pith. sign in

arxiv: 2606.29880 · v1 · pith:K64RLLSTnew · submitted 2026-06-29 · 💻 cs.CV

IREU: Identity-Related Encoder-Only Unlearning for Customized Portrait Generation

Chaoyi Shi , Shanshan Zhang , Jian Yang This is my paper

Pith reviewed 2026-06-30 06:28 UTC · model grok-4.3

classification 💻 cs.CV
keywords identity unlearningcustomized portrait generationimage encoderfeature perturbationprivacy protectioncross-generator generalization
0
0 comments X

The pith

Selective offline perturbation of only identity-related features in the image encoder unlearns target identities while preserving fidelity for retained identities.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper investigates identity unlearning for customized portrait generation to mitigate privacy risks from fake person images. A simple baseline that globally updates the image encoder harms generation quality for identities that should be retained. IREU solves this by first locating identity-related features offline and then restricting perturbations to those features alone. Experiments show this yields stronger unlearning on targets and better preservation on retained identities. The resulting encoder also transfers to other generators using the same encoder without any further tuning.

Core claim

The proposed IREU method first locates identity-related features in an offline manner and then only performs feature perturbations on them. This achieves better identity unlearning performance for target identities to be unlearned, while keeping high fidelity for other identities to be retained. Additionally, the unlearned image encoder is generalizable across different generators with the same encoder without fine-tuning.

What carries the argument

IREU, the offline location of identity-related features followed by selective perturbation restricted to those features in the image encoder.

Load-bearing premise

That identity-related features can be reliably located in an offline manner such that selective perturbation removes the target identity without unintended side effects on retained identities or overall generation quality.

What would settle it

Measuring a statistically significant drop in identity similarity scores for retained identities after IREU is applied, relative to the original encoder, would falsify the preservation claim.

Figures

Figures reproduced from arXiv: 2606.29880 by Chaoyi Shi, Jian Yang, Shanshan Zhang.

Figure 1
Figure 1. Figure 1: Illustration of our identity unlearning paradigm. Left: a typical CPG pipeline, making edits based on the text prompt while keeping the same identity as the input. Right: only the image encoder is updated (w/ the diffusion model frozen) during un￾learning, so that the generated images exhibit a different identity from the input. To protect privacy, some methods primarily focus on image-level protection, su… view at source ↗
Figure 2
Figure 2. Figure 2 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of our proposed method IREU, which consists of two key components: (1) Locating Identity-Related Features: given an image x− to be unlearned and its Face-Swap counterpart xa, we encode them with the original image encoder ε to obtain their embeddings. Then we compute the difference ∆z, and rank |∆z| to build a binary mask M that selects identity-related dimensions for feature perturbation. (2) Unl… view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of difference features across 1,000 Face-Swap pairs. The non￾uniform distribution exhibits a clear bias towards specific dimensions, which motivates our dimensional masking strategy to isolate identity-related features. In the difference vector ∆z, the absolute value of each component reflects how strongly that dimension responds to the identity change: a larger |∆zi | implies a higher identit… view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of visual quality with BIA [33] for unlearning under the oneID setting on FastComposer [45]. For each input image (first column), both BIA and our method perform identity unlearning from the same original model. Text prompts are included below each generated image. split, we randomly sample training images according to the number of iterations. To evaluate unlearning performance on unseen images… view at source ↗
Figure 6
Figure 6. Figure 6: Visualization results for unseen images of the same target identity to be unlearned under the fiveID setting on FastComposer [45]. “Target” refers to the images used during unlearning, while “Unseen” denotes the images belonging to the same target identity to be un￾learned but not used during unlearning. A street art mural of a woman 选择的生成：0，8， 21 遗忘后 原始 30，5 原始 遗忘后 Target Unseen Ul w/ Unlearning GUIDE BIA… view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative results when our unlearned image encoder is plugged into a new generator PhotoMaker [25] without any finetuning under the fiveID setting. Top: an identity to be retained; bottom: a target identity to be unlearned. A street art mural of a woman 选择的生成：0，8， 21 遗忘后 原始 30，5 原始 遗忘后 Target Unseen Ul w/ Unlearning GUIDE BIA Ours w/o Unlearning w/ Unlearning Input Input Reference a =1 a =10 a =100 a =10… view at source ↗
Figure 9
Figure 9. Figure 9: Multi-subject image generation results of FastComposer [45]. Our unlearned image encoder is adapted from the single-subject to multi-subject setting without any finetuning. For each example, the input image marked with a red box is the target identity to be unlearned, and the other one represents the identity to be retained. both BIA and the baseline. These results demonstrate that our method not only pres… view at source ↗
Figure 10
Figure 10. Figure 10: Analysis on the ratio of identity-related features k. (a) The SSIM values show a larger variance (with more small values) as k increases. (b) ∆FID consistently increases as k grows. These results support our argument that restricting perturbations to identity-related elements preserves fidelity better. A street art mural of a woman 选择的生成：0，8， 21 遗忘后 原始 30，5 原始 遗忘后 Source Others Ul Unlearning GUIDE BIA Our… view at source ↗
Figure 11
Figure 11. Figure 11: Impact of the perturbation magnitude α in the identity-related transformation (Eq. 3). Increasing α consistently strengthens identity unlearning performance (lower ID similarity), but it saturates after α = 100. small, the constructed virtual identity z˜− stays close to z−, leading to insufficient unlearning. As α increases, the deviation from the input image grows and the identity similarity of the targe… view at source ↗
Figure 13
Figure 13. Figure 13: Analysis on the magnitude of masked identity-related differences. Cases are sorted by \delimiter "026B30D z_- - \tilde {z}_-\|_2 . The consistently low \protect \text {ID}_\text {target}^{o} and high \protect \text {ID}_\text {other}^{g} show that variations in masked differences do not harm the unlearning-retaining trade-off. that mild variations in the identity-related shift do not degrade either identi… view at source ↗
read the original abstract

