acceptodds
Withdrawn from ICLR 2027

Physical Response Attack: Exposing Hidden Response Discrepancies in Neural Operators

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Neural operators have emerged as efficient surrogates for numerical PDE solvers. However, a reliable neural operator should not only achieve low prediction error, but also preserve the response characteristics dictated by the underlying physical equations. Existing robustness evaluations are largely model-centric and focus on how perturbations affect the model output, while overlooking whether the learned operator responds consistently with the true physical system. To address this gap, we propose Physical Response Attack (PRA), a fully black-box framework that exploits the physical response structure of the governing PDE. Using only forward queries to a reference numerical solver, PRA identifies physically informative per- turbation directions and constructs a compact search subspace, within which it searches for perturbations that maximize the response discrepancy between the neural operator and the solver. PRA requires no access to model parameters, gra- dients, or internal architecture; it does not require a differentiable solver and per- forms no backpropagation through either system. We evaluate PRA on the Poisson equation, Darcy flow, and two-dimensional Navier–Stokes equations across four representative neural operator architectures. Experimental results show that PRA effectively reveals failure directions tied to the underlying physical structure; in particular, for time-dependent systems, PRA can induce pronounced deviations in predicted dynamical trajectories. These results demonstrate that the reliability of neural operators should not be determined by prediction accuracy alone; they must also faithfully reproduce the local response and dynamical behavior dictated by the governing equations.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Neural Operators, Adversarial Robustness, Partial Differential Equations, Black-box Attacks

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

The authors introduce Physical Response Attack (PRA), a black box framework that identifies physically informative perturbation directions on which to attack a trained neural PDE solver. The authors demonstrate their method on the Poisson equation, Darcy flow, and Navier-Stokes.

Critical strengths

While there has been some prior work on neural operator attacks, the topic is nonetheless interesting and novel, and I believe the authors should receive credit for exploring this direction. Further, the proposed PRA attack is interesting in contrast to prior works because it does not rely on solver or model internals and is fully black-box.

Critical weaknesses

The authors frame PRA as a neural operator-specific attack, when it appears to not take advantage of the distinct properties of a neural operator and instead would be more aptly applied to general “neural PDE solvers.” Neural operators specifically are resolution agnostic in their input and output; one could imagine an attack that takes advantage of discrepancies across resolutions for neural operators, or perturbations that function differently at different resolutions. As far as I can see, PRA only assumes access to a model and solver at one specific resolution. As such, I believe the operator-tailored framing of the paper may be slightly inaccurate.

Furthermore, I believe it would be important to emphasize and explain the importance (e.g., is it for safety and security reasons?) to explore attacks against neural PDE solvers, or neural operators in general. Given that this paper sits at the intersection of the neural PDE solver and adversarial attack communities (and given that these communities historically do not overlap much), it would be important to justify to both why this line of work should be explored.

While I believe that the numerical methods are sound and thorough, there are still a few other empirical axes that I would recommend the authors explore:

  • How does the performance of PRA compare with other attacks as a function of the perturbation size?
  • It seems from Table 3 that PRA uses more solver calls and frames than some other methods, such as SPSA. Correct me if I misunderstood, but it seems that this cost is variable and depends on factors such as the size of the search space. As such, I would recommend exploring also how the effectiveness of PRA changes as the computational budget for solver/model passes changes.
Questions for authors

It almost seems as though the solver results in Figure 3 contain some artifacts (e.g., lines or ringing), especially for the smoother PRA perturbation (than the other inputs). Is this a realistic response/can you explain why this is happening? If anything, for instance, it seems that FNO has the better output than the solver in Figure 3b.

1 comment

Q1. Resolution-related issue.

Thank you for the valuable comment. While resolution independence is indeed a characteristic of neural operators, this paper investigates the discrepancies between the responses of neural operators and a ground-truth solver under input perturbations. Furthermore, all of our experiments are conducted on neural operators, and the proposed method is not tied to a specific discretization and can be applied at any resolution.

Q2. Is investigating the adversarial robustness of neural PDE solvers practically meaningful, particularly in terms of safety and security?

Thank you for the valuable comment. Investigating the security and robustness of neural PDE solvers has important practical implications, particularly as neural operators are increasingly adopted as computationally efficient surrogates in safety-critical applications. For example, Roy et al. (2026) investigated adversarial vulnerabilities in neural-operator-based digital twins for nuclear thermal-hydraulic systems, demonstrating that physically plausible input perturbations can substantially increase prediction errors. This highlights that adversarial vulnerabilities in neural operators are not merely theoretical concerns but can pose real risks to the reliability of engineering systems.

Q3. Artifacts in the solver results in Figure 3.

Thank you for the valuable comment. We clarify that our Navier–Stokes trajectories were generated by directly solving the governing equations on a 64×6464 \times 64 grid, rather than by downsampling trajectories from a higher-resolution simulation.

