acceptodds
Under review as a conference paper at ICLR 2027

Can Hybrid Weight-Space Stochasticity Advance Adversarial Robustness in Multimodal Fusion?

Abstract

We introduce Hybrid Weight-Space Convolutional Stochasticity (HyWeCS), a unified randomized defense for multimodal fusion that closes the long-standing gap between provable ℓ2 certification and empirical robustness against strong ℓ∞ attacks, while preserving generalization on natural- and medical data. Specifically, we formulate two complementary mechanisms at disjoint weight locations within every filter: Orthogonally Retracted Random-Projection Weights (ORaW) and Stochastic Hybrid-Space Attention-Noise Weights (SHANW), jointly exploiting projection diversity and self-modulated attention noise during training and inference. Reciprocal intermediate cross fusion reuses terminal stochastic weights without a separate feature-space stochastic module, thereby saving computational cost, with parameters and GFLOPs reduced by up to ≈ 60.5% and ≈ 91.1%, respectively. Shared Lipschitz calibration establishes a uniform 2-Lipschitz score bound for formal certification. Across eight multimodal benchmarks, HyWeCS improves certified and empirical robustness by up to ≈ 12.4% and ≈ 12.8%, respectively, over the strongest competing defenses, with improvements persisting under adaptive attacks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Adversarial Defense, Multimodal Fusion, Medical data

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_1ygQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes HyWeCS, a method for improving adversarial robustness in multimodal fusion by combining two stochastic weight-generation mechanisms, ORaW, which integrates random weight perturbations with Layer-wise Orthogonal Training, and SHANW, which uses Euclidean and hyperbolic attention. The method also reuses stochastic weights for cross-modal interactions. Through experiments on eight multimodal benchmarks, the authors show improvements in certified and empirical robustness. The paper also reports reductions in parameter count and computational cost. The authors also provide theoretical robustness guarantees based on Lipschitz continuity and Randomized Smoothing.

Critical strengths
  1. The proposed method improves certified and empirical robustness across eight multimodal benchmarks.
  2. The method reduces parameters and computational cost compared with existing hybrid stochastic defenses.
Critical weaknesses
  1. The proposed method mainly combines various existing techniques. Although it seems empirically effective, the novelty and functional significance of this particular combination remain unclear.
  2. Whether the trained networks actually satisfy the assumptions underlying the theoretical analysis, including the required Lipschitz bounds, is unclear.
  3. Several design choices lack sufficient justification and validation. In particular, the contributions of self-modulated noise, hyperbolic attention, and reciprocal cross-modal interaction to robustness improvements remain unclear.
Questions for authors
  1. How are the Lipschitz bounds in Eq. (3) and Theorem 1 enforced or verified during training and inference? In particular, how is the input dependence of SHANW weights accounted for, and how is the non-expansiveness of reciprocal cross fusion ensured? Could the authors provide numerical evaluations or other evidence demonstrating that the implemented networks satisfy these assumptions?
  2. What motivates the self-modulated noise in Eq. (17), and what advantages does it offer over conventional Gaussian noise?
  3. Why is hyperbolic attention necessary in SHANW? What theoretical or empirical advantages does it offer over Euclidean attention alone?
  4. Does the reciprocal cross-modal interaction in Eq. (6) improve robustness beyond reducing computational cost? How does it compare with conventional multimodal fusion using the same ORaW and SHANW mechanisms?
3 comments

Dear Reviewer 1ygQ,

We would like to sincerely thank you once again for the very careful reading of our manuscript and for your thoughtful, detailed, and constructive evaluation. As discussion closes, we respectfully invite feedback on the evidence below. References concern the revised manuscript; gains are percentage points. Tables 12–14 use UCF101.

  1. W1 — Prior-defense distinctions. We explain the complementary limitations of CTRW, HyCAS, HySCAN, and LOT (Apps. A.2–A.3, A.5, pp. 16–19; Fig. 1; Table 7, p. 18).

