acceptodds
Under review as a conference paper at ICLR 2027

PRISM: Breaking the Reactive Bottleneck in Retain-Free Unlearning with Proactive Data-Free Structural Priors

Abstract

The practical deployment of Machine Unlearning (MU) is hindered by two fundamental, often intertwined, challenges: (1) a widespread dependency on original retain data, which is often legally and practically untenable, and (2) a severe operational bottleneck in existing methods. This bottleneck renders critical operational functions, such as efficiently calibrating forgetting intensity and handling concurrent requests, a practical challenge largely unaddressed in the field. We introduce PRISM (PRoactive Intrinsic Structural Modeling for Unlearning), a retain-free framework that solves this trade-off with a proactive and model-centric philosophy. At its core is our novel 'Data-Free Structural Prior', a data-agnostic foundation learned offline by modeling the model's intrinsic parameter structure. This proactive design replaces costly iterative optimization with a single, analytical update at request time. This decoupled structure enables amortized handling of concurrent requests and low-cost calibration to meet a predefined forgetting target. Extensive experiments confirm PRISM's state-of-the-art performance and operational efficiency, unifying the erasure of data, classes, and concepts across discriminative and generative models.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Data Protection, Privacy, Machine Unlearning, Concept Unlearning

Reviews

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

This paper proposes PRISM, a retain-free machine unlearning framework designed to reduce the computational cost of unlearning requests. It learns a structural prior from pretrained model parameters using a VAE and combines it with forget-data gradients for a single-step parameter update. Experiments show competitive unlearning performance and improved computational efficiency across discriminative and generative models.

Critical strengths
  1. The paper addresses practical challenges in retain-free unlearning, particularly the computational cost of repeated optimization and forgetting-strength calibration.

  2. The separation of offline structural modeling and request-time updates is interesting and allows the learned prior to be reused across different unlearning requests.

  3. The experiments cover multiple unlearning tasks across discriminative and generative models, with useful ablation and efficiency analyses.

Critical weaknesses
  1. Unclear structural prior construction: The paper does not sufficiently explain how the VAE learns meaningful parameter dependencies from the pretrained model, making the central component difficult to assess and reproduce.

  2. Incomplete calibration cost analysis: The reported calibration times exclude the evaluation cost needed to verify whether a forgetting target has been met, potentially underestimating the practical overhead.

  3. Limited evidence for concurrent requests: The experiments evaluate sequential rather than concurrent requests, and retain accuracy degrades after repeated unlearning, leaving the claimed scalability insufficiently validated.

Questions for authors
  1. How does the VAE learn meaningful structural dependencies from the pretrained model's parameters?

  2. Does the reported calibration time include the cost of evaluating whether the forgetting target is met?

  3. How do the sequential unlearning results support the claimed efficiency of handling concurrent requests?

Reviewer_c2k9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces PRISM (PRoactive Intrinsic Structural Modeling for Unlearning), a retain-free unlearning framework that does not need the original retain data. Before any request, it trains a VAE on the pretrained parameters and uses the decoder Jacobian to obtain a Data-Free Structural Prior. At request time, it compares this prior with a Fisher-based forgetting score on the forget set and attenuates the selected parameters in a single analytical update, where the selection rate γ\gamma controls the forgetting intensity. Experiments on sample-wise and class-wise forgetting on CIFAR-100 and nudity removal in Stable Diffusion demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The proposed method computes the retention signal once before any request, so each request only needs a forget-set Fisher and one analytical update, which makes it cheap to tune the forgetting intensity with a single parameter γ\gamma.
  2. Without any retain data, PRISM keeps retain and test accuracy close to Retrain in class-wise forgetting and gives the lowest unsafe ratio in nudity removal, which shows that the method works for both classifiers and diffusion models.
Critical weaknesses
  1. The contribution of the Data-Free Structural Prior is not clearly isolated. The VAE is trained on a single parameter vector θ0\theta_0, and the paper does not explain why the Jacobian of its decoder should indicate which parameters are important for the retained knowledge. Table 6 only compares with weight norms and β=0\beta=0; adding a uniform prior (si=1s_i=1) and a data-based prior (e.g., the Fisher on DD used by SSD) under the same rule in Eqs. (11)-(14) would show whether the gain comes from the VAE or from the selection rule.

  2. The operational advantage over existing methods seems overstated. SSD has the same O(H⋅P)O(H \cdot P) calibration cost in Table 5, and its importance on DD does not depend on the request, so it could also be computed once before the training data is deleted. For concept erasure, UCE [1] edits the cross-attention weights in closed form and MACE [2] erases up to 100 concepts at once, but neither is discussed or compared.

[1] Gandikota, Rohit, et al. "Unified concept editing in diffusion models." 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2024. [2] Lu, Shilin, et al. "Mace: Mass concept erasure in diffusion models." 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024.

Questions for authors
  1. Appendix B uses a two-layer VAE with hidden size 256 on the flattened θ0\theta_0 of ResNet-18. How is such a high-dimensional input handled (e.g., split into chunks), and is the VAE trained only on masked copies of θ0\theta_0?
  2. The paper follows Fan et al. (2024), but its Avg. Gap only averages UA, RA, and TA, while SalUn also includes MIA. How does Table 2 change if MIA is included, given that PRISM has the second largest ΔMIA\Delta_{MIA} (8.45)?
  3. The nudity evaluation uses only 10 I2P prompts (Table 18), and FID is computed against the original SD outputs. How does PRISM perform on the full I2P nudity set, with FID on COCO prompts?
  4. For random forgetting, γ\gamma is tuned using UA only. What is the target UA? Matching the Retrain UA in Table 1 would require the retrained model.
  5. Only sequential requests are evaluated, and RA drops from 99.17 to 82.48 after five classes in Table 11. Is the prior computed on θ0\theta_0 still valid after several updates, and how does PRISM handle concurrent requests? Minor:
  6. Lines 1145-1147: The text says PRISM has competitive CLIP similarity on non-identity prompts, but Table 16 shows that its CLIP(t, i) and CLIP(i, i) are the lowest among all methods.
  7. Table 5: The calibration cost of the baselines is written as H⋅TpreH \cdot T_{pre}, but they have no pre-request stage.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.