acceptodds
Under review as a conference paper at ICLR 2027

Where Priors Leave Fingerprints: Identifying Learned Regularizers from Reconstruction Response

Abstract

Many learned priors for inverse problems reconstruct equally well, so the reconstruction alone reveals little about which prior produced it. Model fingerprints read the outputs of a model, and the classical analysis of how a solver reacts to its data assumes a known regularizer. In this paper, we ask what the behavior of a solver reveals about its learned prior beyond the reconstruction, and when it suffices to distinguish two priors. We study the response of a solver, namely how its reconstruction changes when the measurement is randomly perturbed. For solvers that combine a least-squares data fit with a learned regularizer, we prove that, when the operator and the perturbation level are known, the covariance of random re-solves determines the Jacobian of the solver in closed form, which identifies the curvature of the regularizer on the subspace that the solver can reach. Guided by this result, we introduce the response fingerprint, built from the centered energies of the response in frequency bands, which applies to any differentiable solver. Across denoising, deblurring, inpainting, super-resolution, MRI and CT, it separates learned priors that are hard to tell apart from their outputs, and it transfers without retraining to external priors, which played no role in its design. It also provides a label-free reliability indicator that flags the priors it separates poorly.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
inverse problems, reconstruction, regularization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_UNpqOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates whether learned priors can be identified from how reconstruction solvers respond to measurement perturbations, even when their outputs are similar. The authors propose a response fingerprint based on centered log energies in frequency bands, estimated through repeated solves with randomly perturbed measurements. Attribution is performed by matching the fingerprint to reference centroids of known candidate priors. For variational solvers, the paper proves that, under local regularity assumptions and with the operator and perturbation level known, the first-order response covariance determines the solver Jacobian and the regularizer’s curvature on the reachable subspace. Experiments across nine inverse problems and external reconstruction models demonstrate promising attribution performance, supported by extensive comparisons and ablations. Overall, I appreciate the proposed method and its theoretical analysis. However, I remain concerned about its practical contribution, particularly how identifying a prior informs meaningful decisions or provides useful guidance for solving imaging inverse problems.

Critical strengths

S1. The research perspective is novel. Starting from the observation that "similar reconstruction results may arise from different priors," the paper clearly formulates the problem of prior identification and provides a relatively detailed theoretical analysis.

S2. The experiments cover various inverse problems and different types of priors, with relatively comprehensive comparisons and ablation studies. The proposed method achieves good attribution performance.

S3. The paper is well-structured and coherently written, clearly presenting the research problem, method design, and underlying ideas.

Critical weaknesses

W1. The practical value needs further clarification. Although I appreciate the proposed method and theory, I find it difficult to identify concrete practical uses or understand what guidance the method provides for solving imaging inverse problems. In imaging inverse problems, we typically care more about reconstruction quality and reliability. If different priors produce equally good reconstructions, the additional value of identifying which prior was used remains unclear. The authors mention provenance tracking and model comparison, but the experiments primarily evaluate attribution accuracy and do not sufficiently explain how attribution results can inform practical decisions. The authors should clarify: in what scenarios is evaluating reconstruction results alone insufficient, making identification of the underlying prior necessary or useful?

W2. The connection between theoretical identifiability and the practical fingerprint requires clarification. Theorem 1 relies on the full first-order response covariance, whereas the practical method retains only a small number of frequency-band energies and removes the overall scale. Consequently, different covariance matrices may still yield identical fingerprints. The authors should clarify to what extent theoretical identifiability supports the practical compressed fingerprint, and which differences between priors may be lost during compression.

W3. The target of attribution needs to be further clarified. I am not entirely sure whether the fingerprint identifies the prior itself or the entire reconstruction process, including the solver and its parameter settings. Although the theoretical analysis focuses on variational solvers that satisfy certain conditions, the actual reconstruction results can also be affected by factors such as regularization strength, the number of iterations, and randomness. This means that even with the same prior, changing the solver settings may produce different fingerprints. Therefore, the authors should further explain to what extent the proposed method can truly identify the prior, rather than simply distinguish between different reconstruction settings.

Questions for authors

Q1. Could the authors provide concrete application scenarios explaining why evaluating reconstruction quality alone is insufficient, and what practical decisions or guidance for solving inverse problems can be supported by identifying the prior?

Q2. Theorem 1 relies on the full response covariance, whereas the practical fingerprint retains only a small number of frequency-band energies. Could the authors explain what identifiable information is preserved after compression and which differences between priors may be lost?

Q3. When the prior remains unchanged but the regularization strength, number of iterations, or solution algorithm changes, can the method still identify it as the same prior? Should the attribution result be interpreted as identifying the prior or the complete reconstruction configuration?

4 comments
AuthorsOpenReview ↗

Author Responses (Part 1)

Thank you for the positive and constructive review. We are glad that you found the research perspective novel, that the paper clearly formulates the problem of prior identification and provides a relatively detailed theoretical analysis, that the experiments offer relatively comprehensive comparisons and ablation studies, and that the paper is well-structured and coherently written.

In short: your review asks what identifying the prior adds when reconstructions are equally good, what the compressed fingerprint keeps, and whether it identifies the prior or the configuration. We added:

  1. concrete scenarios, each with the decision it supports and a measurement, and a direct test of priors whose reconstructions are equally good;
  2. Proposition 5, which states what the compressed fingerprint keeps and what it sets aside;
  3. the same prior under a changed regularization strength, number of iterations, solution algorithm and initialization.

