Pith. sign in

REVIEW 4 major objections 5 minor 21 references

On the Reconstruction of Training Data from Group Invariant Networks

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Pith's one-line read Standard reconstruction methods applied to group-invariant networks converge to symmetric orbit averages, and two symmetry-aware modifications improve reconstruction quality.

desk verdict Worth engaging: first study of reconstruction attacks on invariant nets, with a nice stabilizer-nesting observation, but the printed KKT loss is missing a gradient and SAME-GD's gains don't show up in the CIFAR-10 rows. read the letter →

arxiv 2411.16458 v1 pith:6TDKNNCU submitted 2024-11-25 cs.LG

classification cs.LG
keywords group-invariantneuralnetworksdatareconstructionactivationmaximizationKKTgradientdescentequivarianceorbitopedeepimagepriorprivacyattacks
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Reconstructing training data from a neural network gets harder when the network is invariant to a group of transformations such as reflections or rotations. This paper argues that conventional reconstruction objectives—activation maximization and KKT-based reconstruction—are themselves group-invariant, and gradient descent on such objectives cannot reach arbitrary points: the update is equivariant and the stabilizer of the iterate only grows, so reconstructions drift toward symmetric inputs on which the group acts trivially, typically the average over the orbit. Experiments on MNIST and CIFAR-10 with groups of size 2 to 8 confirm this drift, measured by group-invariant DSSIM. The paper then introduces two modifications, a memory-enhanced gradient update (SAME-GD) and a deep-image-prior parameterization, that break the symmetry pull and improve reconstruction quality. If the account is right, invariant models are substantially harder to invert naively, and future reconstruction attacks must be designed to break symmetry explicitly.

What carries the argument

The load-bearing objects are the orbit of an input under a finite orthogonal group, the orbitope (the convex hull of that orbit), and the invariant subspace $V^G$ of points fixed by every group element. Proposition 2 is the main mechanism: for a $G$-invariant objective, one gradient descent step is $G$-equivariant, and $\mathrm{Stab}_G(x_{t-1}) \subseteq \mathrm{Stab}_G(x_t)$, so iterates cannot decrease their symmetry and are drawn to the orbit average, the point with the largest stabilizer. SAME-GD deliberately violates this nesting by aggregating the current point with a previous point from which the orbit average has been removed, while deep image prior replaces direct optimization over pixels by optimization over a randomly initialized CNN whose natural-image prior breaks the symmetry bias.

What would settle it

Train a two-layer symmetrized ReLU network on MNIST with a two-element reflection group, run KKT reconstruction with many random seeds, and compute the stabilizer of each final reconstruction together with its distance to the orbit average; if a large fraction of reconstructions are not fixed by the reflection and lie far from the average, the paper's claim that reconstructions gravitate to $V^G$ is contradicted.

Watch

Extended reading notes

Core claim

The central claim is that group invariance, not network capacity or dataset size, is what makes standard reconstruction methods collapse. For any finite orthogonal group $G$, if the reconstruction loss is $G$-invariant, then each gradient-descent step is $G$-equivariant and the stabilizer of the current iterate is nested inside the stabilizer of the next, so optimization can only move toward points with larger symmetry; the fixed points of this dynamics lie in the invariant subspace $V^G$, and the orbit average is the natural attractor. The paper proves this in two propositions and shows empirically that reconstructions from AM and KKT objectives concentrate near group averages on orbitopes. It reports that increasing the group size or training set size worsens the effect, and that KKT-based reconstruction beats activation maximization but still fails. Its proposed fixes, SAME-GD and deep image prior, both improve standard error metrics while producing asymmetric, piecewise-smooth reconstructions, especially when combined with the KKT objective.

Load-bearing premise

The KKT-based reconstruction results assume the trained invariant network is homogeneous and sits exactly at a margin-maximizing stationary point, so its weights are a weighted sum of gradients at the true training inputs, $\theta = \sum_i \lambda_i y_i \nabla_\theta \phi(x_i; \theta)$; the paper asserts this holds for the symmetrized ReLU networks but does not verify it.

