acceptodds
Under review as a conference paper at ICLR 2027

Generation-Resistant Data Publishing with Force Field and Finsler Geometry

28 views

Abstract

In the era of generative AI, modern data publishing faces a new privacy challenge: datasets released to preserve utility for downstream tasks can be exploited to train or fine-tune generative AI models, enabling training-example extraction, membership inference, and style copying. This paper proposes a unified framework for generation-resistant data publishing (GRDP) that simultaneously preserves classification utility while degrading generative capability, by leveraging force field, Finsler geometry, and generative statistics theory. To preserve classification utility, we characterize boundary-sensitive local discriminative geometry in the embedding space through a physics-inspired bubble expansion process and construct a distance-decaying force field with intra-class attraction and inter-class repulsion. Under the induced Finsler geometry, anisotropic bubble shells evolve according to a Hamilton–Jacobi formulation, and force invariance is enforced along shell trajectories so that local decision-boundary geometry remains stable even when global distributional statistics are distorted. To degrade generative capability while preserving local geometry, we derive a Gaussian target prior whose covariance shift increases the FID-based feature-distribution mismatch under a quadratic approximation of the force-invariance budget. We further demonstrate that this covariance shift has a closed-form rank-one structure aligned with the minimum-eigenvalue direction of the induced force matrix, i.e., the least force-constrained direction. The published data are then matched to this prior through covariance shaping, distorting global distributional statistics while maintaining classification performance.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Task-Specific Data Publishing, Generative Utility Degradation, Classification Utility Preservation, Force Field, Finsler Geometry, Covariance Prior Matching

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_M5fyOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Generation-Resistant Data Publishing (GRDP), aiming to release a modified dataset that preserves utility for training downstream classifiers while degrading the quality of generative models trained on the same data. The authors motivation is that classification relies primarily on local boundary geometry, whereas generative models rely on global distributional statistics. To formalize this, they model embedded samples as expanding bubbles governed by intra-class attraction and inter-class repulsion, formulating the shell evolution via a Hamilton-Jacobi equation under an induced Finsler metric to enforce local force invariance. To degrade generative utility, they relax the force-invariance constraint into a quadratic covariance budget and derive a rank-one covariance shift along the minimum-eigenvalue direction to maximize the Gaussian FID mismatch.

Critical strengths
  • Well-motivated problem and intuitive high-level insight: Protecting released datasets from unauthorized generative training while maintaining discriminative utility is a timely and practical problem. The conceptual idea of decoupling local decision-boundary geometry from global feature moments is intuitive and appealing.
Critical weaknesses
  • Severe over-complication ("mathiness") and a huge gap between theory and implementation: The main text spends pages building an elaborate mathematical apparatus—NTK Taylor expansions, Finsler geometry, dual norms, and Hamilton-Jacobi PDEs. However, this heavy machinery is largely abandoned in both the subsequent derivations and the actual implementation:
    1. When simulating shell trajectories, the update step simply normalizes the force vector using the standard Euclidean norm. Later, when deriving the covariance constraint, the Finsler metric is upper-bounded by a standard quadratic form, effectively collapsing the anisotropic Finsler structure back to a basic weighted Euclidean metric.
    2. While Algorithm 1 in the main text claims to optimize all dataset embedding centers directly, Appendix D reveals that the actual implementation only modifies a tiny fraction (e.g., 5%) of the training images by passing them through a learned autoencoder with random Gaussian latent noise and an MLP.
    3. Most importantly, the ablation study (Table 18 / Figure 2c) shows that simply injecting random Gaussian noise into the autoencoder latent space without any of the proposed modules (Random Shift) already achieves almost the exact same ResNet accuracy and FID degradation as the full GRDP framework. This strongly suggests that the generative degradation comes almost entirely from replacing 5% of the dataset with blurry, low-fidelity autoencoder reconstructions, rather than from the Finsler geometry or rank-one covariance matching.
  • Flawed threat model and experimental inconsistencies:
    1. Despite being framed as "data publishing", the method only modifies a small subset (1%–10%) of the data while leaving the remaining 90%–99% of raw images untouched. Furthermore, there is no explicit perceptual bound on the modified images, which suffer from severe visual distortion (very low PSNR and SSIM in Table 20). In practice, a downstream user training a generative model could easily filter out this 5% of heavily degraded outliers using basic image quality metrics and train cleanly on the remaining 95% of untouched data.
    2. There are unexplained inconsistencies in the experimental tables. For instance, the baseline DDIM-FID on the raw CIFAR-100 dataset differs drastically between Table 18 and Table 21, and poisoning the dataset with UIB in Table 21 somehow yields a much better (lower) FID than training on the clean raw dataset. Moreover, in the dataset distillation comparison, GRDP does not distill data itself but merely modifies 5% of an existing distilled dataset (D3HR), yielding marginal differences.
Questions for authors
  • Clarification Requests:
    1. True contribution of the theoretical framework: Given that the Random Shift ablation variant (autoencoder + random Gaussian latent noise) achieves nearly identical classification accuracy and FID degradation as Full GRDP, how do you justify the necessity of the Finsler geometry, Hamilton-Jacobi trajectory sampling, and rank-one covariance matching?
    2. Vulnerability to simple outlier filtering: Since only a small fraction of images are modified and they exhibit heavy visual degradation without any pixel-space perturbation bound, what prevents an adversary from simply filtering out the modified subset via basic anomaly or quality detection? What happens if 100% of the published dataset is passed through your pipeline?
    3. Experimental discrepancies: Why does the raw CIFAR-100 DDIM-FID change so dramatically between Table 18 and Table 21, and why does UIB poisoning improve generative quality (lower FID/KID) over the clean dataset in Table 21?
    4. Computational tractability: How are the pairwise force fields, Jacobians, and the dual Finsler gauge optimization actually computed for tens of thousands of samples—are they only approximated within small mini-batches?
  • Suggestions for Future Work:
    1. A realistic data publishing framework should process the entire released dataset under strict perceptual fidelity constraints (e.g., bounded LPIPS or L∞L_\infty perturbations) and provide visual comparisons of the images before and after publishing.
7 comments