  2. W1 — Within-filter construction. We detail r/(n−r) allocation, ORaW’s projection resampling and LOT retraction, SHANW’s feature-conditioned attention noise, and reciprocal terminal-weight reuse (Secs. 2.2–2.3.3, pp. 4–7; App. B.2, p. 20; Sec. 3.1, p. 9, ll. 474–485).

  3. W1 — Component and replacement evidence. Four configurations test individual/joint contributions; a separate ablation isolates random projection and LOT. ORaW is compared with RPF, CTRW, and SN-RPF, and SHANW with RANI-W and RWAN. On RAVDESS/UCF101, joint gains over either component alone are 7.0–12.4 certified / 8.4–13.3 EOTPGD-20 points. Replacement gains are 3.1–6.4 for ORaW and 1.3–4.7 points for SHANW (Tables 10–11; Apps. J.1–J.2, pp. 49–51).

  4. W2/Q1 — SHANW’s input dependence. Lemma 6 bounds the full attention stack; Proposition 2 retains input-dependent attention and geometry-selection coefficients for a fixed draw (Apps. D.2–D.5; Lemma 6 and Proposition 2, pp. 28–30, Eqs. (60)–(63)).

  5. W2/Q1 — Explicit layer calibration. Corollary 3 derives S + M K_SHANW, including the input-dependent term, and the coefficient-to-convolution conversion. Calibration gives Lip ≤ τ ≤ 1 on the stated input domain (Sec. 2.3, Eqs. (3)–(4); App. D.5, Corollary 3, pp. 31–32, Eqs. (64)–(69)).

  6. W2/Q1 — Residual–fusion sensitivity. Proposition 3 retains the source branch’s weight-conditioning pathway and recipient-feature variation. Its 2 × 2 sensitivity matrix bounds the complete stage without assuming automatic non-expansiveness (App. E.1, pp. 33–35, Eqs. (70)–(74)).

  7. W2/Q1 — End-to-end guarantee. Theorem 1 composes complete stage bounds and applies global calibration. Under the stated assumptions and uniform representation bounds, it establishes a 2-Lipschitz prediction mapping and expectation (Sec. 2.3.3; Sec. 2.4, Theorem 1, p. 7, Eq. (15); App. E.2, pp. 35–36).

  8. W2/Q1 — Training, inference, and certification. We detail layer/global calibration, resampling, attack-phase and parameter-update draws, and EOT gradient averaging (App. D.4, p. 27, ll. 1439–1445; App. F, pp. 38–39; App. H.2, pp. 43–45).

  9. W3/Q2 — Noise construction and analysis. Eqs. (6)–(7) distinguish noise construction from its integration into SHANW weights. Lemma 5 gives mean π, variance 3π², and second moment 4π²; Theorem 3 bounds the recursive weight update (Sec. 2.3.2; App. D.4, pp. 26–27; App. D.5, Lemma 5/Theorem 3, p. 28).

  10. W3/Q2 — Isolated Gaussian comparison. HyWeCS-G changes only the noise rule. Self-modulation gains 0.5 / 1.5 / 1.6 / 2.9 points in clean / PGD-20 / EOTPGD-20 / RS-certified accuracy (Table 12; App. J.3, pp. 51–52).

  11. W3/Q3 — Geometry and motivation. We explain Lorentz-geodesic relations to learned prototypes, complementary Euclidean local/channel responses, and their alignment with SHANW weights. We cite Nickel and Kiela (2018) and Park et al. (2023) (App. B.3, pp. 20–21; Apps. D.2–D.4, pp. 24–27).

  12. W3/Q3 — Geometry ablation. Hybrid attention gains 0.9 / 1.2 / 1.6 / 1.7 points over Euclidean-only and 0.4 / 0.5 / 0.8 / 0.9 points over hyperbolic-only in clean / PGD-20 / EOTPGD-20 / RS-certified accuracy (Table 13; App. J.4, p. 52).

