Pith. sign in

REVIEW 3 major objections 4 minor 19 references

Group-level training-data attribution for diffusion models can be approximated by unlearning each candidate group from the full model, matching leave-one-group-out retraining's ability to identify primary influences at a tiny fraction of th

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 06:29 UTC pith:P24FZ7OL

load-bearing objection Group-wise attribution via unlearning is a good idea, but the central equivalence claim rests on a missing importance-weight correction, and the empirical wins are thinner than the tables suggest. the 3 major comments →

arxiv 2601.22651 v2 pith:P24FZ7OL submitted 2026-01-30 cs.LG cs.AI

GUDA: Counterfactual Group-wise Training Data Attribution for Diffusion Models via Unlearning

Naoki Murata , Yuhta Takida , Chieh-Hsin Lai , Toshimitsu Uesaka , Bac Nguyen , Stefano Ermon , Yuki Mitsufuji This is my paper
classification cs.LG cs.AI
keywords training data attributiondiffusion modelsmachine unlearninggroup attributioncounterfactualELBOtext-to-imagedata valuation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that group-level training-data attribution for diffusion models—asking which style or class contributed to a generated image—can be computed efficiently by replacing leave-one-group-out retraining with machine unlearning. The proposed framework, GUDA, starts from the full-data diffusion model, unlearns each group one at a time, and scores a group's influence as the difference in the model's ELBO on the generated sample before and after that group is unlearned. On CIFAR-10 and a 60-style text-to-image benchmark, the resulting rankings agree with the retrained oracle more often than CLIP similarity, gradient-based attribution, or instance-level unlearning, at roughly a hundredth of the retraining wall-clock time on CIFAR-10. The load-bearing move is choosing an unlearning objective that approximates actual retraining without that group—redirecting forget-set denoising toward retain-set targets—rather than a generic erasure loss.

Core claim

The paper's central claim is that the counterfactual attribution estimand—the difference in ELBO between the full model and the model retrained without group k—can be faithfully approximated by the analogous difference using a model obtained by unlearning group k from the full model. It then shows, empirically, that this approximation identifies the most influential groups better than semantic similarity (CLIP prototypes), aggregated gradient-based scores, and instance-level unlearning, in both unconditional (CIFAR-10) and conditional text-to-image (UnlearnCanvas/Stable Diffusion 1.5) settings. For unconditional models the unlearning loss samples noisy latents from the forget group but redir

What carries the argument

The central object is the group-removal counterfactual model θ^logo_{-k}—a model retrained without group k—and its approximation θ^ul_{-k} obtained by unlearning. Attribution is computed as an ELBO difference, ELBO(x0|c;θ_full) − ELBO(x0|c;θ^ul_{-k}), which is cheap once the unlearned models are precomputed. The unlearning operator carries the argument: in the unconditional setting it uses an importance-weighted redirection loss that, for each noisy latent sampled from a forget-group image, targets the weighted-average noise of nearby retain-group images, so the model avoids memorizing the group while preserving retain-set behavior via a score-matching distillation term. In the conditional s

Load-bearing premise

The framework's theoretical foundation is the claim that the unconditional redirection objective (Eq. 15) is an unbiased reformulation of the retain-only training objective; this requires the forget-set and retain-set noise distributions to match, which they generally do not, so the unlearned model is not guaranteed to coincide with the LOGO retrained model.