We would like to thank the reviewer for the helpful and constructive comments. We have tried our best to address your concerns and have revised the manuscript accordingly. Our detailed responses and the corresponding revisions are provided below.

W1:Necessity of Proposed Components We conduct additional ablation experiments on CIFAR-100 under a 5% modification ratio to examine the contributions of force invariance, Finsler geometry, covariance matching, covariance direction selection, and bubble expansion monitor points. All variants use the same classification and DDIM evaluation settings.

1. Contribution of Force Invariance, Finsler Geometry, and Covariance Matching.

MethodAccuracy (%) ↑FID ↑
Clean Dataset77.5045.055
AE Reconstruction Only76.3446.946
Random Gaussian Shift76.4847.360
Random Covariance Direction Only76.8649.096
Covariance Matching Only77.0750.555
FF + CS (Euclidean Throughout)77.0750.723
Full GRDP77.2253.836
GRDP with Euclidean Constraint76.7450.309
GRDP with Random Covariance Direction76.5052.054

Random Gaussian Shift achieves 76.48% classification accuracy and 47.360 FID, compared with 77.22% and 53.836 for Full GRDP. Covariance Matching Only increases FID to 50.555, but remains below Full GRDP. Replacing the Finsler constraint with a Euclidean constraint reduces accuracy from 77.22% to 76.74% and FID from 53.836 to 50.309. Using Euclidean geometry throughout further reduces FID to 50.723. These results show that the Finsler constraint improves the balance between classification accuracy and generative degradation compared with its Euclidean alternatives.

The covariance direction also affects the results. Theorem 4.1 selects the minimum eigenvalue direction of Σ01/2AΣ01/2\Sigma_0^{1/2}A\Sigma_0^{1/2} to maximize the FID objective under the derived covariance constraint. Replacing this direction with a random covariance direction reduces accuracy from 77.22% to 76.50% and FID from 53.836 to 52.054. This supports the use of the force induced covariance direction instead of an arbitrary statistical shift.

2. Contribution of Bubble Expansion Monitor Points.

We further examine the monitor point selection process. In Full GRDP, monitor points are sampled along bubble expansion trajectories. In the ablation, we replace them with randomly sampled points around the same control centers using random directions and a fixed radius of 0.5.

Both variants use the same 2,500 control samples, with 16 directions and three samples per direction, resulting in 120,000 monitor points. The reference forces are recomputed at the random monitor points using all 50,000 embedding centers. Force matching, Finsler geometry, covariance matching, and the publishing network remain unchanged.