  13. W3/Q4 — Reciprocal interaction. We explain fusion vulnerability and how each branch’s terminal stochastic weights transform the paired modality’s representation (Sec. 2.1; Sec. 2.3.3, Eq. (9), p. 6; App. A.1, p. 16; App. J.5, pp. 52–54).

  14. W3/Q4 — Controlled fusion comparisons. With ORaW–SHANW retained, reciprocal fusion gains 4.4–5.9 / 4.6–5.3 / 4.1–4.6 points over conventional late/intermediate fusion and 1.6 / 1.8 / 1.9 points over same-branch reuse in PGD-20 / EOTPGD-20 / RS-certified accuracy. Clean-accuracy reductions are 1.7–2.4 and 0.8 points, respectively (Table 14; App. J.5, pp. 53–54).

  15. Multimodal and efficiency evidence. We also report modality-specific attacks, robustness–complexity comparisons, and training, inference, memory, and adversarial-evaluation costs (Table 8; Figs. 7–9; Table 9; Apps. I.2–I.3 and I.5, pp. 45–48).

We would be very grateful if you could identify any remaining concerns and the specific argument, assumption, derivation, comparison, or result requiring clarification.

Thank you for your time and consideration.

With kind regards,
The Authors

I sincerely appreciate the authors' efforts to address my previous concerns. In particular, I appreciate the detailed additional proofs about the Lipschitz guarantees, including the input dependence of SHANW and the effects of cross-modal fusion, while, to be honest, I have not been able to examine all the proofs in detail within this limited review period.

My remaining concern is whether the provided theoretical guarantee is fully applicable to the actual implementation and whether the overall contribution is sufficiently significant. Therefore, while I do not rule out possibility of raising my rating, I prefer to maintain my original rating at this stage.

Dear Reviewer 1ygQ,

Thank you very much for your thoughtful follow-up and for acknowledging the additional theoretical analysis. We understand the limited time available for reviewing the proofs and are grateful for your continued consideration. We would like to clarify the two remaining concerns.

1. Applicability of the theoretical guarantee

The revised analysis follows the computations described in the method. At the layer level, the calibration in Equation (3) accounts for how SHANW weights change with the input in the convolution defined by Equation (4). At the fusion level, it accounts for both the branch that determines the reused weights and the paired representation to which those weights are applied. The complete residual and fusion contributions, including any increase in sensitivity during fusion, are included when calibrating the final prediction mapping (App. D.5, Corollary 3, pp. 31–32, Eqs. (65)–(69); App. E.1, Proposition 3, pp. 33–35, Eqs. (70)–(74); App. E.2, pp. 35–36).

The corresponding execution procedure is specified in App. F, p. 38, ll. 2020–2051, which describes the calibrated evaluation used by Algorithms 1–3. It explains how bounds are updated for the current learned model, how calibration is applied to each stochastic realization, and how the complete forward computation enters the final calibration. Training, inference, and certification settings are detailed in App. H.2, pp. 43–45.

The guarantee assumes that internal draws follow an input-independent distribution, that expected predictions are well defined, and that finite bounds on inputs and intermediate representations hold throughout the stated evaluation domain, including inputs perturbed by Gaussian noise. The derivations establish the guarantee under these conditions; confirming that the evaluated code satisfies them requires separate verification (Sec. 2.4, Theorem 1, p. 7; App. F, p. 38, ll. 2046–2051).

We also distinguish the expected-logit certificate supported by Theorem 1 from the probability-based randomized-smoothing certificate supported by the Gaussian-shift argument and finite-sample probability estimation (Corollary 1, Eqs. (16)–(17); App. B.4, Lemma 3; App. H.2, p. 44, ll. 2349–2355).

