Peng, Chen Wang; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 7766-7775

Questions for authors

See the weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer Tbz3

CW1: Does DIMU genuinely remove forgotten information or merely suppress its predictive influence while preserving the underlying representations?

We thank the reviewer for highlighting this fundamental distinction between suppressing the predictive influence of forgotten data and removing information from the underlying learned representations. We agree that these are not equivalent. Canonical DIMU does not claim complete information erasure from the frozen backbone. Rather, it is designed as an approximate unlearning approach that attenuates forget-associated contributions to deployed predictions through a fixed, importance-guided CRM mask.

Specifically, DIMU identifies forget-sensitive penultimate channels using gradients computed on the forget set, suppresses their contribution through the fixed CRM mask, and adapts the classifier head using forgetting, retention, and self-knowledge-distillation objectives. Therefore, although the raw backbone representation remains unchanged, the CRM-masked effective representation supplied to the updated classifier and its predictive behavior are modified. This distinction is central to our formulation and is not intended to imply that forgotten information has been physically removed from the underlying feature extractor.

Importantly, we directly evaluate this limitation using adaptive representation-level attacks. On CIFAR-10 (Table 3), canonical DIMU exhibits a raw-feature MIA AUROC* of 0.989, compared with 0.672 for the deployed CRM-masked effective features. We additionally investigate DIMU-L4, which updates the final residual stage and classifier head while retaining the fixed CRM mask. DIMU-L4 further reduces raw-feature and CRM-masked-feature AUROC* to 0.616 and 0.549, respectively, while achieving 0.00% forgotten-class test accuracy and 95.40% retained-class accuracy. Appendix Table 12 confirms similar improvements on CIFAR-100, where DIMU-L4 achieves raw-feature and CRM-masked-feature AUROC* of 0.620 and 0.550.

These results distinguish the predictive suppression achieved by canonical DIMU from the stronger representation adaptation offered by DIMU-L4. We emphasize that reduced empirical distinguishability does not constitute a guarantee of complete information removal. We have further clarified this scope in the Introduction and Section 6, positioning DIMU as an efficient approximate unlearning framework rather than a certified representation-erasure method.

AuthorsOpenReview ↗

Response to Reviewer Tbz3

CW2: Sample-level suppression may cause over-unlearning rather than selective forgetting, raising concerns about generalization, MIA, and relearning robustness.

We thank the reviewer for highlighting the important distinction between sample-level forgetting and over-unlearning. We agree that random sample-level deletion requires a different interpretation from class-level forgetting. When individual training samples are removed, other examples from the same classes remain in the retain set. Consequently, a correctly retrained model may continue to classify removed samples accurately, despite their exclusion from training.

Our manuscript explicitly examines this distinction in Section 5.3 (Sample-Level Unlearning, Table 5). Using identical forget subsets for DIMU and Retrain across 1%, 5%, and 10% random deletion on CIFAR-10, Retrain achieves 91.19–91.87% accuracy on removed samples, whereas DIMU achieves 65.89–68.47%. DIMU also produces larger forgotten-sample confidence reductions (33.13–36.62 percentage points) than Retrain (5.23–6.39 percentage points), while maintaining 96.79–97.28% retained accuracy and 88.26–90.42% overall test accuracy. We emphasize that lower removed-sample accuracy and confidence do not necessarily indicate stronger instance-level information removal. Instead, these differences may reflect targeted predictive suppression or over-unlearning.

To characterize this discrepancy beyond accuracy, Appendix B (Table 23, Comparison with Retrain) reports additional predictive and representation-level analyses. Symmetric KL divergence relative to Retrain is substantially smaller on retained samples (0.189–0.366) than on forgotten samples (1.818–2.177). Representation CKA is likewise higher on retained samples (0.807–0.854) than on forgotten samples (0.740–0.762). Meanwhile, overall test-set confusion-matrix similarity remains 0.998–0.999. These findings demonstrate stronger preservation of retained behavior while revealing substantial differences in forgotten-sample predictions. They do not establish retraining-equivalent instance deletion or exclude localized collateral damage to retained samples.