MethodAccuracy (%) ↑FID ↑KID ↑Gen. PrecisionLPIPS ↑
Full GRDP77.2253.8360.0425220.650360.250819
Random Monitor Points76.6348.4350.0468130.600620.249511

Replacing bubble expansion monitor points with random points reduces classification accuracy by 0.59 percentage points and FID by 5.401. LPIPS also decreases slightly, while KID increases. Since the number of monitor points and all remaining components are unchanged, the differences in classification accuracy and FID reflect the contribution of trajectory based monitor point selection.

W2: Connection Between Theory and Implementation

The mathematical formulation and practical implementation of GRDP are connected through three steps: constructing boundary-sensitive monitor points, deriving the covariance target, and optimizing the published representations.

1. Finsler geometry and Hamilton–Jacobi shell evolution. Theorem 3.2 establishes the Finsler metric in Eq. (9) and its dual gauge in Eq. (10), which characterize the direction-dependent geometry induced by the original force field. The dual gauge defines the theoretical propagation speed in Eq. (11), and Eq. (12) describes the continuous shell evolution. In the numerical implementation, we construct force-guided trajectories using explicit Euler updates. The propagation direction is determined by the normalized Euclidean resultant force, while the propagation speed is approximated using the Euclidean force norm, threshold RR, and coefficient μ\mu, without explicitly solving the dual Finsler gauge optimization in Eq. (10). These trajectories provide the monitor points at which the original and published force fields are compared through the Finsler force-invariance loss in Eq. (16).

2. Finsler geometry and quadratic covariance formulation. The quadratic formulation in Section 4 follows from the original Finsler metric. Since 1<q≤21<q\leq2 and ∑jωj0(p)=1\sum_j\omega_j^0(p)=1, Jensen's inequality gives

F0(p,v)2=(∑jωj0(p)∣⟨v,uj0(p)⟩∣q)2/q≤∑jωj0(p)⟨v,uj0(p)⟩2=v⊤W(p)v.F_0(p,v)^2=\left(\sum_j\omega_j^0(p)|\langle v,u_j^0(p)\rangle|^q\right)^{2/q}\leq\sum_j\omega_j^0(p)\langle v,u_j^0(p)\rangle^2=v^\top W(p)v.

Thus, Eq. (18) provides a quadratic upper bound derived from the Finsler metric. The matrix W(p)W(p) retains the original directional weights, while the force Jacobians in Eq. (22) describe local changes in the resultant force. Combining them gives the force-induced matrix AA in Eq. (23), leading to the covariance constraint in Eq. (24) under the local linear approximation. Theorem 4.1 then identifies the minimum eigenvalue direction of Σ01/2AΣ01/2\Sigma_0^{1/2}A\Sigma_0^{1/2} to construct the rank-one covariance target. The final objective in Eq. (31) combines covariance matching with the original force-invariance loss.

3. Algorithm 1 and budget-constrained publishing. Algorithm 1 describes the embedding-space optimization, while Appendix D implements it through a learnable publishing transformation under a fixed modification budget. Let SS denote the selected sample indices. For selected samples (j∈Sj \in S), we set zj⋆=Pθ(z~j)z_j^\star = P_\theta(\tilde{z}_j), where the publishing MLP transforms the optionally perturbed latent representation. For unselected samples (j∉Sj \notin S), we keep zj⋆=zj0z_j^\star = z_j^0. The force-invariance and covariance-matching objectives are defined over the resulting published embedding configuration, with unselected centers fixed at their original values. A ResNet-18-based encoder maps images into 64-dimensional embeddings, and the decoder converts the modified representations back into images. Under the 5% modification budget on CIFAR-100, 2,500 samples are modified, while the remaining 47,500 images remain unchanged.

We have clarified the numerical shell evolution, covariance construction, and budget constrained publishing implementation in Appendix D.

W3: Computational Efficiency and Scalability We clarify the computational implementation of pairwise forces, Jacobians, covariance construction, and shell evolution. All geometric computations are performed in a 64-dimensional embedding space using a two-level batching strategy to control memory consumption.

1. Monitor-point construction and pairwise forces. Under the 5% modification budget on CIFAR-100, monitor points are constructed from the 2,500 selected samples instead of all 50,000 training samples. Each selected sample uses 16 initial directions and three trajectory sampling points, resulting in 120,000 monitor points. The resultant force in Eq. (7) includes contributions from all 50,000 embedding centers. To avoid constructing the full pairwise interaction matrix, we process 32 monitor points per batch and divide the embedding centers into chunks of 256. The force contributions from successive center chunks are accumulated for each monitor-point batch.