What would settle it

On a synthetic two-cluster dataset where the forget cluster is far from the retain cluster, compare the model obtained by minimizing Eq. 15 against one obtained by minimizing Eq. 10 (true LOGO) with the same architecture and budget; if their parameters or ELBOs on a held-out sample differ systematically, the 'unbiased reformulation' claim is refuted. The paper never checks this directly—it only compares final attribution rankings.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Group-level attribution becomes practical at scale: one full model plus N lightweight unlearned models replaces N+1 full retraining runs, and the unlearning step is amortized over any number of query images.
  • Because scores are computed from likelihood differences under precomputed models, query-time attribution is a few seconds per image, independent of training time, making it usable in copyright, compensation, and debugging workflows.
  • The choice of unlearning objective matters: redirection-style objectives that target the retain-only counterfactual materially outperform erasure-style and instance-level unlearning for head identification, suggesting that attribution fidelity is tied to how closely unlearning reproduces removal.
  • If the empirical agreement with LOGO holds beyond the tested benchmarks, GUDA provides a viable oracle proxy for developing and evaluating future attribution methods without retraining costs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable extension of the paper's logic: since attribution fidelity depends on how closely unlearning approximates true removal, one could tune unlearning objectives on retain-set likelihood stability alone, without ever computing expensive LOGO baselines—the paper's ablations suggest head metrics are stable across the number of neighbors K, hinting that the ranking may be robust to imperfect ap
  • The framework is not diffusion-specific in principle; any likelihood-based generative model with a tractable ELBO or surrogate could reuse the same unlearning-and-score recipe for group attribution, though the unlearning operator would need re-derivation.
  • The conditional anchor construction (keep content, swap style) is essentially a controlled prompt intervention; it could be reused beyond attribution as a tool for counterfactual style analysis or for probing what a model learned about a concept.
  • The paper's oracle is itself a fine-tuned approximation of true from-scratch retraining, so the reported agreement with LOGO is partly an agreement between two approximations; closing that residual gap is an open problem.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes GUDA, a framework for group-wise training-data attribution in diffusion models. The target estimand is Leave-One-Group-Out Attribution (LOGOA), defined as the ELBO difference between the full-data model and a model retrained without the group. GUDA approximates each counterfactual model by unlearning the group from a shared full model: a ReTrack-based redirection loss for unconditional generation (GUDA-U) and a CLIP-weighted anchor-conditioning loss for conditional text-to-image generation (GUDA-C). Experiments on CIFAR-10 and UnlearnCanvas compare GUDA to CLIP similarity, gradient-based attribution, and instance-level unlearning, reporting improved top-k agreement with the LOGO oracle and a substantial wall-clock speedup. The central theoretical claim is that the ReTrack forget loss is an unbiased reformulation of the retain-only LOGO objective.

Significance. If the theoretical link to LOGO were valid, GUDA would be a practical and well-motivated contribution: it defines a clear counterfactual estimand, uses unlearning as a computational proxy, validates against a LOGO oracle, includes extensive ablations, and reports meaningful gains over reasonable baselines. The head-focused evaluation metrics and the exposure-matched oracle design are thoughtful. However, the central derivation in Appendix A.1 does not support the claimed equivalence, and the conditional analogue is asserted rather than derived. The empirical results are interesting, but the paper's main claim—that GUDA approximates LOGOA—currently rests on an invalid reformulation, so the contribution is not established at the level required for publication.