Customized Portrait Generation (CPG) technologies have been widely used to generate high-fidelity person images given an input image indicating the identity and a text prompt indicating the required edits. Yet these methods pose significant privacy risks by spreading fake visual information. Against such risks, each public generator should be able to suppress its generation ability for a particular person when requested. Therefore, in this work we investigate the identity unlearning problem for CPG. Since there are no previous methods in this field, we propose a simple baseline that updates the image encoder by minimizing identity similarity between generated and input images for target identities to be unlearned, while maximizing it for identities to be retained. However, we find such a global perturbation in the feature space harms the fidelity of generated images for other identities to be retained. To solve this problem, we propose a novel method IREU, which first locates identity-related features in an offline manner and then only performs feature perturbations on them. The experimental results show that our proposed method IREU achieves better identity unlearning performance for target identities to be unlearned, and also keeps high fidelity for other identities to be retained. In addition, our unlearned image encoder is generalizable across different generators with the same encoder without fine-tuning, which is friendly for deployment in practice.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper addresses privacy risks in Customized Portrait Generation (CPG) by proposing identity unlearning for specific persons. It introduces a baseline that globally updates the image encoder to minimize identity similarity for target identities while maximizing it for retained ones, but observes that this harms fidelity for retained identities. To address this, IREU locates identity-related features offline and applies perturbations selectively to them. The abstract claims that IREU yields better unlearning for targets, preserves high fidelity for retained identities, and produces an encoder generalizable across different generators without fine-tuning.