Editorial extensions

If this is right

  • Invariant models are harder to invert with off-the-shelf reconstruction: both AM and KKT return images close to the orbit average, with DSSIM around 0.46–0.49 on MNIST for groups of size 2–8.
  • Reconstruction quality degrades as the group or training set grows, so larger symmetry groups provide stronger accidental protection against naive attacks.
  • The KKT objective is a better starting point than activation maximization, and combining it with deep image prior gives the best reported reconstructions.
  • Evaluation of reconstruction attacks on invariant models should use a group-invariant metric such as min-over-orbit DSSIM to fairly compare across group transformations.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The stabilizer-nesting argument applies beyond reconstruction: any gradient-based inversion or latent-space search over a symmetry-invariant objective will inherit the same drift toward fixed points, so the result predicts similar collapse in equivariant generative models.
  • Because the orbit average is a natural attractor, an invariant network leaks the orbit average but hides in-orbit details; a practical privacy implication is that group invariance acts as an accidental defense against naive attacks, though an adversary who knows the group can still recover the orbit.
  • SAME-GD removes the group-average component from a stored previous iterate, which suggests a principled design: any update that decreases the stabilizer of the current point should improve reconstruction, and quantifying how much symmetry breaking is needed per group size would be a direct follow-up.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper studies the reconstruction of training data from group-invariant neural networks. It formalizes the reconstruction problem for invariant models, proves basic invariance properties of reconstruction objectives, and shows that gradient descent iterates on such objectives are equivariant and have nested stabilizers. The central empirical claim is that conventional activation maximization and KKT-based reconstruction methods converge to highly symmetric inputs, typically near the group average, yielding poor reconstructions. The paper then proposes two heuristics, SAME-GD and a deep-image-prior parameterization, and reports preliminary MNIST and CIFAR-10 experiments suggesting that these can improve reconstruction quality.

Significance. If the findings hold, this is a useful first investigation of a realistic and privacy-relevant setting: reconstructing data from networks that are invariant under finite groups. The formal problem definition and the stabilizer-nesting result for gradient descent (Proposition 2) are conceptually clean and provide a plausible explanation for the observed convergence to orbit averages. The proposed methods are heuristic but could open a useful direction. The paper does not appear to be circular: the main qualitative prediction follows from the theory rather than from fitting the model to the data. However, the manuscript as printed contains a dimensionally invalid KKT objective, an unverified KKT-stationarity assumption for the trained symmetrized networks, and a table that does not support one of the main claims about SAME-GD. These issues are load-bearing and need to be fixed before the empirical claims can be verified.

major comments (4)
  1. [Section 2 / Appendix B, Eq. (3)] The printed KKT reconstruction objective is dimensionally invalid: L_rec = ||θ − Σ_i λ_i y_i φ(x_i; θ)|| subtracts scalar network outputs from the parameter vector θ. The KKT stationary condition of the margin-maximization problem in Eq. (2) is θ = Σ_i λ_i y_i ∇_θ φ(x_i; θ), so the printed formula omits the parameter gradient ∇_θ φ. Lemma 3 in Appendix C proves G-invariance of ∇_θ φ, which is exactly the quantity needed for the loss to be invariant, confirming that the intended objective contains the gradient. Because Table 1 and the SAME-GD/DIP variants are evaluated on top of this objective, the experiments as described are not reproducible from the printed equations. The authors should correct Eq. (3) and explicitly state whether the experiments used the corrected objective with ∇_θ φ.
  2. [Section 3.1] The paper asserts that 'in most cases the training of (homogeneous) invariant neural networks is conducted in a way that the conditions of both methods (AM and KKT-based) are met,' but no proof or verification is provided for the symmetrized ReLU networks used in Appendix E. The KKT-based reconstruction method requires convergence of gradient flow to a KKT point of the margin-maximization problem, which is a nontrivial assumption. The authors should either verify the stationary condition numerically on the trained models (for example, by computing the relative KKT residual) or weaken the claim to state that the method is applied heuristically. Without this, the KKT reconstructions are not anchored to the stated margin-maximization theory.
  3. [Table 1 / Section 4] The text states that SAME-GD yields 'notably improved reconstructions,' but Table 1 does not support this uniformly. For CIFAR-10 the KKT+SAME-GD column is worse than vanilla KKT on both training-set sizes (0.369 ± 0.000 vs 0.346 ± 0.001 for n=50; 0.370 ± 0.007 vs 0.371 ± 0.001 for n=100). For MNIST with group size 8 the improvements are very small (0.465 vs 0.471 and 0.469 vs 0.471). The paper should report the comparisons with error bars, state which differences are significant, and adjust the claim to the actual results, or explain the configurations in which SAME-GD does and does not help.
  4. [Section 3.3 / Figure 5] The central empirical claim is that conventional methods 'gravitate toward symmetric inputs' on which the group acts trivially, but the paper never quantifies this directly. Table 1 reports only DSSIM structural dissimilarity, and Figure 5 is a qualitative histogram of reconstruction locations on discretized orbitopes. The authors should add a quantitative symmetry measure, for example ||x − ḡx||/||x|| or the size of the stabilizer of the reconstruction, and report its distribution across methods. This would make the main phenomenon directly testable rather than inferred from image quality and a qualitative figure.
minor comments (5)
  1. [Section 3.2, Proposition 2] Part (ii) states Stab_G(x_{t−1}) ⊆ Stab_G(x_t), meaning that stabilizers can grow as optimization progresses, so points become more symmetric; the accompanying text saying that stabilizers 'may become more restrictive' is backwards and should be reworded.
  2. [Appendix C, Eq. (4)] The proof of Lemma 3 contains a typo: the right-hand side should be (f(g·x; θ+εe_i) − f(g·x; θ))/ε, not (f(g·x; θ+εe_i) − f(x; θ))/ε.
  3. [Appendix D, proof of Lemma 5] The inner-product derivation contains a stray dot in the expression '<g^{-1}∇L(g·x), ·g^{-1}·h>' and should be cleaned up.
  4. [Algorithm 1 / Appendix F] In the x_prev update, 'xprev ← β_t x_t + (1−β_t)∇L(x_t)' mixes a point and a gradient; the paper should specify the intended gradient with respect to x and clarify that ∇L(x_t) is used as a direction, not as a coordinate.
  5. [Figure 3] Figure 3 is referenced in Section 3.3 but appears only after Appendix D in the submitted layout; the figure should be placed with its first citation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation chain found: the central invariance results are proven from definitions, the KKT baseline is adopted from independent prior work, and the proposed methods are experimental rather than fitted predictions.