major comments (3)
  1. [Appendix A.1, Eq. (15)] The claimed unbiased reformulation is invalid. Equation (14) conditions on x_t sampled from the retain mixture p_ret,t, while Eq. (15) draws x_t from the forget mixture p_for,t. The normalized retain weights w_t are posterior weights given x_t, not a correction for the change in the marginal distribution of x_t. An unbiased reformulation would require the importance weight p_ret,t(x_t)/p_for,t(x_t), which is absent. The appendix's statement that ReTrack 'change[s] how we sample x_t without changing the target objective' is exactly the step that is not justified. This is load-bearing because Sec. 4.2.1 and Sec. 5.1.3 rely on this equivalence to connect GUDA-U to the LOGO oracle. The appended caveats (K-NN truncation, finite steps, preservation regularizer) are approximations to the full objective, but the full objective itself is not the LOGO target. The practical damage may be limited by
  2. [Sec. 5.1.2 / Appendix C.1.2] The 'full' CIFAR-10 model is not trained on all data: one class is excluded at random each epoch for exposure matching. Thus the θ_full used in Eq. (5) is not the full-data model that defines LOGOA in Sec. 3.3. The reference model's training distribution differs from that of each LOGO model (random per-epoch omission versus fixed class omission), and the effect of this mismatch on the measured agreement with LOGOA is unquantified. This does not necessarily invalidate the experimental comparison, but it changes the interpretation of the oracle evaluation from 'full data versus leave-one-group-out' to a comparison against a nonstandard, exposure-matched reference.
  3. [Sec. 4.2.2, Eq. (9)] The conditional forget loss is not shown to approximate any conditional LOGO objective. Unlike the unconditional case, there is no derivation analogous to Appendix A.1 connecting anchor redirection to retain-only conditional training. The switch from forget prompt c_f to anchor prompt c_a changes the conditioning distribution in a way that is not accounted for by any importance weight, bound, or asymptotic argument. Since the UnlearnCanvas experiments are the main practical demonstration of GUDA-C, this is a second load-bearing gap in the theoretical support for the framework.
minor comments (4)
  1. [Figures 4 and 5] The labels 'Commit Etch' and 'Dogs' appear inconsistent; the former should presumably be 'Comic Etch', and the latter appears to be misplaced under Crypto Punks. Please check all dataset/style labels for consistency.
  2. [Sec. 3.2 / Appendix C] Equation (3) defines ELBO as a variational lower bound, but Appendix C estimates ELBO as a sum of per-timestep KL divergences. The relationship between this estimator and the variational bound of Eq. (3) is not derived; please clarify the exact estimator and why it is a valid surrogate for the definition.
  3. [Appendix C.1.3] The timestep restriction [2000,3900] for CIFAR-10 unlearning is not motivated. Given the theoretical issue in Appendix A.1, the paper should at least provide an analysis of how close p_for,t and p_ret,t are in this range, or acknowledge that the method is a heuristic in this regime.
  4. [Eq. (9) / Appendix B.1] The anchor condition c_a is used in Eq. (9) but not defined until Appendix B.1; add a forward reference so the main-text loss is self-contained.

Circularity Check

0 steps flagged

No meaningful circularity: GUDA's approximation is validated against independently retrained LOGO oracles, and the flagged Appendix A.1 sampling issue is a correctness concern, not a definitional loop.

full rationale

The central derivation chain is not circular. GUDA's score (Eq. 5) is defined as an ELBO difference between a shared full-data model and an unlearned model, while the LOGOA oracle (Eq. 4) uses an independently retrained leave-one-group-out model. The unlearned models are produced by fine-tuning with the objectives in Sec. 4.2, which do not take the LOGOA rankings, LOGOA scores, or any fitted attribution output as input; therefore no fitted parameter is renamed as a prediction. The evaluation compares GUDA's top-k rankings against rankings derived from LOGO models that were retrained from scratch or from an exposure-matched initialization (CIFAR-10 and UnlearnCanvas respectively). Using the same ELBO estimation protocol and reusing noise across models is a variance-reduction technique for fair comparison, not a construction that forces the GUDA score to equal LOGOA. The paper contains no self-citations that carry the argument; ReTrack (Shi et al., 2025), ESD, DAS, D-TRAK, and other prior works are external references. The main theoretical bridge is Appendix A.1's claim that the ReTrack forget loss (Eq. 15) is an unbiased reformulation of the retain-only LOGO objective (Eq. 10). Whether that equality actually holds under the forget-mixture sampling distribution is a correctness/mathematical issue, not a circularity: Eq. 15 is not defined in terms of the GUDA score or the LOGOA score, and the paper itself flags the subsequent approximations (k-NN truncation, finite steps, preservation regularizer) as biases away from the exact LOGO optimum. Thus the derivation chain is self-contained against an external benchmark, and the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The method depends on several tuned hyperparameters and on a flawed theoretical equivalence. The most load-bearing assumption is the claimed unbiasedness of the ReTrack reformulation, which is disproven by the equations themselves.