Significance. If the empirical claims are substantiated, the work would offer a practical encoder-only mechanism for selective identity suppression in generative portrait models, mitigating privacy harms while maintaining utility for non-target identities and enabling deployment across multiple generators.

major comments (2)
  1. [Abstract] Abstract: The central claim that 'the experimental results show that our proposed method IREU achieves better identity unlearning performance for target identities to be unlearned, and also keeps high fidelity for other identities to be retained' is unsupported by any metrics, datasets, ablation studies, or quantitative comparisons, rendering the performance advantage unverifiable.
  2. [Abstract] Abstract: The method's core step of locating 'identity-related features in an offline manner' is described only at the level of the claim; no procedure, validation that the located features are identity-specific rather than entangled with generation quality, or ablation isolating the contribution of selective (vs. global) perturbation is provided, which is load-bearing for the argument that IREU solves the fidelity problem identified in the baseline.
minor comments (1)
  1. [Abstract] The abstract refers to 'each public generator' and 'different generators with the same encoder' without naming any specific CPG architectures or encoders used, which would aid immediate context.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each point below and will revise the abstract to better substantiate the claims by incorporating key quantitative results and a concise description of the core procedure, while ensuring the full manuscript's experimental details remain the primary support.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that 'the experimental results show that our proposed method IREU achieves better identity unlearning performance for target identities to be unlearned, and also keeps high fidelity for other identities to be retained' is unsupported by any metrics, datasets, ablation studies, or quantitative comparisons, rendering the performance advantage unverifiable.

    Authors: The abstract is a concise summary; the full manuscript provides the supporting quantitative evidence, including identity similarity metrics, fidelity scores, dataset details, baseline comparisons, and ablation studies in the experiments section. We will revise the abstract to include specific metrics and dataset references to make the performance claims more directly verifiable. revision: yes

  2. Referee: [Abstract] Abstract: The method's core step of locating 'identity-related features in an offline manner' is described only at the level of the claim; no procedure, validation that the located features are identity-specific rather than entangled with generation quality, or ablation isolating the contribution of selective (vs. global) perturbation is provided, which is load-bearing for the argument that IREU solves the fidelity problem identified in the baseline.

    Authors: The full manuscript details the offline location procedure, validation experiments confirming identity specificity, and ablations isolating selective perturbation effects in Section 3 and the experiments. We agree the abstract is high-level and will revise it to briefly outline the procedure and reference the supporting analyses. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical method with independent experimental validation

full rationale

The paper describes a baseline global perturbation approach, observes its limitation on retained identities, and proposes IREU as an alternative that performs offline localization followed by selective perturbation. All central claims rest on experimental comparisons rather than any derivation that reduces to fitted inputs, self-definitions, or self-citation chains. No equations or uniqueness theorems are invoked that collapse back to the method's own assumptions by construction. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on the assumption that identity similarity is a well-defined and manipulable quantity in encoder feature space; the offline location step is presented without details on how features are identified or validated.

axioms (1)
  • domain assumption Identity similarity measured in the image encoder feature space is a reliable proxy for generation behavior.
    Both the baseline and IREU optimize directly on this similarity metric.

pith-pipeline@v0.9.1-grok · 5761 in / 1090 out tokens · 34109 ms · 2026-06-30T06:28:48.516218+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

51 extracted references · 4 canonical work pages · 1 internal anchor

  1. [1]

    In: CVPR

    Ahmed, S.M., Basaran, U.Y., Raychaudhuri, D.S., Dutta, A., Kundu, R., Niloy, F.F., Guler, B., Roy-Chowdhury, A.K.: Towards source-free machine unlearning. In: CVPR. pp. 4948–4957 (2025)

    2025

  2. [2]

    In: ICCV

    Bala, A., Chowdhury, R., Jaiswal, R., Roheda, S.: Dct-shield: A robust frequency domain defense against malicious image editing. In: ICCV. pp. 18876–18884 (2025)

    2025

  3. [3]

    In: CVPR

    Chen, M., Gao, W., Liu, G., Peng, K., Wang, C.: Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In: CVPR. pp. 7766–7775 (2023)

    2023

  4. [4]

    In: CVPRW

    Choi, D., Choi, S., Lee, E., Seo, J., Na, D.: Towards efficient machine unlearning with data augmentation: Guided loss-increasing (GLI) to prevent the catastrophic model utility drop. In: CVPRW. pp. 93–102 (2024)