Regarding when the reconstruction alone is not enough (W1, Q1). Thank you for this question, which helped us sharpen the motivation of the paper. Reconstruction quality is measured with the ground truth, which a deployed system does not have, and two priors of equal quality can still react differently to noise in the same measurement (Figure 1). For this reason, the reliability of a learned reconstruction is established for a specific model, by its validation and clearance, and the question in practice is which model produced an image. Each scenario below is a decision that rests on this question (Appendices J and L):

ScenarioDecisionMeasurement
A physicist or a regulator checks a deployed systemwhether it runs the cleared model or must be re-validatedverification AUC 0.896 to 0.946 (Table 3), and 0.975 to 0.995 with the per-prior standardization on held-out test images (Table 17); from 8 measurements, attribution at 0.950 to 1.000
A model may have been updated or retrainedwhether a new validation is neededtwo training seeds of one model told apart in 81.2% of test clouds (random guess 50%), compared with 66.6% from the reconstruction (the mean of the re-solves) and 61.2% from the total response
Two systems are comparedwhether they run the same prior, with neither among the candidatesAUC 0.962 / 0.973 (N=8 / 64)
A reconstruction is publishedwhich model produced itthe test below

Inspired by your question, we tested the case of equally good reconstructions directly. On the internal set, on the clouds of Table 18 that no hyperparameter selection used, we collected every test image on which two candidate priors reconstruct within 0.5 dB PSNR of each other (3,067 cases, 6,134 two-way decisions):

Two priors within 0.5 dB PSNRCorrect prior
Matching the PSNR of the test reconstruction, computed with the ground truth, to that of the reference reconstructions51.8% (random guess 50%)
Our fingerprint, which needs no ground truth (N=64, per-prior standardization)84.4%

On these cases, which reconstruction quality cannot separate, the fingerprint identifies the prior in 84.4% of the decisions. For solving inverse problems, the response also shows where a reconstruction is set by the prior rather than by the data: by Example 1, the band energies are set by the operator at the frequencies it passes and by the prior at those it attenuates. Following your comment, Section 1 now states these scenarios, Appendix J describes them, and Appendix L reports their measurements.

AuthorsOpenReview ↗

Author Responses (Part 2)

Regarding what the compression keeps and what it loses (W2, Q2). We appreciate this question on what the compression loses. Inspired by it, we added Proposition 5 (Appendix K, summarized in Section 4): two responses have the same fingerprint if and only if their band energies are proportional. Theorem 1 states what the response contains, the curvature of the prior on the subspace that the solver can reach, and the fingerprint keeps how this response is distributed across the 12 frequency bands: how strongly the prior regularizes each band that the operator attenuates, relative to the other bands (Example 1). It sets aside three properties:

  • The overall size of the response, by design, which makes the fingerprint independent of the perturbation level; in our ablation, the centered log band energies raise the accuracy from 0.773 to 0.808 (Table 19).
  • Where in the image and in which direction the response lies within a band, and how the bands are correlated. A circular translation of the spatially varying weights of a prior changes C by 30% to 37% (over seeds and operators) and leaves the fingerprint unchanged (Proposition 5(ii)). The direction depends on the acquisition: with the acquisition rotated by 90°, our fingerprint attributes at 0.837 / 0.853, while the Cartesian cells with the same standardization decrease to 0.270 / 0.280 (Table 21).
  • The curvature outside the reachable subspace, which no method that observes the response can recover (Theorem 1, converse), and which is therefore not a loss of the compression.

With exact covariances, all 1,200 random pairs of priors per operator are separated, and with nine random priors the fingerprint reaches the accuracy of the maximum-likelihood classifier that knows every exact covariance (1.000 at N=64): the compression loses no accuracy for nine random priors at N=64, and pooling makes each band energy measurable from 8 re-solves (Proposition 3).

Regarding the prior or the reconstruction configuration (W3, Q3). Thank you for this question. The attribution identifies the prior. Once the solver has converged to a solution, its response there depends on the objective it minimizes, the data fit plus λ times the regularizer, and not on the algorithm, the number of iterations or the initialization that reached it (Theorem 1). The strength λ scales the prior inside this objective, and the auditor covers an unknown λ by running each candidate prior at a few strengths. Following your question, we tested both directly on 9 inverse problems (N=64, random guess 0.111; Tables 20 and 22 and Appendix L):

The same prior, run withConfigurations among the referencesConfiguration not among the references
L-BFGS instead of Adam*0.8700.811
0.5× / 2× / 4× the iterations0.853 / 0.848 / 0.8360.775 / 0.830 / 0.778
Zero / random initialization*0.822 / 0.8810.731 / 0.851
0.25× / 0.5× / 2× / 4× its λ0.810 / 0.849 / 0.850 / 0.8330.829 / 0.755 (λ between / outside the referenced strengths, standardized, Table 20)

In the first column, every candidate prior is run under the listed configurations, and the re-solves of a test measurement are attributed to the closest (prior, configuration) reference; in the second, the configuration of the test is removed from the references. *For the 7 explicit regularizers, whose objective defines the solution. Under every change in this table, the same prior is attributed to itself in at least 73.1% of the test clouds, 6.6 times the random guess or more. When the configuration or λ of the test is not among the references (Table 20), the total response decreases from 0.737 to between 0.367 and 0.534, whereas our standardized fingerprint attributes the same test clouds at 0.755 to 0.829. With all 9 candidate priors matched in total response, by calibrating the size of the measurement perturbation for each prior, the total response attributes at 0.311, while our fingerprint attributes at 0.860 (N=64, Table 23). The total response follows the configuration, and the fingerprint follows the prior; following your question, Section 5.4 now summarizes this result. A random initialization leaves the attribution at 0.881 / 0.851 (Table 22), and when the objective itself is stochastic, re-solving the unperturbed measurement measures the randomness of the solver, which is subtracted band by band from the response (Appendix E.2).


Thank you again for your positive recommendation and your questions, which helped us clarify the scope and the practical value of the paper. We hope that these additions address your concerns and that you will consider revising your score; we are happy to discuss any remaining questions.

Thank you for the detailed responses and additional experiments. The concrete application scenarios clarify the practical motivation. I appreciate the authors' efforts to address my concerns. After considering the responses and revisions, I will maintain my original recommendation.