2. Analytical Jacobians and covariance construction. The force Jacobians are computed directly using the analytical expression in Eq. (22), without PyTorch autograd. We apply the same two-level batching strategy, processing 32 monitor points and 256 embedding centers per chunk. Each chunk contributes to the Jacobians of the current monitor-point batch. The resulting Jacobians are immediately combined with the Finsler-related matrix W(p)W(p) to accumulate the force-induced matrix AA in Eq. (23). Since AA is a 64×6464\times64 matrix, only the current batch of Jacobians and the accumulated matrix need to be stored, without retaining Jacobians for all monitor points.

3. Numerical shell evolution and publishing optimization. The Hamilton–Jacobi equation in Eq. (12) describes the continuous shell evolution. In the numerical implementation, we construct force-guided trajectories using explicit Euler updates, with the propagation speed approximated from the Euclidean resultant-force norm, threshold RR, and coefficient μ\mu. This provides a numerical approximation for trajectory sampling without explicitly solving the dual Finsler gauge optimization in Eq. (10). The Finsler metric remains part of the force-invariance loss and covariance target construction. The publishing transformation is then optimized through a three-hidden-layer MLP with 256 hidden units per layer, using a batch size of 128, a learning rate of 5×10−45\times10^{-4}, and 10 training epochs. The implementation retains contributions from all original embedding centers without storing the full pairwise interaction matrix or the Jacobians of all monitor points. Chunked force computation and batch-wise accumulation of AA reduce memory usage while preserving the original force-field formulation.

We have added the monitor point construction, chunked force computation, analytical Jacobian calculation, and covariance matrix construction details in Appendix D.

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

The paper proposes a generation-resistant data publishing framework that preserves classification utility while degrading generative capabilities. It achieves this by preserving local discriminative geometry and modifying global feature statistics.

Critical strengths
  1. Well-Written: The paper is well-organized and clearly written, making it easy to follow.

  2. Interesting Topic: The paper addresses an interesting and important problem of preserving classification utility while degrading generative capabilities.

Critical weaknesses
  1. Loss Function Design: The proposed loss function relies on FID, which is not the only metric for evaluating generative models. The authors should better justify this design choice.

  2. Limited Generative Models: The experiments only consider DDIM-based diffusion models. More generative architectures, such as GANs and VAEs, should be evaluated.

  3. Limited Dataset Diversity: The experiments are restricted to low-resolution image datasets. Evaluation on higher-resolution datasets would strengthen the results.

  4. Generalizability: It remains unclear whether the proposed framework can be extended to text or multimodal datasets. The authors should discuss its applicability beyond images.

Questions for authors

See weakness 4

3 comments

We would like to thank the reviewer for the helpful and constructive comments. We have tried our best to address your concerns and have revised the manuscript accordingly. Our detailed responses and the corresponding revisions are provided below.

W1: Loss Function Design

We choose FID for the generative degradation objective because its formulation in terms of feature means and covariances directly matches the covariance constraint derived from force invariance. Specifically, Eqs. (18)–(24) connect the original force Jacobians and Finsler metric to a covariance constraint. Under this constraint, Theorem 4.1 derives the covariance direction that maximizes the FID objective, corresponding to the minimum eigenvalue direction of Σ01/2AΣ01/2\Sigma_0^{1/2}A\Sigma_0^{1/2}. This provides a closed form target covariance, which is incorporated into the publishing loss through covariance matching in Eq. (31). Thus, FID is selected because its mathematical form allows us to derive and optimize the generative degradation objective under the classification preserving constraint.

We also evaluate the resulting generative degradation using metrics beyond FID. In Table 21, GRDP increases DDIM FID from 49.708 to 53.836 and KID from 0.0483 to 0.0524. For StyleGAN-XL, FID increases from 77.335 to 83.016 and KID from 0.008869 to 0.010860. We additionally report LPIPS to evaluate perceptual differences. These results show that the distributional changes introduced by GRDP are reflected in KID as well as FID across both diffusion and GAN models.

We have added an explanation of the FID objective and its connection to the covariance constraint in Appendix D.

W2: Limited Generative Models

We have evaluated GRDP on StyleGAN-XL in addition to DDIM, with the results reported in Appendix E.3 (Table 21). We further conduct experiments using a ResNet-based convolutional VAE on CIFAR-100.