We also appreciate the reviewer's reference to Ha et al. (2026). Their over-unlearning metric, OU@ε\mathrm{OU}@\varepsilon (Over-Unlearning at ε), is particularly relevant because it evaluates collateral damage near the forget set, which global retained accuracy may not capture. Our existing evaluations do not directly measure OU@ε\mathrm{OU}@\varepsilon, and we therefore do not claim that DIMU eliminates localized over-unlearning.

Accordingly, we clarify that class-level forgetting is DIMU's primary demonstrated strength, while the sample-level experiments provide complementary evidence of targeted suppression, retained utility, and limitations relative to Retrain. We acknowledge that stronger instance-level privacy and localized over-unlearning evaluations are needed before claiming retraining-equivalent sample deletion. This distinction is consistent with the approximate-unlearning scope and limitations discussed in Section 6 (Limitations and Future Directions).

AuthorsOpenReview ↗

Response to Reviewer Tbz3

CW3: DIMU's resistance to Prototypical Relearning Attacks should be evaluated against other approximate unlearning methods, not just Retrain.

We thank the reviewer for emphasizing the importance of evaluating resistance to stronger relearning attacks, particularly those exploiting feature representations preserved by a frozen backbone. We agree that comparisons with Retrain alone are insufficient and that representative approximate unlearning methods should also be considered.

Our submitted manuscript already evaluates one-epoch forgotten-class relearning (Appendix Table 18), where forgotten-class accuracy increases by 22.40 percentage points (pp) for DIMU, compared with 9.35 pp for Retrain and 41.37 pp for Head-only Retain-FT. However, prototype-based attacks provide a complementary evaluation of forgotten-class recoverability. To investigate this concern, we additionally evaluate the Prototypical Relearning Attack (PRA) proposed by Ha et al. (ICML 2026) on CIFAR-100 using ResNet-18, with class 0 as the unlearning target. We compare Retrain, Head-only Retain-FT, DELETE, DIMU, and DIMU-L4 using 1, 5, and 10 forgotten-class support samples.

The results show that DIMU and DIMU-L4 exhibit lower forgotten-class recovery than Retrain and the evaluated approximate methods across all three support sizes. Under five-shot PRA, DIMU achieves 26.00% forgotten-class accuracy, compared with 30.00% for Retrain, 49.00% for DELETE, and 58.00% for Head-only Retain-FT. DIMU-L4 further reduces recovery to 17.00%. This trend persists under ten-shot PRA, where DIMU and DIMU-L4 obtain 38.00% and 28.00%, respectively, compared with 43.10% for Retrain. Importantly, retained-class test accuracy after five-shot PRA remains 77.94% for DIMU and 77.35% for DIMU-L4. These findings provide additional evidence that DIMU's importance-guided CRM masking and classifier adaptation can limit recovery under the evaluated prototype-based attack. The lower recovery observed for DIMU-L4 further suggests that limited backbone adaptation can strengthen resistance in this setting.

Table 28. Prototypical Relearning Attack (PRA) on CIFAR-100 using ResNet-18. Forgotten-class test accuracy before and after PRA with different support sizes, together with retained-class test accuracy after five-shot PRA. All values are percentages. Lower post-attack forgotten-class accuracy indicates stronger resistance to the evaluated attack.

MethodAccft\mathrm{Acc}_{ft} Before ↓Accft\mathrm{Acc}_{ft} PRA-1 ↓Accft\mathrm{Acc}_{ft} PRA-5 ↓Accft\mathrm{Acc}_{ft} PRA-10 ↓Accrt\mathrm{Acc}_{rt} PRA-5 ↑
Retrain0.0012.5030.0043.1078.00
Head-only Retain-FT12.7527.1058.0072.0077.00
DELETE0.0019.5049.0064.0076.50
DIMU0.0011.0026.0038.0077.94
DIMU-L40.008.5017.0028.0077.35

PRA and the existing one-epoch fine-tuning experiment evaluate distinct relearning procedures, providing complementary evidence of post-unlearning recoverability. The observed improvements demonstrate resistance under the evaluated PRA settings rather than guaranteeing robustness against all possible relearning attacks. We have incorporated the PRA analysis into Appendix Table 28 and summarized the findings in Section 5.3.

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

The paper is presenting a method for machine unlearning. The method works by identifying channels that are important for the forget data, and then it suppresses these channels using a masking approach when learning the unlearned model. The final loss for the retrain algorithm is then a combination of knowledge distillation (using KL divergence of logits), cross entropy of the forget set and cross entropy on the retain set. The unlearning algorithm is only adjusting the final classification layer in the entire model and then leaves all previous layers untouched.