Dear Reviewer UNpq,

We thank you for the continued support in our work, and for your actionable and detailed feedback, which we feel helped us to improve the quality of our paper.

With kindest regards,

Authors.

Reviewer_4BruOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether learned priors can be identified by looking at how their reconstructions change under small measurement perturbations. The authors propose a fingerprint based on the frequency content of these changes and provide a theoretical connection to the curvature of the underlying regularizer. They evaluate the approach across several inverse problems, including MRI and CT.

Critical strengths
  1. I think the idea of identifying learned priors through reconstruction responses is interesting and well motivated. In particular, examining how reconstructions change under small measurement perturbations provides a useful alternative to conventional model fingerprinting methods that primarily examine reconstruction outputs.
  2. I appreciate the extensive experimental evaluation across different inverse problems, including denoising, deblurring, inpainting, super-resolution, MRI, and CT. I also find the experiments on external pretrained priors and real MRI measurements particularly valuable, as they provide evidence that the proposed approach generalizes beyond the models used in its development.
Critical weaknesses
  1. I find the practical attribution setting somewhat restrictive. The proposed method requires repeated access to the unknown reconstruction solver, as well as access to the candidate solvers to construct reference fingerprints. Furthermore, attribution is evaluated only when the true prior belongs to the known candidate set. I would have liked a clearer discussion of practical scenarios where these requirements can be satisfied, particularly for the proposed provenance applications. Although the theoretical analysis is interesting, it also establishes identifiability from the full response covariance under specific assumptions, whereas the actual fingerprint compresses this information into a small number of frequency-band statistics. I find it unclear how much of the theoretical identifiability carries over to this compressed representation.
  2. Although the experimental results are encouraging, I find the attribution performance less convincing for some of the strongest reconstruction methods. For example, the external evaluations achieve 58.9-68.0% accuracy, and Appendix E.4 reports that the most accurate unrolled MRI reconstruction networks are identified with approximately 53% accuracy. The authors acknowledge and theoretically investigate this limitation, which I appreciate. Nevertheless, distinguishing similarly performing reconstruction methods is a central motivation of the paper, and these results raise questions about the practical effectiveness of the proposed fingerprint in precisely such settings. Furthermore, some of the improvements for difficult cases require substantially larger perturbations and more solver iterations, potentially increasing the computational cost of attribution.
Questions for authors
  1. Could the authors discuss how the proposed framework would handle a solver whose prior is not included among the known candidates? In particular, could the fingerprint distances be used to identify an unknown prior rather than always assigning it to one of the existing candidates?
  2. As I understand it, the reliability indicator ρk\rho_k is computed from fingerprints associated with each candidate prior, so it measures how well that candidate can be distinguished from the others in general. However, it is unclear how this translates into confidence for an individual attribution. Could the authors clarify how reliability would be assessed for a single test fingerprint from an unknown solver? For example, could the relative distances to the nearest and second-nearest centroids provide a useful confidence measure?
5 comments
AuthorsOpenReview ↗

Author Responses (Part 1)

Thank you for the constructive review. We are glad that you found the idea of identifying learned priors through reconstruction responses interesting and well motivated, that it provides a useful alternative to conventional model fingerprinting methods, and that you found the experiments on external pretrained priors and real MRI measurements particularly valuable.

In short: your review asks whether attribution is practical: what access it needs, whether it extends beyond a closed set of candidates, and how accurate it is on accurate priors, at what cost. We added:

  1. concrete scenarios, each supported by a measurement;
  2. Proposition 5, which states what the compressed fingerprint keeps;
  3. higher accuracy on the accurate external priors at the default budget;
  4. an open-set test for priors outside the candidate set;
  5. the confidence measure you suggested.

Regarding the practical setting (W1). Thank you for asking for the scenarios behind the setting. Attribution answers "which model produced this image" when candidate priors reconstruct equally well: to verify that a deployed system runs the cleared version of a learned reconstruction, to detect a silent update, and to state the provenance of a published reconstruction. The auditor, such as a regulator, a hospital physicist or the vendor, holds the candidate models, the clinical protocol fixes the type of acquisition, and repeated queries of the solver under test are the same kind of access that quality-assurance measurements of clinical reconstructions use. For provenance, the auditor queries the solver that produced a published reconstruction, and the candidates can be publicly released models, as in our three external sets. The reference clouds are computed once per candidate and reused for every test image. Each scenario is supported by a measurement (Appendices J and L):

ScenarioMeasurement
A silent update or a retrained versiontwo training seeds of the same model told apart in 81.2% of test clouds (total response 61.2%)
Candidates that reconstruct alikewithin 0.5 dB PSNR, which reconstruction quality cannot separate, our fingerprint, which needs no ground truth, decides 84.4% of 6,134 two-way decisions correctly (random guess 50%; per-prior standardization)
Another algorithm, iteration budget, initialization or operator instancethe same prior attributed to itself at 0.822 to 0.896, and at 0.731 to 0.851 when its configuration is not among the references (Table 22, random guess 0.111)

The candidate set need not contain the true prior either: from 8 images of the solver, a prior outside the candidate set is separated from the known priors with AUC 0.959 to 1.000 (per-prior standardization, Appendix J). Following your comment, Section 1 now states these scenarios, Appendix J describes them, and Appendix L reports their measurements.

Regarding what the compressed fingerprint keeps (W1). We appreciate this question on the compressed fingerprint. Inspired by it, we added Proposition 5 (Appendix K, summarized in Section 4): two responses have the same fingerprint if and only if their band energies are proportional. The fingerprint keeps how the response is distributed across frequency bands and sets aside, by design, its overall scale, which depends on the perturbation level, its arrangement within each band, and the correlations between bands. Within an analytic family of priors in which two members differ in fingerprint, equal fingerprints are exceptional (a set of measure zero). With exact covariances, all 1,200 random pairs of priors per operator are separated, and with nine random priors the fingerprint reaches the accuracy of the maximum-likelihood classifier that knows every exact covariance (1.000 at N=64). On the internal set, when the acquisition of three inverse problems is rotated by 90° at test time, our 12 numbers attribute at 0.762 / 0.760 (N=8 / 64), while the full 2D power spectrum (one number per frequency cell) decreases to 0.286 / 0.290 with the same distance (Table 21). Pooling also makes each band energy measurable from 8 re-solves (Proposition 3).