full rationale

I found no step in this paper where a 'prediction' or derived result is equivalent to its inputs by construction. Proposition 2 (stabilizer nesting under GD) is proven from the definition of G-invariance and orthogonality of the representation, not from fitted parameters or from the later experiments. The empirical observation that reconstructions concentrate near the orbit average is presented as an observed phenomenon and only loosely connected to Proposition 2; even if that connection is overstated, it is not a circular reduction. The KKT objective is taken from Haim et al. (2022) and Lyu & Li (2020); the former shares an author with this paper, but it is a parameter-free, externally stated stationarity condition and is not used to forbid alternative explanations, so the self-citation is not load-bearing. SAME-GD and DIP are proposed algorithmic modifications whose improved DSSIM values are experimental outcomes, not quantities forced by the loss definition. Two non-circular but important flaws should be flagged: (1) Eq. (3) and Appendix B print the KKT loss as ||θ − Σ λ_i y_i φ(x_i; θ)|| without the parameter gradient ∇_θ φ, making the expression dimensionally invalid; Lemma 3 correctly concerns ∇_θ φ, so the printed formula appears to be a typo that nevertheless prevents verification of the reported KKT experiments. (2) Section 3.1 asserts without proof that trained symmetrized ReLU networks satisfy the KKT conditions of both methods. These are correctness and reproducibility concerns, not circularity: none of the paper's claims reduce to its own definitions or to a self-citation chain.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper's central empirical claim relies on the margin-maximization KKT theory, on orthogonality and differentiability of the group action, and on the validity of the orbitope binning and DSSIM evaluation. The reconstruction methods add ad hoc hyperparameters (SAME-GD, DIP, m) that are not swept or fully specified.

free parameters (3)
  • SAME-GD hyperparameters (α_t, β_t, T_save, T_update) = not specified
    Algorithm 1 introduces these coefficients and schedules ad hoc; no values or sensitivity study are given, and the update rule is not derived from the stability theory.
  • number of reconstruction candidates m = 1000 (500 per class)
    Chosen in Appendix E; reconstruction quality could depend on this overparameterization (m ≫ n) and is not swept.
  • DIP network architecture and iterations = not specified
    The deep image prior method requires a randomly initialized CNN parameterization; architecture, noise level, and optimization schedule are not reported.
assumptions (4)
  • domain assumption The trained symmetrized ReLU network is homogeneous with respect to θ and its gradient flow converges to a KKT point of the margin maximization problem (Eq. 2).
    Invoked in Section 3.1 and Appendix B to justify the KKT reconstruction objective; not verified for the symmetrized invariant models used in Appendix E.
  • standard math The group representation is orthogonal and the reconstruction loss is differentiable with respect to x.
    Required for Lemma 5 and Prop 2; orthogonality is used to move group elements through inner products, and differentiability is assumed in the derivation.
  • domain assumption The reconstruction objectives are G-invariant functions of the inputs.
    Prop 1 establishes this for AM and, for the KKT objective, depends on the invariance of the parameter gradient; as printed Eq. 3 holds trivially for outputs.
  • domain assumption Nearest-neighbor binning over discretized orbitopes is a valid characterization of where reconstructions land.
    Used in Figure 5 to claim concentration near the group average; discretization resolution and nearest-neighbor rule are not described.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the Reconstruction of Training Data from Group Invariant Networks." pith.science (2026). https://pith.science/paper/6TDKNNCU