The method is evaluated on CIFAR-10, CIFAR-100, and ImageNet-1K, and both class-level unlearning and sample-level unlearning is considered. DIMU performs well compared to the reference methods, and from this perspective is a viable method.

Critical strengths

The method is well motivated and performs well according the the baselines.

Critical weaknesses

The main critical weakness in this paper is (unfortunately) the design choice of keeping entire neural network frozen except the last fully connected layer. Essentially it means that a logistic regression is being altered. The authors acknowledge this in their limitations section by writing "information associated with forgotten data can remain accessible in the raw representation even when its contribution to deployed predictions". This means that the method is not really an unlearning methods that can be compatible with ordinary AI regulations since representations are not erased.

Questions for authors

Could the authors carefully explain the utility of their method from a real-life perspective? When would I use this method and be able to ignore the fact that the learned representations of data is kept within the model?

3 comments
AuthorsOpenReview ↗

Response to Reviewer VSUy

CW1: Freezing the backbone leaves forgotten representations intact, raising concerns about genuine unlearning and regulatory compliance.

We thank the reviewer for raising this important concern regarding the frozen-backbone design and residual information in learned representations. We acknowledge that canonical DIMU does not modify the underlying feature extractor and therefore cannot guarantee complete removal of forgotten information from raw representations, as discussed in Section 6.

However, DIMU differs from conventional classifier-head fine-tuning. It identifies forget-sensitive channels using gradient-based importance estimation, suppresses their contribution through a fixed CRM mask applied to the penultimate representation, and adapts the classifier head using forgetting, retention, and self-knowledge-distillation objectives. Appendix Table 17 further shows that head-only Retain-FT retains 12.75% forgotten-class accuracy, whereas DIMU achieves 0.00%, demonstrating the contribution of the importance-guided CRM-masked adaptation beyond ordinary head-only fine-tuning.

Importantly, DIMU-L4 extends this framework by updating the final residual stage alongside the classifier head while maintaining the fixed CRM mask. On CIFAR-10 (Table 3), DIMU-L4 reduces raw-feature MIA AUROC* from 0.989 to 0.616 and CRM-masked-feature AUROC* from 0.672 to 0.549, while maintaining 0.00% forgotten-class accuracy and 95.40% retained-class accuracy. Similar improvements are observed on CIFAR-100 (Appendix Table 12), where raw-feature AUROC* decreases from 0.960 to 0.620 and CRM-masked-feature AUROC* from 0.660 to 0.550, with 77.85% retained-class accuracy.

These findings demonstrate that the DIMU framework can incorporate limited backbone adaptation to reduce residual feature-level distinguishability. Nevertheless, we interpret these results as empirical privacy diagnostics rather than guarantees of complete information erasure or regulatory compliance. We have further clarified this distinction in Section 6.

AuthorsOpenReview ↗

Response to Reviewer VSUy

Q1: What are DIMU's practical applications, and when is retaining forgotten information in the backbone acceptable?

We thank the reviewer for requesting clarification of DIMU's practical utility and deployment assumptions. We agree that residual information in frozen backbone representations cannot be ignored when complete removal of sensitive training information is required. Canonical DIMU instead targets efficient approximate unlearning in controlled inference systems where predictions are generated through the fixed CRM mask and updated classifier head, without exposing raw backbone representations or model parameters.

Real-world computer-vision applications provide practical motivation for this setting. For example, industrial visual inspection systems may need to suppress obsolete defect categories, retail product-recognition systems may need to stop predicting discontinued products, and automated recycling systems may need to retire unsupported material categories while preserving recognition of retained classes. These scenarios illustrate potential applications of DIMU rather than existing deployments of the method.

In such settings, DIMU provides a mechanism for suppressing targeted predictive contributions through forget-set gradient-based importance estimation, fixed CRM masking of the effective penultimate representation, and classifier-head adaptation regularized by self-knowledge distillation. This approach is particularly relevant when full retraining is computationally expensive. On ImageNet-1K using ResNet-50, DIMU requires 1.5563 hours compared with 38.7310 hours for full Retrain, corresponding to a 24.89× speedup under the evaluated training schedules (Appendix Table 16).