AuthorsOpenReview ↗

Author Responses (Part 2)

Regarding accurate priors and the cost of attribution (W2). Thank you for pointing to these hard cases. With the same fingerprint and the per-prior standardized distance (Eq. 16), the same 8 solver calls per test image and the default perturbation and iteration budget, the external priors of Table 3 (58.9% to 68.0%) are attributed at 0.828 to 0.935 on held-out test images, and the unrolled MRI networks (near 0.531 in Appendix E.4) at 0.688 to 0.859 from a single slice. As you note, Proposition 4 shows why accurate priors are harder for any method that reads the behavior of the solver, and ρ_k flags them in advance; Proposition 3 adds that the band energies are estimated with different precision. Following your comment, we measure distances in units of the spread of the reference fingerprints around their centroids (standardization), pooled over the candidates and, in addition, for each prior (per-prior standardization, Eq. 16, with three hyperparameters selected on development data). This changes only the distance of the nearest-centroid step, uses the reference clouds and adds no call to the solver (Table 17):

Held-out test imagesAttribution of Table 3StandardizedPer-prior standardized [95% CI]Verification AUC
Denoising, 29 images per prior0.6110.8920.921 [0.887, 0.956]0.990
CT, 48 images per prior0.6670.8780.935 [0.908, 0.958]0.995
Real MRI, volumes 5 to 80.6380.7940.828 [0.788, 0.866]0.975
Real MRI, volumes 1 to 40.7220.8530.897 [0.866, 0.925]0.993

No hyperparameter selection used these test images, and the standardized fingerprint has no hyperparameter to select, because its shrinkage is computed from the reference fingerprints. The verification AUC measures how well the fingerprint separates a true claim "this solver uses prior k" from a false one (1 is perfect, 0.5 is a random guess). From all slices of one test volume, 9 of the 10 MRI networks are attributed correctly on every volume, and the remaining unrolled network on 6 of the 8 volumes. WCRR on CT, which Appendix E.3 separates with a perturbation of 4% instead of 1% and up to 48,000 iterations, reaches 0.917 at the default budget, compared with 0.062 on the 16 test images of Table 11.

Cost, and more evidence at a known cost. Attribution with N=8 calls the solver under test 8 times, as in Table 1, and adds one 2D FFT per re-solve: on one GPU, one reconstruction of the external denoising priors takes 0.06 s (LPN) to 8.7 s (PatchNR), and a test image is attributed in 0.5 to 70 s (Appendix J). An auditor who needs more evidence queries the solver on 8 measurements, at 8 × 8 solver calls, and attributes the external priors at 0.950 to 1.000 with the per-prior standardization. Following your comment, Appendix J now reports the cost and this option.

AuthorsOpenReview ↗

Author Responses (Part 3)

Regarding a prior outside the candidate set (Q1). Thank you for this question. The fingerprint distances flag a prior that is not among the candidates, instead of assigning it to one of them. Following your question, the distance of the standardized fingerprint to each centroid, divided by a threshold that the auditor sets from its own runs of that candidate and minimized over the candidates, gives an open-set test that reports "none of the candidates". Holding out each prior in turn, as if it were unknown (AUC: the probability that an unknown prior scores above a known one; Appendices J and L):

Open-set AUCInternalDenoisingCTReal MRI
One test image0.8430.8260.8650.800

The four priors introduced after 2022 (Appendix D.4) are separated from the known priors with AUC 0.901 (Euclidean fingerprint, thresholds from calibration images), and from 8 images of the unknown solver, with the per-prior standardization, the AUC on the external sets is 0.959 to 1.000. Inspired by your question, we also tested recognizing the same unknown prior again: holding out two priors at a time, the distance between two test fingerprints tells whether they come from the same unknown prior with AUC 0.962 (N=8) and 0.973 (N=64). The revision reports this open-set test in Section 3.4 and Appendices J and L.

Regarding the confidence of a single attribution (Q2). Thank you for this suggestion, which we adopted. As you note, ρ_k states in advance, from runs of the candidate priors alone, which candidates the response separates well (Section 3.4). The ratio of the distances to the nearest and second-nearest centroids complements it for a single attribution, from the test fingerprint alone: the smaller the ratio, the more confident the attribution. Ranked by this ratio (Appendices J and L):

The more confident half of the attributionsCorrect
Internal set, N=64 (all attributions: 83.2%)98.0%
External denoising, standardized (Table 16)100%
External CT, standardized (Table 16)100%
Real MRI, standardized (Table 16)96.9%
Larger denoising and CT test sets, standardized100%

Section 3.4 and Appendix J of the revision now report this confidence measure.


We sincerely appreciate your suggestions on the open-set test and the confidence measure, which strengthened the presentation of our contribution. We hope that these additions address your concerns and that you will consider revising your score; we are happy to discuss any remaining questions.