@misc{pith2026241116458,
  author       = {Pith},
  title        = {Pith review of: On the Reconstruction of Training Data from Group Invariant Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6TDKNNCU}},
  note         = {Machine review of arXiv:2411.16458}
}
read the original abstract

Reconstructing training data from trained neural networks is an active area of research with significant implications for privacy and explainability. Recent advances have demonstrated the feasibility of this process for several data types. However, reconstructing data from group-invariant neural networks poses distinct challenges that remain largely unexplored. This paper addresses this gap by first formulating the problem and discussing some of its basic properties. We then provide an experimental evaluation demonstrating that conventional reconstruction techniques are inadequate in this scenario. Specifically, we observe that the resulting data reconstructions gravitate toward symmetric inputs on which the group acts trivially, leading to poor-quality results. Finally, we propose two novel methods aiming to improve reconstruction in this setup and present promising preliminary experimental results. Our work sheds light on the complexities of reconstructing data from group invariant neural networks and offers potential avenues for future research in this domain.

Figures

Figures reproduced from arXiv: 2411.16458 by the authors.

Figure 1
Figure 1. Visualization of or￾bitope with G = D4 Recent works (Haim et al., 2022; Oz et al., 2024; Loo et al., 2023) have shown that it is possible to reconstruct training data from standard neural networks. However, the reconstruction from group invariant neural networks, such as networks applied to point clouds (Zaheer et al., 2017; Qi et al., 2017), graph data (Gilmer et al., 2017) or images with rotation and reflection sy… view at source ↗
Figure 2
Figure 2. Pairs of training samples and their corresponding nearest neighbors reconstruc￾tions on their left, where n = 50, |G| = 2 is the group of right-left reflections. questions about the behavior of reconstruction methods applied to these networks. Al￾though we provide some insights and novel approaches, It is still unclear why standard reconstruction methods fail for invariant models. There are many future directions to… view at source ↗
Figure 3
Figure 3. The mean DSSIM over MNIST training subsets with varying size and different groups. Appendix E. Experimental setting Setting. We focus on image data and considered 4 groups for our experiments - the trivial group, group of 2 elements acting as horizontal reflection, the group of 4 elements acting as horizontal and vertical reflection G4 (Klein four-group), and the Dihedral group D4 (rotations and reflections). To con… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Pairs of training samples and their corresponding nearest neighbors reconstruc￾tions on their right , where n = 50, |G| = 2. (a) (b) (c) (d) (e) (f ) [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: The empirical distribution of reconstructions across orbitopes using KKT-based method on MNIST-trained invariant networks. Orbitopes are discretized into bins, each representing a convex combination of orbit elements. Reconstructions are assigned to bins based on their…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 9 canonical work pages

  1. [1]

    Baker, Alexander Pinard, and Dorit M

    Allison H. Baker, Alexander Pinard, and Dorit M. Hammerling. Dssim: a structural similarity index for floating-point data, 2023. URL https://arxiv.org/abs/2202.02616

  2. [2]

    Deconstructing data reconstruction: Multiclass, weight decay and general losses

    Gon Buzaglo, Niv Haim, Gilad Yehudai, Gal Vardi, Yakir Oz, Yaniv Nikankin, and Michal Irani. Deconstructing data reconstruction: Multiclass, weight decay and general losses. In Advances in Neural Information Processing Systems, volume 36, pages 51515--51535, 2023

  3. [3]

    Group equivariant convolutional networks

    Taco Cohen and Max Welling. Group equivariant convolutional networks. In International conference on machine learning, pages 2990--2999. PMLR, 2016

  4. [4]

    Model inversion attacks that exploit confidence information and basic countermeasures

    Matt Fredrikson, Somesh Jha, and Thomas Ristenpart. Model inversion attacks that exploit confidence information and basic countermeasures. In Proceedings of the 22nd ACM SIGSAC conference on computer and communications security, pages 1322--1333, 2015

  5. [5]