VAE Experimental Setup. The VAE uses residual blocks, Group Normalization, and a latent dimension of 256. We independently train two models on the clean and GRDP datasets for 200 epochs using Adam, a learning rate of 2×10−42\times10^{-4}, and a batch size of 128. The training objective combines pixel reconstruction error and KL divergence with β=1\beta=1 and a linear warmup over the first 20 epochs. Both datasets contain 50,000 images, with 5% modified in the GRDP dataset. Each trained VAE generates 50,000 images, which are evaluated against the original CIFAR-100 dataset.

Training DatasetFID ↓KID ↓Gen. Precision ↑
Clean Dataset149.04820.1247230.24786
GRDP 5%140.42090.1225550.26422

Results. GRDP increases both FID and KID for DDIM and StyleGAN-XL, as reported in Appendix E.3. The VAE results differ: FID decreases from 149.0482 to 140.4209, KID decreases from 0.124723 to 0.122555, and generative precision increases from 0.24786 to 0.26422.

GRDP constructs its covariance target by maximizing an FID objective under the force constraint in Theorem 4.1. This objective changes the feature statistics of the published training data but does not directly optimize the reconstruction loss or latent distribution learned by a VAE. In our experiment, the clean VAE already has an FID of 149.0482, indicating a substantial difference between its generated distribution and the original CIFAR-100 distribution. The VAE's reconstruction objective and KL regularization may affect how changes in the training data are reflected in its generated samples.

We have added the VAE experimental setup, results, and analysis in Appendix E.3, following the existing DDIM and StyleGAN-XL experiments.

W3: Limited Dataset Diversity

We have included additional experiments on an ImageNet-1K subset covering all 1,000 classes, with results reported in Appendix E.1 (Table 19). Classification performance is evaluated using a pretrained ResNet-18, and generative performance is evaluated using an unconditional DDIM at 64 × 64 resolution.

At the 5% modification ratio, GRDP achieves 52.12% top-5 classification accuracy, 104.943 FID, and 0.566 LPIPS. At the 10% ratio, it achieves 51.37% top-5 accuracy, 120.124 FID, and 0.573 LPIPS. GRDP achieves the highest classification accuracy and LPIPS among the compared methods at both modification ratios, while obtaining the second highest FID.

W4: Extension to Text and Multimodal Data

GRDP is formulated in an embedding space. The force field in Eq. (7) depends on feature distances and class labels, while the Finsler metric in Eq. (9) describes the local geometry of these representations. Eq. (16) preserves local class relationships through force invariance, and Theorem 4.1 determines the covariance direction used to modify global feature statistics. These formulations do not depend on image pixels and can also be defined for text and multimodal representations.

For text data, a text encoder can map sentences into a semantic embedding space, where the same force and covariance constraints can be applied. The modified embeddings must then be decoded into discrete token sequences while maintaining their original class labels and semantic content. This requires a text decoder and additional constraints on the decoded sentences, since preserving embedding geometry alone does not guarantee valid text.

For multimodal data, the force field and covariance target can be constructed in a shared embedding space. In addition to preserving class relationships, the publishing process would need to maintain the similarity between corresponding representations from different modalities, such as paired images and text descriptions. This can be incorporated as an additional alignment constraint during publishing.

The current implementation uses an image encoder and decoder, as described in Appendix D. The force and covariance formulations can be transferred to other embedding spaces, while text decoding and multimodal alignment require changes to the publishing process.

We have added a discussion of the extensions to text and multimodal data in Appendix E.6.

We would like to thank the reviewer for taking the time to review our paper and for the valuable comments. We hope our response has adequately addressed your concerns raised in the review.

Please kindly let us know if anything is unclear. We truly appreciate this opportunity to improve our work and shall be most grateful for any feedback you could give us.

Reviewer_9zHZOpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a unified framework for generation-resistant data publishing (GRDP) that simultaneously preserves classification utility while degrading generative capability, by leveraging force field, Finsler geometry, and generative statistics theory. To preserve classification utility, authors characterize boundary-sensitive local discriminative geometry in the embedding space through a physics-inspired bubble expansion process and construct a distance-decaying force field with intra-class attraction and inter-class repulsion.

Critical strengths
  • The proposed method is novel and is based on a physics-inspired bubble expansion process for generation-resistant data publishing.
  • The proposed method is associated with theoretical analysis that demonstrate effectiveness of approach.
  • The proposed method achieves remarkable performance compared to previous methods.
Critical weaknesses
  • The experiments are conducted on small datasets (CIFAR-100, Tiny ImageNet). It will be interesting to see the comparison on larger datasets.
Questions for authors
  • can an adversary try to reverse the approach and generate/estimate original distribution?
3 comments

We would like to thank the reviewer for the helpful and constructive comments. We have tried our best to address your concerns and have revised the manuscript accordingly. Our detailed responses and the corresponding revisions are provided below.