    2024

  5. [5]

    arXiv preprint arXiv:2405.11135 (2024)

    Feng, W., Zhou, W., He, J., Zhang, J., Wei, T., Li, G., Zhang, T., Zhang, W., Yu, N.: Aqualora: Toward white-box protection for customized stable diffusion models via watermark lora. arXiv preprint arXiv:2405.11135 (2024)

    work page arXiv 2024

  6. [6]

    In: ICLR (2023)

    Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image gener- ation using textual inversion. In: ICLR (2023)

    2023

  7. [7]

    In: ICCV

    Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. In: ICCV. pp. 2426–2436 (2023)

    2023

  8. [8]

    In: WACV

    Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. In: WACV. pp. 5111–5120 (2024)

    2024

  9. [9]

    In: CVPR

    Gao, G., Huang, H., Fu, C., Li, Z., He, R.: Information bottleneck disentanglement for identity swapping. In: CVPR. pp. 3404–3413 (2021)

    2021

  10. [10]

    In: ICCV

    Gao, H., Pang, T., Du, C., Hu, T., Deng, Z., Lin, M.: Meta-unlearning on diffusion models: Preventing relearning unlearned concepts. In: ICCV. pp. 2131–2141 (2025)

    2025

  11. [11]

    In: ECCV

    Gong, C., Chen, K., Wei, Z., Chen, J., Jiang, Y.G.: Reliable and efficient concept erasure of text-to-image diffusion models. In: ECCV. pp. 73–88 (2024)

    2024

  12. [12]

    In: NeurIPS

    Gu, H., Ong, W.K., Chan, C.S., Fan, L.: Ferrari: Federated feature unlearning via optimizing feature sensitivity. In: NeurIPS. vol. 37, pp. 24150–24180 (2024)

    2024

  13. [13]

    In: NeurIPS

    Heusel,M.,Ramsauer,H.,Unterthiner,T.,Nessler,B.,Hochreiter,S.:Ganstrained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS. vol. 30 (2017)

    2017

  14. [14]

    In: CVPR

    Huang, Y., Wang, Y., Tai, Y., Liu, X., Shen, P., Li, S., Li, J., Huang, F.: Curric- ularface: Adaptive curriculum learning loss for deep face recognition. In: CVPR. pp. 5900–5909 (2020)

    2020

  15. [15]

    In: CVPR

    Huang,Z.,Chan,K.C.K.,Jiang,Y.,Liu,Z.:Collaborativediffusionformulti-modal face generation and editing. In: CVPR. pp. 6080–6090 (2023)

    2023

  16. [16]

    In: CVPR (2025)

    Kang, T., Jeong, S., Jang, H., Choo, J.: Zero-shot head swapping in real-world scenarios. In: CVPR (2025)

    2025

  17. [17]

    In: ICLR (2018)

    Karras, T., Aila, T., Laine, S., Lehtinen, J.: Progressive growing of gans for im- proved quality, stability, and variation. In: ICLR (2018)

    2018

  18. [18]

    In: CVPR

    Khalil, Y.H., Brunswic, L.M., Lamghari, S., Li, X., Beitollahi, M., Chen, X.: Not: Federated unlearning via weight negation. In: CVPR. pp. 25759–25769 (2025)

    2025

  19. [19]

    In: CVPR

    Kim, J., Gu, G., Park, M., Park, S., Choo, J.: Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on. In: CVPR. pp. 8176– 8185 (2024)

    2024

  20. [20]

    In: CVPR

    Lee, B.H., Lim, S., Chun, S.Y.: Localized concept erasure for text-to-image diffu- sion models using training-free gated low-rank adaptation. In: CVPR. pp. 18596– 18606 (2025) IREU for Customized Portrait Generation 17

    2025

  21. [21]

    In: NeurIPS

    Lee, T.Y., Seo, J., Ko, J.H., Park, G.M.: Perturb a model, not an image: Towards robust privacy protection via anti-personalized diffusion models. In: NeurIPS. vol. 38, pp. 92410–92442 (2025)