At ν=10−4\nu = 10^{-4}, the weak viscous dissipation allows increasingly fine-scale vorticity structures to develop, which may not be adequately resolved by the 64×6464 \times 64 discretization. Although the 2/3 dealiasing rule mitigates aliasing errors, it does not eliminate truncation errors associated with unresolved spatial scales. Consequently, the direct 64×6464 \times 64 solver may exhibit high-wavenumber spectral accumulation and ringing artifacts, which are absent or substantially reduced at higher resolutions (We will provide further explanation in the appendix)

In contrast, FNO learns an approximate solution operator through spectral convolutions in the Fourier domain rather than explicitly reproducing the numerical integration procedure. Its use of truncated Fourier modes introduces a spectral bias toward low-frequency structures, which suppresses poorly resolved high-frequency components while preserving the dominant large-scale flow patterns. As a result, FNO may produce smoother predictions than the numerical solver.

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

This paper proposes Physical Response Attack (PRA), a black-box adversarial framework for evaluating the robustness of neural operators by identifying input perturbations that maximize the discrepancy between neural operator predictions and reference PDE solutions.

The main idea is to use the physical response of the PDE solver to construct a low-dimensional, Lyapunov-informed perturbation subspace, followed by response-based optimization without requiring gradients or differentiable solvers.

Numerical experiments on Poisson, Darcy flow, and Navier–Stokes equations demonstrate that PRA can effectively expose neural operator prediction errors, particularly for time-dependent PDEs.

Critical strengths
  1. Incorporating physical response information into black-box adversarial attacks on neural operators is an interesting and potentially useful idea.

  2. The proposed approach does not require differentiable PDE solvers or access to neural operator gradients, making it applicable to a broader range of settings.

  3. Experiments across several PDEs demonstrate the potential of PRA to identify neural operator prediction failures, particularly for time-dependent problems.

Critical weaknesses
  1. The forward-propagated QR basis does not necessarily capture the most sensitive initial perturbation directions, particularly for non-normal dynamics.

  2. PRA substantially underperforms existing attacks on Poisson, raising concerns about its effectiveness across different PDEs.

  3. PRA requires substantially more solver evaluations than SPSA in some settings. More rigorous subspace analysis and matched-budget comparisons are needed to support its claimed advantages.

Questions for authors
  1. How does the QR-based subspace capture sensitive perturbation directions, especially for non-normal dynamics?
  2. How does the proposed subspace compare with singular-vector-based or random subspaces?
  3. Why does PRA underperform existing attacks on Poisson?
  4. Can the authors provide matched-budget comparisons with SPSA to justify the claimed query efficiency?
2 comments

Q1. How does the QR-based subspace capture sensitive perturbation directions, especially for non-normal dynamics?

Thank you for the valuable comment. The sensitivity information arises from solver-driven perturbation evolution rather than from QR itself. Specifically, repeated Jacobian-vector products propagate candidate directions through the physical dynamics, while QR orthogonalization maintains numerical stability and prevents directional collapse. For non-normal dynamics, where transient amplification may occur even under asymptotic stability, this process identifies directions associated with finite-time perturbation growth rather than relying solely on eigenvalue magnitudes. The subsequent SVD of the response matrix further refines the attack direction based on the observed physical responses.

Q2. How does the proposed subspace compare with singular-vector-based or random subspaces?

Thank you for the valuable comment. We conducted additional experiments across four neural operators at ν=10−3\nu = 10^{-3} , all under a 100% perturbation budget:

ν\nuModelOurs (QR-based)Singular-vectorRandom subsbase
10−310^{-3}FNO23.8523.561.35
10−310^{-3}UNO24.2224.752.56
10−310^{-3}Transolver3.963.382.43
10−310^{-3}UPT13.0818.231.95

Both QR-based and singular-vector-based subspaces generally outperform random subspaces, demonstrating the importance of identifying sensitive perturbation directions. However, neither consistently dominates the other across all configurations. We acknowledge that QR-based construction cannot universally replace singular-vector-based methods; instead, we view the two as complementary. These findings suggest treating subspace construction as a modular component of PRA rather than restricting it to QR-based construction, allowing different methods to be selected based on the underlying PDE dynamics and neural operator architecture. We will revise the manuscript accordingly to clarify this limitation.

Q3. Why does PRA underperform existing attacks on Poisson?

Thank you for the valuable comment. As acknowledged in the paper, PRA achieves lower total error amplification than some baselines on Poisson. However, conventional attacks achieve this higher total error predominantly by inducing errors outside the dominant solution modes, whereas PRA more strongly perturbs the dominant spatial structures. To quantify this distinction, we introduce the structural share S\mathcal{S}, defined as the fraction of attack-induced error concentrated in the dominant modes, which are identified solely from the reference solution.

At the 100% perturbation budget, Random, SPSA, and DE amplify the total error by 200200–500×500\times on FNO and UNO, yet their structural share remains extremely low (S≈0.0011\mathcal{S} \approx 0.0011). In contrast, PRA achieves S=0.55\mathcal{S} = 0.55–0.880.88, with errors in the dominant modes 3.53.5–4.1×4.1\times larger than those of the strongest baseline. Similar trends are observed on Transolver and the other architectures. These results show that PRA primarily disrupts the dominant spatial structures of the solution rather than merely amplifying high-frequency noise.