2. Significance of the contribution

Distinct methodological construction. HyWeCS coordinates ORaW and SHANW at complementary, non-overlapping weight locations within each filter and reuses the terminal stochastic weights for reciprocal cross-modal interaction. This combines projection-driven diversity with input-conditioned Euclidean and hyperbolic attention noise. The comparison with CTRW, HyCAS, HySCAN, and LOT explains the distinctions motivating this organization (Sec. 2.3; App. A.5, Table 7).

Evidence for the coordinated design. On RAVDESS and UCF101, combining the two mechanisms improves RS-certified accuracy by 7.0 to 12.4 percentage points and EOTPGD-20 accuracy by 8.4 to 13.3 points over either mechanism alone. Comparisons with stochastic replacements additionally show gains of 3.1 to 6.4 points for ORaW and 1.3 to 4.7 points for SHANW (Apps. J.1–J.2, Tables 10–11).

Additional contribution of reciprocal fusion. With the encoders and both stochastic mechanisms retained, reciprocal fusion improves UCF101 accuracy over conventional late and intermediate fusion by 4.4 to 5.9 points under PGD-20, 4.6 to 5.3 points under EOTPGD-20, and 4.1 to 4.6 points for RS certification. Compared with same-branch weight reuse, the corresponding gains are 1.6, 1.8, and 1.9 points. We also report the accompanying clean-accuracy reductions of 1.7 to 2.4 points relative to conventional fusion and 0.8 points relative to same-branch reuse (App. J.5, Table 14).

Practical benefit. Across four natural-domain benchmarks, HyWeCS reduces training time by 29.6 to 50.5 percent, inference latency by 30.2 to 56.3 percent, memory consumption by 37.3 to 48.9 percent, and adversarial-evaluation time by 38.6 to 56.0 percent, relative to HyCAS-M and HySCAN-M (App. I.5, Table 9).

We respectfully believe these controlled comparisons support the significance of the coordinated design and its practical benefits, while the theoretical result should be assessed within its stated scope.

As the author discussion approaches its close, we would be grateful for any specific indication of an assumption, computational step, or comparison that would benefit from further clarification. We would be glad to respond promptly and carefully. We respectfully ask that these clarifications and the cited evidence be considered in your final assessment.

Thank you for your time and consideration.

With kind regards,
The Authors

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

This paper proposes HyWeCS, a hybrid weight-space stochastic defense for multimodal fusion that combines orthogonally retracted random-projection weights (ORaW) with stochastic hybrid-space attention-noise weights (SHANW) within each convolutional filter. By integrating Lipschitz calibration and reciprocal intermediate cross-fusion, the method aims to improve certified (ℓ2\ell_2) robustness and empirical (ℓ∞\ell_\infty) robustness while reducing computational overhead. Experiments on eight natural and medical multimodal benchmarks demonstrate improvements over existing defenses, including under adaptive attacks.

Critical strengths

1 HyWeCS combines two complementary weight-space stochasticity mechanisms within each convolutional filter, offering a coherent approach to improving both certified and empirical adversarial robustness.

2 The paper provides a Lipschitz-based analysis and randomized-smoothing certificates, connecting the proposed architecture to formal robustness guarantees beyond empirical attack evaluation.

3 Experiments cover eight multimodal benchmarks across natural and medical domains, with evaluations under adaptive attacks, ablation studies, and comparisons of parameter counts and computational costs. The reported results suggest a favorable robustness–efficiency trade-off.

Critical weaknesses

1 The method introduces several specialized components and acronyms, including ORaW, SHANW, LOT, and reciprocal intermediate cross-fusion, within a relatively compact main text. Some mathematical formulations are difficult to interpret without repeatedly consulting the figures and appendices, while the distinction between architectural design, theoretical assumptions, and experimentally demonstrated benefits could be made more explicit. A clearer presentation of the central algorithm, a more intuitive explanation of the hybrid weight construction, and a sharper separation between formal guarantees and empirical observations would improve readability.