    2025

  22. [22]

    arXiv preprint arXiv:2406.15305 (2024)

    Li, A., Mo, Y., Li, M., Wang, Y.: Pid: Prompt-independent data protection against latent diffusion models. arXiv preprint arXiv:2406.15305 (2024)

    work page arXiv 2024

  23. [23]

    In: CVPR

    Li, M., Chen, J., Feng, W., Li, B., Dai, F., Zhao, S., He, Q.: Hyperlora: Parameter- efficient adaptive generation for portrait synthesis. In: CVPR. pp. 13114–13123 (2025)

    2025

  24. [24]

    IEEE Trans

    Li, N., Zhou, C., Gao, Y., Chen, H., Zhang, Z., Kuang, B., Fu, A.: Machine un- learning: Taxonomy, metrics, applications, challenges, and prospects. IEEE Trans. Neural Networks Learn. Syst.36(8), 13709–13729 (2025)

    2025

  25. [25]

    In: CVPR

    Li, Z., Cao, M., Wang, X., Qi, Z., Cheng, M., Shan, Y.: Photomaker: Customizing realistic human photos via stacked ID embedding. In: CVPR. pp. 8640–8650 (2024)

    2024

  26. [26]

    In: ICCV

    Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: ICCV. pp. 3730–3738 (2015)

    2015

  27. [27]

    In: CVPR

    Lu, S., Wang, Z., Li, L., Liu, Y., Kong, A.W.K.: Mace: Mass concept erasure in diffusion models. In: CVPR. pp. 6430–6440 (2024)

    2024

  28. [28]

    In: CVPR

    Lyu, M., Yang, Y., Hong, H., Chen, H., Jin, X., He, Y., Xue, H., Han, J., Ding, G.: One-dimensional adapter to rule them all: Concepts diffusion models and erasing applications. In: CVPR. pp. 7559–7568 (2024)

    2024

  29. [29]

    In: AAAI

    Moon, S., Cho, S., Kim, D.: Feature unlearning for pre-trained gans and vaes. In: AAAI. pp. 21420–21428 (2024)

    2024

  30. [30]

    In:ICML.vol.139, pp.8748–8763 (2021)

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models fromnatural languagesupervision. In:ICML.vol.139, pp.8748–8763 (2021)

    2021

  31. [31]

    In: CVPR

    Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dream- booth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR. pp. 22500–22510 (2023)

    2023

  32. [32]

    In: CVPR

    Seo, J., Lee, S., Lee, T., Moon, S., Park, G.: Generative unlearning for any identity. In: CVPR. pp. 9151–9161 (2024)

    2024

  33. [33]

    In: CVPR

    Shaheryar, M., Lee, J.T., Jung, S.K.: Black hole-driven identity absorbing in dif- fusion models. In: CVPR. pp. 28544–28554 (2025)

    2025

  34. [34]

    Shamshad,F.,Naseer,M.,Nandakumar,K.:Clip2protect:Protectingfacialprivacy usingtext-guidedmakeupviaadversariallatentsearch.In:CVPR.pp.20595–20605 (2023)

    2023

  35. [35]

    In: CVPR

    Song, Y., Yang, P., Ci, H., Shou, M.Z.: Idprotector: An adversarial noise encoder to protect against id-preserving image generation. In: CVPR. pp. 3019–3028 (2025)

    2025

  36. [36]

    In: CVPR

    Spartalis, C.N., Semertzidis, T., Gavves, E., Daras, P.: Lotus: Large-scale machine unlearning with a taste of uncertainty. In: CVPR. pp. 10046–10055 (2025)

    2025

  37. [37]

    In: ECCV

    Tian, L., Wang, Q., Zhang, B., Bo, L.: EMO: emote portrait alive generating expressive portrait videos with audio2video diffusion model under weak conditions. In: ECCV. vol. 15141, pp. 244–260 (2024)

    2024

  38. [38]