Reviewer_E2iwOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose to identify which of several known learned priors an inverse-problem solver uses, with only black-box access to the solver, a task they call attribution. Their key idea is that solvers with different priors can produce nearly identical reconstructions yet react differently when the measurement is slightly perturbed. The paper makes three contributions: (i) a "response fingerprint" that re-runs the solver on randomly perturbed measurements and summarizes how the reconstruction varies across spatial frequencies, combined with a nearest-centroid attribution rule and a reliability indicator; (ii) a theoretical result showing that, for variational solvers with a known forward operator and noise level, this response determines the Jacobian of the solver and therefore the curvature of the regularizer on the subspace the solver can reach, but not outside it; and (iii) an empirical study. The validation compares the fingerprint with model-fingerprinting baselines and with other statistics of the response on several imaging problems (denoising, deblurring, inpainting, super-resolution, MRI and CT), using priors trained by the authors: explicit learned regularizers (Fields of Experts, adversarial regularizers, an input-convex network and a learned quadratic) and plug-and-play denoisers (DnCNN and a gradient-step denoiser). It then tests transfer, without retraining, to externally trained priors: the learned regularizers of a recent comparative study (CRR, WCRR, ICNN, TDV, NETT, PatchNR, LPN and EPLL) and pretrained MRI reconstruction networks, mostly unrolled (e.g., VarNet, RecurrentVarNet, RIM and KIKI-net, along with a U-Net), applied to real MRI data. Finally, it evaluates robustness to errors in the operator, the perturbation law, and the regularization strength

Critical strengths
  1. A precise statement of what can be identified: Theorem 1 and its converse state exactly what any method that observes the response of a solver can and cannot learn about a variational regularizer, its curvature on the subspace that the solver can reach, and nothing outside it. Although the ingredients are classical, the characterization is correct and useful, because it sets the limits of black-box attribution independently of the specific fingerprint.

  2. Careful and honest experimental practice: Reference and test images are disjoint, the main comparison is matched in the number of solver calls, confidence intervals are reported, and the external models are used with their authors' weights and without retraining, including on real MRI measurements. The authors also report and analyze their failure cases, such as a convex/weakly convex pair that is confused on CT, the small advantage at denoising, and the lower accuracy on the most accurate priors.

Critical weaknesses
  1. Limited novelty of both the method and the theory. The fingerprint is, in essence, the noise power spectrum (NPS) of the reconstruction, grouped into rings of spatial frequency and normalized to remove its overall magnitude, followed by a nearest-centroid rule. The paper mentions the NPS connection only in Section 3 and does not discuss that this normalized radial NPS is the standard medical-physics tool for characterizing and distinguishing reconstruction algorithms, including commercial deep-learning reconstructions. In model fingerprinting, ProbeSpec already identifies models from the frequency content of their response to input perturbations; yet this closest method is absent from the main comparison and appears only in the appendix, re-implemented with modifications and under a different protocol. The remaining differences (random perturbations at a single level, covariance instead of mean response, normalization) are design choices rather than a new principle, and no ablation shows which of them matters. On the theory side, Theorem 1 combines classical results: the first-order covariance of Fessler, the implicit function theorem, and the uniqueness of the positive semidefinite square root; for denoising, it reduces to the known linear-smoother argument. The difficulty it solves, recovering the Jacobian from the covariance alone, only arises when the perturbations are unknown. In the proposed algorithm, the authors draw the perturbations themselves, so the Jacobian can be estimated directly from the correlation between perturbations and re-solves (Stein's lemma), the principle behind Monte-Carlo SURE (Ramani et al., 2008) and generalized degrees of freedom (Ye, 1998). Moreover, the theory does not cover the fingerprint that is actually used: the theorem requires the operator and the noise level, whereas the fingerprint uses neither and keeps only a few isotropic band energies. The analysis therefore gives a necessary property of the fingerprint, but no condition under which it separates two priors, and much of the evaluation (plug-and-play and unrolled networks, unverified assumptions, MRI magnitude images) lies outside it.

  2. Limited significance and overstated claims: the fingerprint may identify the solver configuration rather than the prior. Attribution is restricted to a closed set of known candidates and requires running each of them with the same operator, regularization strength, perturbation level, and solver settings as the unknown solver, while the practical use case (provenance in medical imaging) is stated in a single sentence. Several results suggest that the fingerprint mostly captures how the solver is configured rather than which prior it uses: the total magnitude of the response alone, a single number largely governed by the regularization strength, already attributes most internal priors correctly; a mismatch in the regularization strength substantially reduces accuracy; a confusion on CT disappears only when the perturbation is increased, and the solver is run much longer to convergence; and for one prior, more than half of the measured response comes from the randomness of the solver itself. The claimed robustness to unknown acquisition conditions is also overstated, because the invariance to errors in the operator holds by construction: the fingerprint never uses the operator.

Questions for authors
  1. Novelty of the fingerprint (Section 3.1, Eq. (8), Table 1, Appendices C and D.2). The fingerprint of Eq. (8) is essentially the normalized radial noise power spectrum ("noise texture") that is routinely used to characterize and compare reconstruction algorithms, including commercial deep-learning CT reconstructions (e.g., Solomon et al., Med. Phys. 47(9), 2020; Greffier et al., Eur. Radiol., 2020), which Section 3.1 mentions only once. ProbeSpec (Lou et al., 2026), the closest fingerprinting method, is absent from Table 1, is re-implemented with a modified input selection (Appendix C), and is reported only in Appendix D.2 under a different protocol. What is new beyond these two, and can you add ProbeSpec to the budget-matched comparison of Table 1, together with an ablation that moves from ProbeSpec to your fingerprint one ingredient at a time (covariance vs. mean response, a single perturbation level, radial vs. zig-zag bands, centering)?

  2. Novelty and scope of the theory (Section 4, Appendices A.1, D.5, E.2 and E.4). Theorem 1 combines the first-order covariance of Fessler (1996), the implicit function theorem and the uniqueness of the PSD square root, and for denoising it reduces to the linear-smoother argument of Buja et al. (1989). Because Algorithm 1 draws the perturbations itself, the Jacobian is directly available from the cross-covariance between perturbations and re-solves (Stein's lemma; Ramani et al., 2008; Ye, 1998), which would also remove solver noise (cf. Appendix E.2). What does Theorem 1 add beyond the case of unobserved perturbations? Moreover, Theorem 1 needs A and σp\sigma_p, whereas Eq. (8) uses neither: beyond Example 1, when do band energies separate two priors whose curvatures differ on T, given that Proposition 4 is only an upper bound? Since many evaluated solvers lie outside the theory (Appendices E.2 and E.4) and a Cartesian partition outperforms the radial one (Section 5.2, Appendix D.5), in what sense is the fingerprint "guided by" Theorem 1, as the abstract states?