    Inverting gradients-how easy is it to break privacy in federated learning? Advances in Neural Information Processing Systems, 33: 0 16937--16947, 2020

    Jonas Geiping, Hartmut Bauermeister, Hannah Dr \"o ge, and Michael Moeller. Inverting gradients-how easy is it to break privacy in federated learning? Advances in Neural Information Processing Systems, 33: 0 16937--16947, 2020

  6. [6]

    Neural message passing for quantum chemistry

    Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In International conference on machine learning, pages 1263--1272. PMLR, 2017

  7. [7]

    Reconstructing training data from trained neural networks

    Niv Haim, Gal Vardi, Gilad Yehudai, Ohad Shamir, and Michal Irani. Reconstructing training data from trained neural networks. Advances in Neural Information Processing Systems, 35: 0 22911--22924, 2022

  8. [8]

    Deep models under the gan: information leakage from collaborative deep learning

    Briland Hitaj, Giuseppe Ateniese, and Fernando Perez-Cruz. Deep models under the gan: information leakage from collaborative deep learning. In Proceedings of the 2017 ACM SIGSAC conference on computer and communications security, pages 603--618, 2017

Show all 21 references
  1. [9]

    Evaluating gradient inversion attacks and defenses in federated learning

    Yangsibo Huang, Samyak Gupta, Zhao Song, Kai Li, and Sanjeev Arora. Evaluating gradient inversion attacks and defenses in federated learning. Advances in Neural Information Processing Systems, 34: 0 7232--7241, 2021

  2. [10]

    Directional convergence and alignment in deep learning

    Ziwei Ji and Matus Telgarsky. Directional convergence and alignment in deep learning. In Advances in Neural Information Processing Systems (NeurIPS), 2020

  3. [11]

    Dataset distillation fixes dataset reconstruction attacks

    Noel Loo, Ramin Hasani, Mathias Lechner, and Daniela Rus. Dataset distillation fixes dataset reconstruction attacks. arXiv preprint arXiv:2302.01428, 2023

  4. [12]

    Gradient descent maximizes the margin of homogeneous neural networks

    Kaifeng Lyu and Jian Li. Gradient descent maximizes the margin of homogeneous neural networks. In International Conference on Learning Representations (ICLR), 2020

  5. [13]

    Reconstructing training data from real world models trained with transfer learning

    Yakir Oz, Gilad Yehudai, Gal Vardi, Itai Antebi, Michal Irani, and Niv Haim. Reconstructing training data from real world models trained with transfer learning. arXiv preprint arXiv:2407.15845, 2024

  6. [14]

    Pointnet: Deep learning on point sets for 3d classification and segmentation

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652--660, 2017

  7. [15]

    Orbitopes

    Raman Sanyal, Frank Sottile, and Bernd Sturmfels. Orbitopes. Mathematika, 57 0 (2): 0 275–314, June 2011. ISSN 2041-7942. doi:10.1112/s002557931100132x. URL http://dx.doi.org/10.1112/S002557931100132X

  8. [16]

    Deep image prior

    Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. International Journal of Computer Vision, 128 0 (7): 0 1867–1888, March 2020. ISSN 1573-1405. doi:10.1007/s11263-020-01303-4. URL http://dx.doi.org/10.1007/s11263-020-01303-4

  9. [17]

    Adapting membership inference attacks to gnn for graph classification: Approaches and implications, 2021

    Bang Wu, Xiangwen Yang, Shirui Pan, and Xingliang Yuan. Adapting membership inference attacks to gnn for graph classification: Approaches and implications, 2021

  10. [18]

    Neural network inversion in adversarial setting via background knowledge alignment

    Ziqi Yang, Jiyi Zhang, Ee-Chien Chang, and Zhenkai Liang. Neural network inversion in adversarial setting via background knowledge alignment. In Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, pages 225--240, 2019

  11. [19]

    Deep sets

    Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Russ R Salakhutdinov, and Alexander J Smola. Deep sets. Advances in neural information processing systems, 30, 2017

  12. [20]

    Graphmi: Extracting private graph data from graph neural networks, 2021

    Zaixi Zhang, Qi Liu, Zhenya Huang, Hao Wang, Chengqiang Lu, Chuanren Liu, and Enhong Chen. Graphmi: Extracting private graph data from graph neural networks, 2021

  13. [21]

    Deep leakage from gradients

    Ligeng Zhu, Zhijian Liu, and Song Han. Deep leakage from gradients. Advances in Neural Information Processing Systems, 32, 2019

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.