2 The reported reductions in parameter count and GFLOPs are promising, but these metrics alone do not fully characterize the computational cost of stochastic weight generation, recursive attention-noise computation, and repeated evaluations required by adaptive attacks. A more comprehensive comparison of training cost, inference latency, memory consumption, and attack-time evaluation overhead would better substantiate the claim that HyWeCS offers a practically superior robustness–efficiency trade-off.

3 The paper attributes both improved multimodal robustness and reduced computational overhead to reusing terminal stochastic weights for cross-modal interactions. However, the reported component ablations primarily focus on ORaW, SHANW, and the components of ORaW. More direct comparisons isolating the effect of reciprocal cross-fusion, while keeping the remaining architecture and training protocol fixed, would strengthen the causal evidence for this design choice.

Questions for authors

1 See weakness. I'll adjust my rating according to the authors' rebuttal.

1 comment

Dear Reviewer NLTx,

We would like to sincerely thank you once again for the very careful reading of our manuscript and for your thoughtful, detailed, and constructive evaluation. We are grateful for your encouraging assessment and helpful suggestions. As the discussion draws to a close, we respectfully invite your feedback on the revisions below. References concern the revised manuscript; accuracy differences are percentage points.

  1. W1 — Clearer method progression. We reorganized the presentation from within-filter construction to ORaW and SHANW, then their integration into the multimodal learner. The component descriptions explain LOT’s role and the progressive attention-noise formulation (Secs. 2.3–2.3.2, pp. 4–6).

  2. W1 — Explicit weight allocation. We specify the non-overlapping r/(n−r) allocation, explain its rationale relative to whole-filter and overlapping assignments, and distinguish the calibrated hybrid weights in Eq. (3) from the stochastic convolution in Eq. (4) (Sec. 2.3, pp. 4–5, ll. 184–230).

  3. W1 — Dedicated learning framework. The new subsection groups residual integration, reciprocal cross fusion, deterministic–stochastic mappings, losses, and the training objective; references to the training and adversarial-optimization procedures accompany that objective (Sec. 2.3.3, pp. 6–7, ll. 294–345; Algorithms 1–3).

  4. W1 — Guarantees and empirical evidence. We group the formal results with their assumptions, clarify Theorem 1’s role in expected-logit certification, and distinguish certified ℓ₂ guarantees from the adversarial ℓ∞ expected-loss bound. Attack-based results are presented separately (Sec. 2.4, Theorem 1, Corollary 1, Theorem 2, pp. 7–8; Sec. 3, pp. 8–10).

  5. W2 — Matched computational setting. We compare HyWeCS, HyCAS-M, and HySCAN-M on an A100 80 GB GPU with one ResNet18 encoder per modality and matched processing, training, and evaluation protocols. Costs are averaged across RAVDESS, UCF101, KS, and VGGSound (App. H.1, p. 41; App. H.2, p. 43; App. I.5, Table 9, p. 48).

  6. W2 — Complete cost accounting. Training time includes adversarial training; latency covers a complete multimodal input; adversarial-evaluation time includes both example generation and model evaluation (App. G, pp. 40–41; App. H.2, pp. 44–45; App. I.5, p. 48, ll. 2560–2563).

  7. W2 — Measured practical savings. HyWeCS reports 89.5 h training, 15.7 ms/input inference, 107.03 MB memory, and 12.1 h adversarial evaluation. Relative to HyCAS-M and HySCAN-M, reductions are 29.6–50.5%, 30.2–56.3%, 37.3–48.9%, and 38.6–56.0%, respectively (App. I.5, Table 9, p. 48).

  8. W2 — Architectural explanation and broader efficiency evidence. The within-filter construction avoids a separate feature-space stochastic pathway; terminal-weight reuse avoids an additional stochastic fusion-weight generator. We also report parameter/GFLOP comparisons (Secs. 2.3 and 2.3.3; App. H.2, p. 43, ll. 2308–2315; App. I.3, Figs. 7–9, pp. 46–47; App. I.5, p. 48; App. K, p. 57, ll. 3050–3052).