free parameters (7)
  • lambda_forget (CIFAR-10 ReTrack) = 0.03
    Forget loss weight tuned by ablation (Table 5); controls strength of redirection.
  • lambda_pres (UnlearnCanvas) = 2.0
    Preservation loss weight for conditional unlearning; no ablation reported.
  • K (ReTrack neighbors) = 10
    Truncation number for importance-weighted target; ablation shows stability across K.
  • tau (style-selection temperature) = 2.0
    Temperature for CLIP-weighted style selection, chosen via hyperparameter sweep.
  • eta_mix (style-selection mixing) = 0.1
    Uniform mixing weight for exploration, chosen via sweep.
  • alpha (descriptor filtering) = 0.5
    Penalty exponent for cross-style descriptor sharing; sensitivity not reported.
  • exposure-matched full-model dropout = one random class per epoch
    CIFAR-10 full model excludes one random class per epoch to match LOGO training size; this changes the definition of θ_full.
axioms (5)
  • domain assumption ELBO differences preserve log-likelihood ordering for attribution
    Sec. 3.3 explicitly states ELBO does not guarantee preservation of likelihood ordering, yet it is used as the attribution score.
  • ad hoc to paper ReTrack loss with p_for,t sampling is an unbiased reformulation of the retain-only objective
    Appendix A.1 Eq. (15) changes sampling distribution from p_ret,t to p_for,t without an importance-weight correction; the equality does not hold.
  • ad hoc to paper Anchor-conditioned redirection approximates retain-only conditional training
    Sec. 4.2.2 proposes a heuristic conditional analogue without derivation connecting it to the LOGO target.
  • domain assumption Fine-tuned LOGO proxy is a valid oracle for counterfactual group removal
    Sec. 5.2.1 instantiates the oracle by fine-tuning from the same SD 1.5 initialization, which already contains style knowledge; this is not a clean counterfactual.
  • domain assumption Generated images from θ_full are valid queries
    θ_full on CIFAR-10 is exposure-matched with random class dropout, so queries are from a non-standard model.

pith-pipeline@v1.3.0-alltime-deepseek · 24009 in / 12887 out tokens · 128310 ms · 2026-08-03T06:29:33.732692+00:00 · methodology

0 comments
read the original abstract

Training-data attribution for vision generative models aims to identify which training data influenced a given output. While most methods score individual examples, practitioners often need group-level answers (e.g., artistic styles or object classes). Group-wise attribution is counterfactual: how would a model's behavior on a generated sample change if a group were absent from training? A natural realization of this counterfactual is Leave-One-Group-Out (LOGO) retraining, which retrains the model with each group removed; however, it becomes computationally prohibitive as the number of groups grows. We propose GUDA (Group Unlearning-based Data Attribution) for diffusion models, which approximates each counterfactual model by applying machine unlearning to a shared full-data model instead of training from scratch. GUDA quantifies group influence using differences in a likelihood-based scoring rule (ELBO) between the full model and each unlearned counterfactual. Experiments on CIFAR-10 and artistic style attribution with Stable Diffusion show that GUDA identifies primary contributing groups more reliably than semantic similarity, gradient-based attribution, and instance-level unlearning approaches, while achieving ~100x speedup on CIFAR-10 over LOGO retraining.

Figures

Figures reproduced from arXiv: 2601.22651 by Bac Nguyen, Chieh-Hsin Lai, Naoki Murata, Stefano Ermon, Toshimitsu Uesaka, Yuhta Takida, Yuki Mitsufuji.