  3. Prior versus solver configuration (Tables 2, 14 and 15, Figure 8, Section 5.4, Appendices E.2 and E.3). The total response tr C alone reaches 73.6%, versus 83.5% for the full fingerprint (Table 2); a λ\lambda mismatch drops accuracy from 0.856 to 0.498 at λ/4\lambda/4 (Table 15); the WCRR/CRR confusion on CT disappears only with a larger perturbation and up to 48,000 iterations (Appendix E.3); and 54% of the band energy of PatchNR is solver stochasticity (Appendix E.2). In addition, the invariance to operator error (Section 5.4(i), Table 14) holds by construction, since the fingerprint never uses A. Is the same prior still attributed to itself when it is deployed with a different λ\lambda, iteration budget, initialization or operator than its references, and how much accuracy remains when the candidates are matched in total response? Please qualify the robustness claim of the abstract accordingly.

  4. Practical relevance (Sections 1 and 3.4, Table 3, Appendices E.1, E.4 and E.5). The use case is stated in a single sentence (Section 1). Accuracy is lowest on accurate modern priors: 58.9–68.0% on the external sets (Table 3), with only 16 test images per prior on the regularizer sets (Appendix E.1), and about 53% within the unrolled MRI family (Appendix E.4); Proposition 4 implies that this worsens as priors improve. Moreover, the label-free indicator ρk\rho_k requires a cut-off "calibrated per problem" (Section 3.4, Appendix E.1). Can you describe a concrete scenario that requires closed-set attribution with matched configurations, report results on larger external test sets, explain how the cut-off is set without labels, and show what happens when the true prior is not among the candidates?

  5. Baselines, controls and reproducibility (Tables 1, 2 and 5, Appendices C and I). ModelDiff and SAC stay near chance and decrease with N (Table 5), the CNN classifier reaches 39.6% (Table 2), and the random-basis control (13.0%, Table 2) appears to be at chance by construction, since random directions carry on average the same energy, which centering removes. How were the classifier baselines adapted to image-valued outputs (Appendix C), and how much training data did the CNN have? Would a simple baseline, such as the nearest centroid on per-pixel variance maps or on their full 2D power spectra, close the gap?

11 comments
AuthorsOpenReview ↗

Author Responses (Part 1)

Thank you for the detailed and expert review. We are glad that you found Theorem 1 and its converse a precise statement of what can be identified, a characterization that is correct and useful, because it sets the limits of black-box attribution independently of the specific fingerprint, and our experimental practice careful and honest.

In short: your review asks what the fingerprint adds to the noise power spectrum, ProbeSpec and the classical theory, and whether it identifies the prior or the solver configuration. We ran the experiments you requested:

  1. ProbeSpec, with its published perturbation law, in the budget-matched comparison of Table 1, and the ablation from ProbeSpec to our fingerprint one ingredient at a time;
  2. the noise power spectrum and the Stein estimate of the Jacobian as attribution features, and a new Proposition 5 on when two responses share a fingerprint;
  3. the same prior under another regularization strength, iteration budget, initialization and operator, and candidates matched in total response;
  4. larger external test sets, priors outside the candidate set and the cut-off of ρ_k without labels;
  5. the simple baselines you proposed.

Our fingerprint is Eq. (8) with the nearest centroid of Eq. (10); "standardized" uses the Mahalanobis distance of the pooled reference covariance, and "per-prior standardized" adds the covariance of each prior (Eq. 16, Appendix J). The new results answer both questions in the same way: reading the response in frequency bands carries the attribution, its covariance carries more than its mean (Table 19), and when the strength or the configuration of the test is not among the references, the standardized fingerprint attributes at 0.755 to 0.829 while the total response decreases to 0.367 to 0.534 (Table 20).

Regarding the relation to the noise power spectrum (W1, Q1). Thank you for the references to the NPS literature. The fingerprint reads the same kind of measurement as the NPS, re-solves from repeated acquisitions or Monte Carlo re-noising (Section 3), and asks a different question. In medical physics, the NPS evaluates the image quality of a reconstruction algorithm (noise magnitude, noise texture, detectability), and Solomon et al. [1] and Greffier et al. [2] compare a deep-learning CT reconstruction with filtered back projection and iterative reconstruction. Our contribution regards which learned prior a solver uses: (i) for variational solvers, the response determines λH on the reachable subspace T, and nothing outside it (Theorem 1 and its converse); (ii) the fingerprint keeps the distribution of the response across bands, with an exact condition for when the fingerprints of two priors coincide (Proposition 5); and (iii) an attribution, verification and reliability protocol, validated on 18 externally trained priors without retraining. Following your suggestion, we evaluated the NPS characterizations of these studies as attribution features, on the same clouds and references (9 priors, 9 inverse problems, random guess 0.111; Appendix L):

FeaturesN=8N=64
NPS figures of merit (peak frequency, mean frequency, magnitude)0.7100.760
Normalized radial NPS (band energies divided by their sum)0.7610.801
Centered log band energies (ours)0.8000.832

Dividing the band energies by their sum lets the largest bands dominate the distance, whereas centered log energies compare every band by its relative change, whose estimation error Proposition 3 states. Following your comment, Section 2 now cites both studies, and Appendix H relates the fingerprint to the NPS.