  9. W3 — Why fusion needs protection. We explain how perturbed modality evidence can affect the joint prediction through fusion. Eq. (9) applies each branch’s terminal stochastic weights to the paired modality, extending ORaW–SHANW transformations to cross-modal exchange (Sec. 2.1, p. 3; Secs. 2.3.1–2.3.3, pp. 5–7; App. A.1, p. 16; App. H.2, pp. 43–45; App. J.5, pp. 52–53).

  10. W3 — Controlled fusion comparisons. On UCF101, we retain the encoders, ORaW–SHANW construction, and training/evaluation protocols while changing fusion: late fusion (LF), concatenation-based intermediate fusion with learned projections (CF), same-branch terminal-weight reuse (SR), and reciprocal fusion. The SR comparison isolates cross-modal application beyond weight reuse (App. J.5, Table 14, p. 53, ll. 2809–2825, 2841–2856).

  11. W3 — Empirical and certified gains. Reciprocal fusion improves PGD-20 / EOTPGD-20 / RS-certified accuracy by 4.4–5.9 / 4.6–5.3 / 4.1–4.6 points over LF/CF and 1.6 / 1.8 / 1.9 points over SR. Both attacks use ε = 8/255; RS certification uses (r, σ) = (0.75, 0.25) (App. J.5, Table 14, pp. 53–54, ll. 2857–2867).

  12. W3 — Clean-accuracy trade-off. We also report 1.7–2.4 points lower clean accuracy than LF/CF and 0.8 points lower than SR, alongside the robustness improvements (App. J.5, pp. 53–54, ll. 2868–2874).

If your schedule permits, we would be very grateful to hear whether these revisions address your concerns and which specific explanation, measurement, or comparison still requires clarification. Even a brief indication would help us provide a focused response before the discussion concludes. We would be glad to respond promptly and carefully.

Thank you for your time and consideration.

With kind regards,
The Authors

Reviewer_mmZ8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HyWeCS, which combines two existing methods (ORaW and SHANW) for robust multimodal fusion. Experiments show that the combination of these methods performs well across eight robustness benchmarks.

Critical strengths

This paper conducts a broad experiments over 8 multimodal benchmarks. The benchmarks covers both certified and empirical adversarial robustness under different attacks. The ablation study also supports that the proposed method individually contribute to robustness improvements.

Critical weaknesses

Novelty

The proposed method is a combination of existing methods (random weight perturbation, LOT and attention-based noise injection). It does not have a convincing explanation of why this combination should improve adversarial robustness. Further more, stochastic weight modelling is closely related to Bayesian neural networks and stochastic model averaging, but the comparison between the proposed method and them are missing. Hence, we do not know whether the gains comes from the proposed mechanisms or generic weight randomization.

Theoretical Analysis is not related to the robustness improvements

The theoretical results are about orthogonality, boundedness, and conditional Lipschitz guarantees. However, none of these results explains why ORaW, SHANW, or their combination should yield adversarial robustness.

Questions for authors

(1) What is the difference between HyWeCS and Bayesian methods (such as Bayesian weight sampling or simpler stochastic weight ensembles)? Can the authors provide controlled comparisons?

(2) Table 7 just show that ORaW and SHANW improve the robustness individually. Can authors demonstrate improvements attributable to their combination?

1 comment

Dear Reviewer mmZ8,

We would like to sincerely thank you once again for the very careful reading of our manuscript and for your thoughtful, detailed, and constructive evaluation. We are grateful for the time and care you have devoted to assessing our work and for the opportunity to strengthen it through your questions.

Following your suggestions, we have clarified the novelty and complementary roles of ORaW and SHANW, added the requested Bayesian and stochastic-averaging comparisons, and expanded the component and replacement ablations. We sincerely hope these additions make the contribution and its supporting evidence clearer.