Figure 1
Figure 1. Figure 1: Overview of the GUDA (Group Unlearning-based Data Attribution) framework. Instead of retraining N + 1 models from scratch (LOGO), GUDA starts from the all-group model θ full and applies machine unlearning to obtain approximate counterfactual models θ ul −k for each group k. Attribution is computed by comparing the ELBO of generated samples under the all-group model versus each unlearned model, providing an… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison of group-wise attribution methods on UnlearnCanvas. For two generated images, we show the top-3 attributed styles for each method. Each row corresponds to: LOGOA (oracle target), GUDA (Ours), CLIPA, and Wang et al.. Representative training images from each attributed style are displayed. Green boxes indicate agreement with the LOGOA top-3 styles. phabetically. Each LOGO model θ logo … view at source ↗
Figure 3
Figure 3. Figure 3: CIFAR-10 generations under group removal. Each panel pairs the all-class model with a counterfactual model trained without class 1 (automobile) or class 7 (horse). We use DDIM with 100 steps and the same initial noise for each pair. The top row shows LOGO models, and the bottom row shows ReTrack and ESD unlearned models. Remark on the LOGO oracle (implementation). While Sec. 3.3 defines θ logo −k as the co… view at source ↗
Figure 4
Figure 4. Figure 4: UnlearnCanvas training data examples for the Dogs class across the 16 evaluation styles. Images within the same style exhibit consistent visual characteristics, supporting style-level attribution analysis. Abstractionism Artist Sketch Blossom Season Blue Blooming Bricks Byzantine Cartoon Cold Warm Color Fantasy Commit Etch Crayon Dogs Crypto Punks Cubism Dogs Dadaism Dapple Defoliation [PITH_FULL_IMAGE:fi… view at source ↗
Figure 5
Figure 5. Figure 5: Evaluation images generated by the all-group model using the evaluation prompts. Due to partial descriptor overlap across styles, some generations reflect a dominant style while others exhibit mixed visual characteristics. using openai/clip-vit-large-patch145 with style selection parameters τ = 2.0 (temperature sharpness) and ηmix = 0.1 (uniform mixing weight for exploration). C.2.3. TRAINING AND EVALUATIO… view at source ↗

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

19 extracted references · 13 linked inside Pith

  1. [1]

    Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923,

    Bai, S., Chen, K., et al. Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923,

    Pith/arXiv arXiv

  2. [5]

    Erasing concepts from diffusion models.arXiv preprint arXiv:2303.07345,

    Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., and Bau, D. Erasing concepts from diffusion models.arXiv preprint arXiv:2303.07345,

    Pith/arXiv arXiv

  3. [6]

    and Zou, J

    Ghorbani, A. and Zou, J. Data shapley: Equitable valuation of data for machine learning. InProceedings of the 36th International Conference on Machine Learning, 2019a. Ghorbani, A. and Zou, J. Equitable valuation of data for ma- chine learning.arXiv preprint arXiv:1904.02868, 2019b. Ginart, A., Guan, M. Y ., Valiant, G., and Zou, J. Making AI forget you: ...

    Pith/arXiv arXiv 1904

  4. [7]

    Denoising diffusion proba- bilistic models.arXiv preprint arXiv:2006.11239,

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion proba- bilistic models.arXiv preprint arXiv:2006.11239,

    Pith/arXiv arXiv 2006

  5. [11]

    Estimat- ing training data influence by tracing gradient descent

    Pruthi, G., Liu, F., Kale, S., and Sundararajan, M. Estimat- ing training data influence by tracing gradient descent. arXiv preprint arXiv:2002.08484,

    Pith/arXiv arXiv 2002

  6. [13]

    Retrack: Data unlearn- ing in diffusion models through redirecting the denoising trajectory.arXiv preprint arXiv:2509.13007,

    Shi, Q., Jin, C., Zhang, J., and Gu, Y . Retrack: Data unlearn- ing in diffusion models through redirecting the denoising trajectory.arXiv preprint arXiv:2509.13007,

    arXiv

  7. [14]

    Forget- me-not: Learning to forget in text-to-image diffusion models

    Zhang, G., Wang, K., Xu, X., Wang, Z., and Shi, H. Forget- me-not: Learning to forget in text-to-image diffusion models. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 1755–1764. IEEE, 2024a. Zhang, Y ., Fan, C., Zhang, Y ., Yao, Y ., Jia, J., Liu, J., Zhang, G., Liu, G., Kompella, R. R., Liu, X., and Liu, S. ...

    Pith/arXiv arXiv

  8. [16]

    {object} in {style} style

    (runwayml/stable-diffusion-v1-5) with frozen text encoder and V AE, full fine-tuning of UNet. Training uses DDPMScheduler with default configuration andϵ-prediction type. Loss function is mean squared error between predicted and ground-truth noise. All-group model: Trained on all 60 styles for 10,000 steps with learning rate 1×10 −6 (constant schedule, no...