W1: Evaluation on Larger Datasets Our original submission already includes experiments on an ImageNet-1K in Appendix E.1 (Table 19), covering all 1,000 classes. Under both 5% and 10% modification ratios, GRDP achieves the highest top-5 classification accuracy and LPIPS among the compared methods, while maintaining competitive FID results. These results demonstrate that GRDP remains effective on a dataset with substantially greater class diversity.

Q1: Recovery of the Original Distribution

1. Recovery Analysis. GRDP modifies global feature statistics while preserving local discriminative geometry. According to Theorem 4.1, the target covariance is

Σprior=Σ01/2(I+κλmin⁡umin⁡umin⁡⊤)Σ01/2.\Sigma_{\mathrm{prior}}=\Sigma_0^{1/2}\left(I+\frac{\kappa}{\lambda_{\min}}u_{\min}u_{\min}^{\top}\right)\Sigma_0^{1/2}.

Here, Σ0\Sigma_0 is the original covariance, while umin⁡u_{\min} and λmin⁡\lambda_{\min} are obtained from the minimum eigenpair of Σ01/2AΣ01/2\Sigma_0^{1/2}A\Sigma_0^{1/2}. The matrix AA is constructed from the original force Jacobians and Finsler geometry at the monitor points (Eqs. (19)–(23)). Directly reversing the covariance shift therefore requires the original covariance and force geometry, which are not available to the attacker. Eq. (16) constrains the resultant forces at a finite set of monitor points, which does not uniquely determine the original global distribution. The publishing process modifies selected embeddings through a learned MLP and decoder instead of applying a fixed invertible transformation. The attacker also does not know the modification ratio or modified sample indices, so the original and modified samples cannot be directly separated using known modification information.

2. Adaptive Recovery Attack. We evaluate whether an attacker can restore generative performance by identifying and removing modified samples. We conduct the experiment on CIFAR-100, where GRDP modifies 2,500 of the 50,000 training images. The attacker accesses only the published dataset, without knowing the original images, modification ratio, or modified sample indices.

We extract 512 dimensional features using an ImageNet pretrained ResNet-18, compute 20 nearest neighbor distances, and apply a two component Gaussian Mixture Model (GMM) to identify suspicious samples. Without using the true modification ratio, the GMM identifies and removes 4,512 images. We compare this attack with random filtering and clean filtering using the same number of remaining images. DDIM is trained from scratch for 100 epochs with a batch size of 128, and generative performance is evaluated against the original CIFAR-100 distribution.

Training DatasetImagesFID ↓KID ↓Gen. Precision ↑
Clean Dataset50,00045.05530.0492340.60404
Full GRDP50,00053.83600.0425220.65036
Clean Filtered45,48845.91410.0430450.59196
Random Filtered45,48848.05650.0402620.62732
GMM Filtered45,48849.96140.0460050.60008

3. Recovery Results. GMM filtering reduces FID from 53.8360 to 49.9614, but it remains higher than the clean dataset FID of 45.0553. GMM filtering also performs worse than random filtering (48.0565) and clean filtering (45.9141), despite retaining the same number of training images. The attacker removes 4,512 images, nearly twice the actual number of modified samples, without knowing the modification ratio. KID and generative precision show different trends, indicating that the recovery outcome varies across evaluation metrics. The tested filtering attack does not restore the original FID performance, and identifying suspicious samples provides no advantage over random filtering in this experiment.

We have added the adaptive GMM filtering experiment and its results in Appendix E.5.

We would like to thank the reviewer for taking the time to review our paper and for the valuable comments. We hope our response has adequately addressed your concerns raised in the review.

Please kindly let us know if anything is unclear. We truly appreciate this opportunity to improve our work and shall be most grateful for any feedback you could give us.

Thanks authors for their response which mostly addressed my concerns.

Comments on the paper

1 comment
Area_Chair_zcsUOpenReview ↗

Dear Reviewers

Since the author-reviewer discussion ends soon, I encourage the reviewer to review the authors' rebuttals and let them know whether their responses are satisfactory and whether any concerns remain, if you haven't done so.

Thank you for your time and effort in reviewing for ICLR 2027.

Best wishes,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.