AuthorsOpenReview ↗

Author Responses (Part 2)

Regarding ProbeSpec in the budget-matched comparison of Table 1 (W1, Q1). We appreciate this request. Following your suggestion, we added ProbeSpec to Table 1 as a lower block, with the perturbation law and features of [3] (Sec. 3.2 and 3.3), and with every method on the same measurements and 30 test images per prior (9 priors, 9 inverse problems, random guess 11.1%). Each probe is perturbed on the standardized input by Gaussian noise with ε ∈ {0, 0.05, 0.10, 0.15, 0.20}, with fresh noise for each probe and level, the z-scored outputs form the response matrix, and its 2D DCT gives the six published feature types, matched by cosine. An image-valued solver needs two adaptations: the class-defined probes are replaced by the test measurement, because references and tests share no input, and each output image is z-scored on a fixed 64-dimensional orthonormal sketch in place of the class logits (Appendices C and L).

Method (Table 1, lower block)Top-1 (%) [95% CI]
ModelDiff [4], random perturbations15.5 [13.9, 17.4]
ModelDiff, selected perturbations14.9 [13.8, 15.9]
ProbeSpec, 5 solver calls23.3 [18.6, 30.0]
Response fingerprint (ours), 8 solver calls80.6 [71.3, 87.3]

The paired difference over ProbeSpec is +0.573 [0.475, 0.662], positive on each of the 9 inverse problems (Appendix D.1). On the same clouds, ProbeSpec on all pixels, without the sketch, attributes at 21.1% with 5 solver calls (Appendix C). In the submission, ProbeSpec read clouds in which reference and test share their noise draws, where it reaches 0.354 (Appendix D.2); here, the test perturbations are drawn independently of the references. On the design of Table 5 (Table 18):

Method (Table 18)Solver callsAccuracy
ProbeSpec, 16 probes, 5 levels650.205
ProbeSpec, 2 probes, 5 levels90.219
ProbeSpec, 1 probe, 5 levels50.229
Response fingerprint (ours)80.800
Response fingerprint (ours)640.832

ProbeSpec does not improve with more probes, and Table 1 reports it with 5 calls, where it attributes more test clouds than with 9 or 65. In the revision, Appendix D.1 now reports ProbeSpec per inverse problem, and Appendix C its adaptation.

Regarding the ablation from ProbeSpec to our fingerprint (W1, Q1). Thank you for proposing this ablation. As you proposed, we moved from ProbeSpec to our fingerprint changing one ingredient per step, including each ingredient you listed, on the same clouds (mean of three realizations of the noise and the perturbations, Table 19):

StepChangeAccuracy
0ProbeSpec with its published law and features (16 probes, 5 levels)0.205
1a single nonzero level0.214
2perturbation level σp0.215
3energies of the mean response in 12 zig-zag bands instead of the ProbeSpec features0.692
4energies of the responses instead of their mean (covariance)0.760
5radial bands instead of zig-zag bands0.763
6Fourier power spectrum instead of the DCT0.773
7centered log band energies (centering)0.808
8Euclidean instead of cosine distance0.801
9re-solves centered at their mean, N=16 (ours)0.813

Reading the response as energies in frequency bands carries the attribution (+0.48), its covariance instead of its mean adds +0.07, and centering adds +0.035. For variational solvers, this is what Theorem 1 indicates: the covariance of the response determines the curvature of the prior on T. Centering also removes the scale σp² from the fingerprint. With references at σp and test clouds at another level (N=64, Appendix L):

Test clouds at0.5× σp2× σp
Log band energies without centering0.4660.442
Centered log band energies (ours)0.7860.779

The revision reports this ablation in Table 19 and the change of perturbation level in Appendix L.

AuthorsOpenReview ↗

Author Responses (Part 3)

Regarding what Theorem 1 adds when the draws are known (W1, Q2). We appreciate the question on what Theorem 1 adds. Theorem 1 builds on the first-order covariance of Fessler [5], the implicit function theorem [6] and the uniqueness of the positive semidefinite square root [7]. When the draws are known, the Jacobian can be estimated from their cross-covariance with the re-solves, as in Monte-Carlo SURE [8] and generalized degrees of freedom [9]. What Theorem 1 adds is its identifiability statement and its converse, which hold for every estimator, the Stein estimator [10] included: no method that observes the response, with or without the draws, recovers the curvature outside T, which, as you note, sets the limits of black-box attribution independently of the fingerprint. Its square-root formula also covers draws that are not observed, as in the repeated acquisitions from which the NPS is measured: C then determines J and λH on T. C alone fixes J only up to a rotation, and the symmetric positive semidefinite Jacobian of the predicted measurement removes this ambiguity (Section 4 and Appendix A.1; for A = I it reduces to the linear-smoother argument of Buja et al. [11]). Our fingerprint needs neither the draws nor the unperturbed reconstruction.

Regarding the Stein estimate of the Jacobian (W1, Q2). Thank you for suggesting the estimate from the known draws. Following your suggestion, we computed, on the same clouds, bands and references, the band energies of σp²ĴĴᵀ for the Stein estimate Ĵ = ZΞᵀ/(Nσp²), with the estimator that keeps only pairs of distinct draws, which removes solver noise independent of the draws. Because these estimates can be negative, they are divided by their absolute sum over the bands instead of entering a logarithm (Table 21):

Band energies ofEuclidean, N=8 / 64Per-prior standardized, N=8 / 64
σp²ĴĴᵀ (Stein estimate, known draws)0.220 / 0.5660.297 / 0.649
C (our fingerprint)0.800 / 0.8320.834 / 0.872

Our fingerprint is ahead on each of the 9 inverse problems, at both N. A band energy of C is the trace of a band of the covariance, estimable from N ≪ n re-solves (Proposition 3), whereas the Stein estimate first recovers the Jacobian from N draws, with a variance that increases with the dimension of the measurements. Following your comment, Appendix L now reports this comparison in Table 21.