The key deployment assumption is that the CRM mask remains enforced during inference and that users cannot directly access the raw backbone features or model parameters. Under these conditions, DIMU can be considered when the objective is to suppress targeted predictions while maintaining retained-class utility, rather than to guarantee erasure of all underlying information. Our adaptive privacy analysis explicitly demonstrates this distinction. On CIFAR-10, canonical DIMU exhibits raw-feature AUROC* of 0.989, which decreases to 0.672 when evaluating the CRM-masked representation (Table 3). DIMU-L4 further reduces raw-feature and masked-feature AUROC* to 0.616 and 0.549, respectively. Similar behavior is observed on CIFAR-100 (Appendix Table 12).

We emphasize that enforcing the CRM mask does not eliminate residual information or guarantee protection against all adversarial attacks. Accordingly, canonical DIMU is intended for controlled deployments where approximate predictive suppression is sufficient and the residual representation risk is acceptable under the application's requirements. For applications requiring verified representation erasure, unrestricted model access, or strict data-removal guarantees, stronger unlearning procedures or full retraining may be necessary.

AuthorsOpenReview ↗

Thank You for Your Insightful Feedback on DIMU's Scope and Limitations

We sincerely thank the reviewer for the valuable and constructive feedback. The reviewer's concerns helped us clarify the distinction between predictive suppression and complete representation erasure, as well as the practical scope of DIMU. We have strengthened the discussion of residual information risks, incorporated the DIMU-L4 representation-level evaluation, and clarified deployment assumptions and limitations in the revised manuscript. We hope these analyses and clarifications provide a more balanced understanding of DIMU's contributions and address the reviewer's concerns. We would be grateful if the reviewer would kindly consider updating the assessment.

Reviewer_7zKhOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
2(+1)
Read review
Paper summary

The paper proposed an unlearning method for classifiers that is applied to the feature representation right before the classifier head. The main idea is to rank the feature channels using forget-set gradients and then update the classifier head by masking the selected channels. Several experimental cases such as class unlearning and sample unlearning, are presented.

Critical strengths

The paper is well-structured and the methodology is presented clearly. In the experiments, the settings including the dataset, backbones, and metrics, are fairly presented and discussed.

Critical weaknesses
  • As mentioned in the paper, the proposed method is applied just to the representation right before the classifier head and so adapts the classifier head while keeping the rest frozen. This is a big limitation as the information regarding the forget set could exist in the frozen part and relearning methods like the one presented in "Source-Free Class Relearning: Diagnosing Forgetting in Class Unlearning by Dehghani et al." may retrieve this information. In addition, due to this, talking about privacy and MIA is unrelated.
  • In Appendix A.2 it is mentioned that the hyperparameters were selected using the CIFAR-10/100 test splits... . This means the results for the CIFAR-10/100 are biased.
  • The sample-level unlearning is not a good way to evaluate the unlearning method as it targets the generalizability of the model and not the erasing targeted sample(s) effects. This can be seen in the presented results (even the Retain model) as one has to degrade retain accuracy to degrade forget. One substitute would be feature unlearning.

Minor:

  • In Table 2, for the 10 classes, looking at the H-Mean values, the DIMU should not be boldface as it is not the best one in that list (considering the Retrain model).
  • Figure 1 is presented as the DIMU framework but it lacks details of the proposed method. It is too generic.
Questions for authors

Please see the critical weaknesses. My main question is how the results would change if the proposed method is evaluated on unseen data (and not the data used for hyperparameter tuning).

6 comments
AuthorsOpenReview ↗

Response to Reviewer 7zKh

CW1: Frozen backbone may retain recoverable forgotten information, raising concerns about relearning attacks and MIA validity.

We thank the reviewer for raising this important concern. We agree that canonical DIMU's frozen backbone may retain forgotten-class information. DIMU targets approximate unlearning through importance-guided suppression of the effective penultimate representation, rather than complete erasure of underlying features. This limitation is explicitly acknowledged in Section 6.

Source-free relearning evaluation. To directly address the concern, we conducted an additional Source-Free Relearning Audit (SFRA) on CIFAR-100 using ResNet-18, forget class 0, using the same model checkpoints as those evaluated in the submitted manuscript.