    A Practical Guide, 1st Ed., Cham: Springer International Publishing10(3152676), 10–5555 (2017)

    Voigt, P., dem Bussche, A.V.: The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing10(3152676), 10–5555 (2017)

    2017

  39. [39]

    arXiv preprint arXiv:2303.09522 (2023)

    Voynov, A., Chu, Q., Cohen-Or, D., Aberman, K.: P+: Extended textual condi- tioning in text-to-image generation. arXiv preprint arXiv:2303.09522 (2023)

    work page arXiv 2023

  40. [40]

    Shi et al

    Wang, H., Weng, Y., Li, Y., Guo, Z., Du, J., Niu, S., Ma, J., He, S., Wu, X., Hu, Q., Yin, B., Liu, C., Liu, Q.: Emotivetalk: Expressive talking head generation 18 C. Shi et al. through audio information decoupling and emotional video diffusion. In: CVPR. pp. 26212–26221 (2025)

    2025

  41. [41]

    InstantID: Zero-shot Identity-Preserving Generation in Seconds

    Wang, Q., Bai, X., Wang, H., Qin, Z., Chen, A.: Instantid: Zero-shot identity- preserving generation in seconds. arXiv preprint arXiv:2401.07519 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  42. [42]

    In: CVPR

    Wang, Z., Guo, J., Zhu, J., Li, Y., Huang, H., Chen, M., Tu, Z.: Sleepermark: Towards robust watermark against fine-tuning text-to-image diffusion models. In: CVPR. pp. 8213–8224 (2025)

    2025

  43. [43]

    In: CVPR

    Wu, J., Le, T., Hayat, M., Harandi, M.: Erasing undesirable influence in diffusion models. In: CVPR. pp. 28263–28273 (2025)

    2025

  44. [44]

    In: AAAI

    Wu, Y., Zhou, S., Yang, M., Wang, L., Chang, H., Zhu, W., Hu, X., Zhou, X., Yang, X.: Unlearning concepts in diffusion model via concept domain correction and concept preserving gradient. In: AAAI. pp. 8496–8504 (2025)

    2025

  45. [45]

    Xiao, G., Yin, T., Freeman, W.T., Durand, F., Han, S.: Fastcomposer: Tuning- free multi-subject image generation with localized attention. Int. J. Comput. Vis. 133(3), 1175–1194 (2024)

    2024

  46. [46]

    In: ICML

    Yang, J., Wang, Y., Fang, Y., Dai, Y., Huang, F.: Variance as a catalyst: Effi- cient and transferable semantic erasure adversarial attack for customized diffusion models. In: ICML. vol. 267 (2025)

    2025

  47. [47]

    In: CVPR

    Zeng, J., Song, D., Nie, W., Tian, H., Wang, T., Liu, A.: CAT-DM: controllable accelerated virtual try-on with diffusion model. In: CVPR. pp. 8372–8382 (2024)

    2024

  48. [48]

    In: CVPRW

    Zhang, G., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. In: CVPRW. pp. 1755–1764 (2024)

    2024

  49. [49]

    In: NeurIPS

    Zhang, Y., Chen, X., Jia, J., Zhang, Y., Fan, C., Liu, J., Hong, M., Ding, K., Liu, S.: Defensive unlearning with adversarial training for robust concept erasure in diffusion models. In: NeurIPS. vol. 37, pp. 36748–36776 (2024)

    2024

  50. [50]

    In: CVPR

    Zhong, Z., Bao, W., Wang, J., Zhang, S., Zhou, J., Lyu, L., Lim, W.Y.B.: Unlearn- ing through knowledge overwriting: Reversible federated unlearning via selective sparse adapter. In: CVPR. pp. 30661–30670 (2025)

    2025

  51. [51]

    In: CVPR

    Zhou, Y., Zheng, D., Mo, Q., Lu, R., Lin, K.Y., Zheng, W.S.: Decoupled distillation to erase: A general unlearning method for any class-centric tasks. In: CVPR. pp. 20350–20359 (2025)

    2025