    2025

  9. [17]

    how far the method missed the gold Top-1

    and the model’s learned reverse process. Timestep sampling: Uniform grid with stride 10 overt∈[10,999], yielding 100 timesteps per ELBO computation. Noise sampling: Single noise sample per timestep, with the same noise reused across all models (all-group, LOGO, unlearned) for variance reduction. Precision: UNet forward pass uses mixed precision (bf16 via ...

    2021

  10. [19]

    method to investigate the impact of (i) ridge regularization parameter λ and (ii) aggregation strategies for converting instance-level scores to group-level attribution. D.2.1. EFFECT OFRIDGEREGULARIZATIONPARAMETERλ Table 8 presents the effect of ridge regularization parameter λ∈ {0.1,0.01,0.001} with mean aggregation fixed. The results show that λ= 0.01 ...

    2024

  11. [20]

    This suggests that stronger redirection is beneficial under the current preservation loss, without degrading ranking stability

    Increasing λ strengthens forgetting and improves head metrics up to λ= 0.03 . This suggests that stronger redirection is beneficial under the current preservation loss, without degrading ranking stability. 20 GUDA: Counterfactual Group-wise Data Attribution Table 6.ReTrack ablation: effect of K-nearest neighbors (lr = 1×10 −5, λ= 0.03, epochs = 20). Best ...

    arXiv

  12. [256]

    Dropout 0.3 Attention resolutions 16, 8 Diffusion Process Scheduler DDPM Timesteps 4,000 Noise schedule Squared cosine Prediction typeϵ-prediction Loss function MSE Training Optimizer AdamW Weight decay10 −4 Learning rate10 −4 LR schedule Cosine (5,000 warmup) Gradient clipping 1.0 Batch size 256 Epochs 2,400 C. Experimental Setup Remark on loss weighting...

    2022

  13. [2002]

    Koh, P. W. and Liang, P. Understanding black-box predictions via influence functions.arXiv preprint arXiv:1703.04730,

    Pith/arXiv arXiv

  14. [2019]

    Extracting training data from diffusion models.arXiv preprint arXiv:2301.13188,

    Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V ., Tram`er, F., Balle, B., Ippolito, D., and Wallace, E. Extracting training data from diffusion models.arXiv preprint arXiv:2301.13188,

    Pith/arXiv arXiv

  15. [2020]

    A., Jia, H., Travers, A., Zhang, B., Lie, D., and Papernot, N

    Bourtoule, L., Chandrasekaran, V ., Choquette-Choo, C. A., Jia, H., Travers, A., Zhang, B., Lie, D., and Papernot, N. Machine unlearning.arXiv preprint arXiv:1912.03817,

    Pith/arXiv arXiv 1912

  16. [2021]

    High-resolution image synthesis with la- tent diffusion models.arXiv preprint arXiv:2112.10752,

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with la- tent diffusion models.arXiv preprint arXiv:2112.10752,

    Pith/arXiv arXiv

  17. [2022]

    M., Georgiev, K., Ilyas, A., Leclerc, G., and Madry, A

    Park, S. M., Georgiev, K., Ilyas, A., Leclerc, G., and Madry, A. Trak: Attributing model behavior at scale.arXiv preprint arXiv:2303.14186,

    Pith/arXiv arXiv

  18. [2023]

    URL https: //openreview.net/forum?id=fxNQJVMwK2. Dai, Z. and Gifford, D. K. Training data attribution for dif- fusion models.arXiv preprint arXiv:2306.02174,

    Pith/arXiv arXiv

  19. [2025]

    net/forum?id=9EqQC2ct4H

    URL https://openreview. net/forum?id=9EqQC2ct4H. Nguyen, T. T., Huynh, T. T., Ren, Z., Nguyen, P. L., Liew, A. W.-C., Yin, H., and Nguyen, Q. V . H. A survey of machine unlearning.arXiv preprint arXiv:2209.02299,

    Pith/arXiv arXiv