Table 25: Source-Free Class Relearning Audit (SFRA) on CIFAR-100 using ResNet-18 with forget class 0. Accuracies are percentages. Superscripts pre and post denote evaluations before and after relearning. RS measures forgotten-class recovery while accounting for retained-accuracy preservation. ΔRS denotes the difference in RS relative to Retrain. Lower RS and ΔRS indicate lower absolute and relative recoverability, respectively, under the evaluated attack. The deployed CRM mask is enforced for DIMU and DIMU-L4.

MethodAccftpre\mathrm{Acc}_{ft}^{\mathrm{pre}} ↓Accrtpre\mathrm{Acc}_{rt}^{\mathrm{pre}} ↑Accftpost\mathrm{Acc}_{ft}^{\mathrm{post}} ↓Accrtpost\mathrm{Acc}_{rt}^{\mathrm{post}} ↑RS ↓ΔRS ↓
Retrain0.0078.5320.0078.150.3330.000
Head-only Retain-FT12.7577.5051.5077.000.558+0.225
DELETE0.0077.0818.0077.650.305−0.028
DIMU0.0078.3810.0078.950.182−0.151
DIMU-L40.0077.8513.0076.900.230−0.103

DIMU achieves a lower RS than Retrain (0.182 vs. 0.333), with ΔRS = −0.151. DIMU-L4 also shows lower recoverability (RS = 0.230, ΔRS = −0.103). These results indicate reduced forgotten-class recoverability relative to Retrain under the evaluated source-free attack, without implying complete information erasure or resistance to unrestricted raw-feature access.

Relevance of privacy and MIA. We agree that MIA cannot establish complete forgetting. Nevertheless, it remains relevant as a complementary empirical measure of membership distinguishability. The submitted manuscript separately evaluates output-level, raw-feature, and masked-feature attacks (Tables 3 and 12). On CIFAR-100, Table 12 reports a raw-feature MIA AUROC* of 0.960 for DIMU, indicating substantial residual membership distinguishability, whereas DIMU-L4 reduces this to 0.620. These results are explicitly interpreted as empirical diagnostics rather than formal privacy guarantees.

We have incorporated the SFRA evaluation into Appendix B.5 and clarified the distinction between deployed prediction suppression, residual representation information, and privacy guarantees.

AuthorsOpenReview ↗

Response to Reviewer 7zKh

CW2: Hyperparameter selection using CIFAR-10/100 test splits may bias the reported results.

We thank the reviewer for identifying this concern. We agree that selecting hyperparameters using the official test sets would compromise the independence of the reported evaluation. Upon reviewing our implementation and experimental procedure, we identified an incorrect description in Appendix A.2, for which we apologize. For CIFAR-10/100, the official training and test partitions were kept separate, and test-set results did not influence the selection of the mask ratio (τ), distillation temperature (T), or loss weights (λf\lambda_f, λr\lambda_r, λSKD\lambda_{\mathrm{SKD}}). The test sets were used for evaluation rather than hyperparameter selection. We have corrected the inaccurate statement in Appendix A.2 of the revised manuscript to clarify the actual hyperparameter-selection and evaluation procedures. The reported experimental results remain unchanged.

AuthorsOpenReview ↗

Response to Reviewer 7zKh

CW3: Sample-level unlearning may measure reduced generalization rather than true sample forgetting, motivating feature-level evaluation.

We thank the reviewer for suggesting feature-level unlearning as an additional evaluation of DIMU. To examine whether DIMU can suppress a targeted feature association while preserving unrelated predictive information, we designed a controlled experiment on CIFAR-10 using ResNet-18. Specifically, a fixed 4×4 colored patch is associated with a designated class, and feature unlearning is evaluated by comparing Original, a clean-data Retrain reference, and a feature-level adaptation of DIMU. The protocol uses three random seeds and evaluates target-association strength, feature leakage, retained prediction divergence, retained accuracy, and test accuracy.

Table 26: Feature-unlearning results on CIFAR-10 (ResNet-18).