Q4. Can the authors provide matched-budget comparisons with SPSA to justify the claimed query efficiency?

Thank you for the valuable comment. PRA and SPSA differ fundamentally in their use of the numerical PDE solver. While SPSA directly optimizes perturbations through target-model queries, PRA interacts with the physical solver to construct sensitive perturbation directions, which introduces additional solver evaluations. Importantly, these solver-based computations can be performed offline without accessing the target neural operator, separating solver evaluations from online model queries. We acknowledge that these additional evaluations contribute to the total computational cost. However, since solver-guided subspace construction is an essential component of PRA but is absent in SPSA, directly matching the computational budgets of these fundamentally different procedures is not straightforward. We will clarify this distinction and the scope of our query-efficiency claims in the revised manuscript and, where feasible, provide an additional computational cost analysis.

Thank you for the detailed response and additional experiments. I appreciate the authors' efforts to clarify the method and provide further comparisons, particularly the structural analysis.

However, I still have two concerns:

  1. Attack subspace: The additional results show that low-frequency Fourier subspaces can match or outperform PRA, leaving the specific advantage of the solver-guided construction unclear. The connection between the QR-based subspace and the most relevant attack directions also remains insufficiently justified.

  2. Computational efficiency: While the methods differ in their computational procedures, a comparison accounting for both solver evaluations and model queries is still needed to establish PRA's practical efficiency.

Overall, the response addresses some of my concerns, but I remain unconvinced of the method's distinctive advantages and therefore maintain my current rating.

Reviewer_9X9dOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Physical Response Attack (PRA), a black-box framework for evaluating neural operator robustness. It uses forward numerical-solver queries to construct a low-dimensional perturbation subspace, then searches for discrepancies between model and physical responses. Experiments cover three PDE families and four neural operator architectures, with strong reported results on Navier–Stokes and Darcy flow.

Critical strengths

The solver-guided search requires neither model gradients nor a differentiable solver, making it applicable to black-box neural surrogates.

Experiments cover multiple PDEs and architectures, and the controlled random-subspace ablation provides useful evidence that the search space matters.

Critical weaknesses

The specific benefit of physics guidance remains unclear without a low-frequency subspace baseline. The claimed structural advantage also lacks quantitative validation, particularly on Poisson, where DE consistently achieves larger error amplification.

Attack comparisons use unequal query budgets and perturbation constraints. For Darcy, DE’s maximum perturbation norm is below even PRA’s smallest budget, weakening conclusions about comparative attack strength.

Questions for authors

Does the solver-derived subspace outperform a dimension-matched low-frequency Fourier subspace under the same optimization procedure, and how is its claimed structural advantage quantified?

Do the reported gains persist when attacks use the same admissible perturbations and comparable computational budgets, including subspace construction?

1 comment

Q1. Does the solver-derived subspace outperform a dimension-matched low-frequency Fourier subspace under the same optimization procedure?

Thank you for the valuable comment. We compared PRA's solver-derived subspace against dimension-matched low-frequency Fourier and random subspaces under the same optimization procedure. The results demonstrate that PRA consistently outperforms random subspaces and achieves competitive performance against low-frequency Fourier subspaces, with stronger amplification in several configurations. While low-frequency structure contributes to attack effectiveness, PRA explicitly exploits solver-derived physical sensitivity rather than relying solely on predefined frequency components, enabling a more comprehensive characterization of physically sensitive perturbation directions.

Q2. How is its claimed structural advantage quantified?

Thank you for the valuable comment. We agree that our original analysis relied primarily on visual comparisons. To provide quantitative evidence, we introduce a structural metric for the Poisson problem. Since the solution is dominated by a few smooth modes (with three modes capturing over 95% of its energy), we define the structural share S\mathcal{S} as the fraction of attack-induced error concentrated in these dominant modes, which are identified solely from the reference solution.

At the 100% perturbation budget, Random, SPSA, and DE amplify the total error by 200200–500×500\times on FNO and UNO, yet their structural share remains extremely low (S≈0.0011\mathcal{S} \approx 0.0011), indicating that fine-scale noise dominates the errors. In contrast, PRA achieves S=0.55\mathcal{S} = 0.55–0.880.88, with errors in the dominant modes 3.53.5–4.1×4.1\times larger than those of the strongest baseline. Similar behavior is also observed on Transolver and the other evaluated architectures. These results demonstrate that, while conventional attacks primarily introduce high-frequency noise, PRA more effectively disrupts the dominant spatial structures of the predicted solutions.

Q3. Do the reported gains persist when attacks use the same admissible perturbations and comparable computational budgets, including subspace construction?

Thank you for the valuable comment. Unlike conventional attacks, PRA explicitly interacts with the physical solver to construct a physics-guided subspace, which is a core component of our method. Baselines such as SPSA and DE do not require solver-based subspace construction or joint interaction with the solver. Therefore, enforcing identical computational budgets, including subspace construction, would not reflect these fundamental methodological differences. We acknowledge this limitation and clarify that our comparisons evaluate attack effectiveness rather than strict computational equivalence.

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

Related papers

Discussion (0)

Sign in to comment.