Regarding when band energies separate two priors (W1, Q2). Thank you for asking when band energies separate two priors. Proposition 4 states that solvers close to a common reference have close fingerprints, an upper bound on their distance. Inspired by your question, we added Proposition 5 (Appendix K, summarized in Section 4), which states when two fingerprints coincide. With e_b = tr(P_b C) and φ the centered log band energies:

  1. φ(C₁) = φ(C₂) if and only if the band energies are proportional, e_b(C₁) = c·e_b(C₂) for all b and some c > 0.
  2. φ is invariant to a common scale and to orthogonal maps that preserve every band.
  3. For a real-analytic family θ ↦ C(θ) on an open connected parameter set, if two members have different fingerprints, the pairs with equal fingerprints have Lebesgue measure zero. For variational solvers, C is real analytic in λH.

Two priors are therefore separated exactly when their band energies are not proportional, and a coincidence within an analytic family is non-generic. At a finite N, Proposition 3 gives the relative error of each estimated band energy, which the study below compares with the distance between the fingerprints of two priors. With exact covariances on 32×32 images (Appendix K):

Exact covariancesDenoisingBlur
Random pairs of priors separated (1,200 per operator)allall
Nine random priors, N=80.9690.999
Nine random priors, N=641.0001.000
Priors that share a common component, N=640.9470.991

At N=64, the nine random priors are attributed as by the maximum-likelihood classifier that knows every exact covariance, and each random pair is separated by more than three times the estimation spread.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all reviewers for their careful evaluations and constructive suggestions, which helped us strengthen and clarify the paper. We are glad that the reviewers found the research perspective novel (UNpq), the idea of identifying learned priors through reconstruction responses interesting and well motivated (4Bru), the identifiability result correct and useful (E2iw), our experimental practice careful and honest (E2iw), and the experiments on external pretrained priors and real MRI measurements particularly valuable (4Bru).

The reviews share one question: whether the response of a solver identifies its learned prior, rather than its configuration or its reconstruction quality, and what this is worth in practice. During the rebuttal, we added 21 new experiments and analyses and one new proposition. The fingerprint of Eq. (8) and the submitted protocols are unchanged; the new results add candidate priors run under several configurations, several images of the solver, and distances in units of the spread of the reference fingerprints (standardization). The standardization uses the reference clouds; its per-prior form (Eq. 16) has three hyperparameters selected on development data, and we report it on held-out test images.

  • Beyond the noise power spectrum and ProbeSpec (E2iw). In the budget-matched comparison of Table 1, our fingerprint attributes at 80.6%, compared with 23.3% for ProbeSpec [1] and 14.9% to 15.5% for ModelDiff [2]. On the clouds of Table 18, the normalized noise power spectrum reaches 0.761 / 0.801, compared with 0.800 / 0.832 for our fingerprint (N=8 / 64; Appendix L).
  • The prior, not the solver configuration (E2iw, UNpq). The same prior is attributed to itself at 0.822 to 0.896 under a changed algorithm, iteration budget, initialization or operator instance, and at 0.731 to 0.851 when the configuration of the test is not among the references (Table 22, random guess 0.111). When the configuration or λ of the test is not among the references (Table 20), the total response decreases to 0.367 to 0.534, while the standardized fingerprint attributes at 0.755 to 0.829. With all 9 candidates matched in total response, the total response attributes at 0.311 and our fingerprint at 0.860 (Table 23).
  • What the compressed fingerprint keeps (E2iw, 4Bru, UNpq). Two responses share a fingerprint if and only if their band energies are proportional (new Proposition 5, Section 4); with exact covariances, all 1,200 random pairs of priors per operator are separated.
  • Accurate external priors (E2iw, 4Bru). On held-out test images, the external priors are attributed at 0.794 to 0.892 with the standardized fingerprint and at 0.828 to 0.935 with its per-prior form (Table 17), and at 0.950 to 1.000 from 8 images of the same solver (Appendix J).
  • Priors outside the candidate set (E2iw, 4Bru). They are separated from the known priors with AUC 0.800 to 0.865 from one image and 0.959 to 1.000 from 8 images of the solver (Appendix J).
  • Practical use (E2iw, 4Bru, UNpq). Two training seeds of the same model are told apart in 81.2% of test clouds, and when two priors reconstruct a test image within 0.5 dB PSNR of each other, our fingerprint attributes the correct prior in 84.4% of the two-way decisions (random guess 50%; per-prior standardization; Appendix L).

Added material, marked in blue in the revised PDF. Main text: practical scenarios (Section 1; all reviewers); the noise power spectrum (Section 2 and Appendix H; E2iw); confidence and priors outside the candidate set (Section 3.4; E2iw, 4Bru); what the fingerprint keeps, Proposition 5 (Section 4 and Appendix K; all reviewers); ProbeSpec in Table 1 (E2iw); solver configurations and candidates matched in total response (Section 5.4; E2iw, UNpq); a specific robustness statement in the contributions, Section 5.4, Figure 9 (Figure 8 in the original submission), Appendix F and the conclusion (E2iw). Appendix C: the adaptation of the baselines and the training data of the CNN (E2iw). D.1: ProbeSpec per inverse problem (E2iw). E.3: the CT pair on 64 images and at the default budget (E2iw, 4Bru). J: standardization, larger test sets, several images, open set, confidence, the cut-off of ρ_k, cost, scenarios and dual use (all reviewers). L, Tables 18 to 23: ProbeSpec, the ablation, NPS and Stein features, the rotated acquisition, configurations and strengths, the matched comparisons, training seeds, equal-quality pairs and unseen priors (all reviewers). Our response to Reviewer E2iw also answers the ethics flag.

We thank all reviewers again and hope that these additions address your concerns and that you will consider revising your scores.

References:

[1] ProbeSpec: Robust Model Fingerprinting via Dynamic Perturbation Response Spectrum

[2] ModelDiff: Testing-Based DNN Similarity Comparison for Model Reuse Detection

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.