MethodTarget Assoc. ↓ (pp)Feature Leakage AUC ↓Sym. KLr\mathrm{Sym.\,KL}_{r} ↓Accr\mathrm{Acc}_{r} ↑Acctest\mathrm{Acc}_{\mathrm{test}} ↑
Original32.0 ± 1.2 0.900 ± 0.012 0.120 ± 0.015 98.0 ± 0.2 91.0 ± 0.3
Retrain0.5 ± 0.2 0.530 ± 0.010 0.000 ± 0.000 98.3 ± 0.2 91.5 ± 0.2
DIMU4.0 ± 0.5 0.580 ± 0.015 0.210 ± 0.020 97.1 ± 0.3 89.9 ± 0.3

These results indicate that DIMU substantially attenuates the targeted patch–class association, reducing its strength from 32.0 percentage points in Original to 4.0 points, approaching the 0.5-point Retrain reference. Feature leakage similarly decreases from 0.90 to 0.58 AUROC, approaching the Retrain value of 0.53. Meanwhile, DIMU preserves 97.1% retained accuracy and 89.9% test accuracy, compared with 98.3% and 91.5% for Retrain, respectively. The retained prediction divergence of 0.21 indicates a remaining discrepancy relative to Retrain, highlighting the trade-off between targeted feature suppression and preservation of predictive behavior.

This controlled evaluation is intended to complement our existing class-level and sample-level experiments by examining feature-association attenuation directly. We emphasize that the experiment uses a feature-specific adaptation of DIMU rather than the unchanged sample-level procedure, and therefore interpret it as evidence of framework extensibility rather than proof of exact retraining-equivalent unlearning.

Additional Sample-Specific Forgetting Audit (Table 27). To further address the concern that reduced forgotten-sample accuracy alone does not establish sample-specific unlearning, we conducted an additional audit on CIFAR-10 using ResNet-18 at 1%, 5%, and 10% deletion ratios. Table 27 evaluates symmetric KL divergence relative to Retrain on the complete forget set, highly memorized forgotten samples, and matched retained samples. DIMU reduces divergence on highly memorized forgotten samples compared with Original, from 2.000 to 1.150, 2.300 to 1.250, and 2.500 to 1.350 across the three deletion ratios, respectively. Meanwhile, matched-retain divergence remains low (0.190–0.260), with retained accuracy of 96.79–97.28%. Although DIMU remains more divergent from Retrain on the complete forget set (1.818–2.177 versus 0.820–1.100 for Original), these results provide additional evidence of selective attenuation on highly memorized forgotten samples while preserving substantial retained utility. We therefore distinguish sample-specific predictive attenuation from exact retraining-equivalent forgetting.

Table 27. Sample-Specific Forgetting Audit on CIFAR-10 Using ResNet-18

RatioMethodSym. KLf\mathrm{Sym.\,KL}_{f} ↓Sym. KLfhigh\mathrm{Sym.\,KL}_{f}^{\mathrm{high}} ↓Sym. KLrmatch\mathrm{Sym.\,KL}_{r}^{\mathrm{match}} ↓Accr\mathrm{Acc}_{r} ↑Acctest\mathrm{Acc}_{\mathrm{test}} ↑
1%Original0.820 ± 0.070 2.000 ± 0.160 0.180 ± 0.020 100.00 ± 0.00 95.25 ± 0.00
Retrain0.000 ± 0.000 0.000 ± 0.000 0.000 ± 0.000 99.80 ± 0.07 91.55 ± 0.12
DIMU1.818 ± 0.081 1.150 ± 0.140 0.190 ± 0.030 97.28 ± 0.30 90.42 ± 0.13
5%Original0.960 ± 0.090 2.300 ± 0.220 0.160 ± 0.020 99.98 ± 0.00 95.27 ± 0.00
Retrain0.000 ± 0.000 0.000 ± 0.000 0.000 ± 0.000 99.73 ± 0.09 91.27 ± 0.19
DIMU2.045 ± 0.119 1.250 ± 0.160 0.230 ± 0.030 96.93 ± 0.12 89.27 ± 0.19
10%Original1.100 ± 0.090 2.500 ± 0.200 0.220 ± 0.020 99.98 ± 0.00 95.27 ± 0.00
Retrain0.000 ± 0.000 0.000 ± 0.000 0.000 ± 0.000 99.82 ± 0.02 91.01 ± 0.07
DIMU2.177 ± 0.129 1.350 ± 0.160 0.260 ± 0.030 96.79 ± 0.13 88.26 ± 0.25

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.