As the discussion period is approaching its close, may we respectfully ask whether you might have an opportunity to share your thoughts on our responses? If your schedule permits, we would be very grateful to learn whether the additional explanations address your concerns or whether any point still requires clarification. A brief comment would help us focus the remaining discussion on the questions most important to your assessment.

We would be glad to respond promptly and carefully to any further questions. Thank you very much for considering this follow-up and for your time and kind consideration throughout the review process.

With kind regards,
The Authors

Comments on the paper

1 comment

Dear Reviewer mmZ8,

Thank you for considering our rebuttal and for acknowledging that some concerns have been addressed.

Could you please identify which concerns remain unresolved and why the corresponding explanations or additional evidence are insufficient? For clarity, our detailed responses and revised manuscript provide the following distinct analyses and results.

  1. Prior-defense distinctions and novelty. We separately discuss the limitations of CTRW, HyCAS, HySCAN, and LOT and explain the complementary capabilities coordinated by HyWeCS (Apps. A.2–A.5, pp. 16–19; Fig. 1; Table 7, p. 18).

  2. Construction and mechanistic explanation. We explain the disjoint allocation of ORaW and SHANW to r and n−r locations within each filter, their projection-driven and feature-conditioned roles, and terminal-weight reuse for reciprocal fusion. The main paper also contains a mechanism-based interpretation of the results (Secs. 2.2–2.3.3, pp. 4–7; Sec. 3.1, p. 9, ll. 474–485).

  3. Requested Bayesian and averaging comparisons. We compare with Adv-BNN-M, IG-BNN-M, RSE-M, and SWAAT-M. HyWeCS improves clean, PGD-20, EOTPGD-20, and RS-certified accuracy by 1.2–2.0, 6.0–11.9, 4.6–12.5, and 11.5–20.0 percentage points, respectively (Table 5, p. 10; App. I.4, p. 47).

  4. Evaluation controls. We document the common data, encoders, training and evaluation settings, applicable prediction-sampling budgets, EOT-aware attacks, and common Gaussian randomized-smoothing certification procedure (App. H.2, pp. 43–45; App. I.4, p. 47).

  5. Independent component contributions. The component ablation includes neither mechanism, ORaW alone, SHANW alone, and both mechanisms. Disabled components are replaced by trainable deterministic weights, preserving filter dimensions and trainable capacity. Both mechanisms improve robustness individually (Table 10(a); App. J.1, p. 49, ll. 2592–2616).

  6. Additional benefit of the combination. The joint configuration explicitly outperforms either component alone by 7.0–12.4 points in RS-certified accuracy and 8.4–13.3 points under EOTPGD-20, across RAVDESS and UCF101 (Table 10(a); App. J.1, p. 49, ll. 2617–2624).

  7. Separate roles of random projection and LOT. An additional ablation enables projection resampling and LOT separately and jointly while holding SHANW fixed; their combination performs best among these configurations (Table 10(b); App. J.1, p. 49, ll. 2625–2636; App. B.2, p. 20).

  8. Five controlled stochastic replacements. With the complementary component held fixed, ORaW exceeds RPF, CTRW, and SN-RPF by 3.1–6.4 points, while SHANW exceeds RANI-W and RWAN by 1.3–4.7 points across the reported certified and EOTPGD measures. These comparisons examine the specific stochastic mechanisms under common backbone, allocation, fusion, calibration, and evaluation settings (Table 11; App. J.2, pp. 49–51).

  9. Analytical connection to robustness. The revised analysis accounts for input-dependent weights and complete residual–fusion sensitivity. Under the stated assumptions, it provides the calibrated Lipschitz guarantee, expected-logit certificate, and adversarial-loss bound; the probability-based randomized-smoothing certificate is stated separately (Sec. 2.4, Theorem 1, Corollary 1, and Theorem 2, pp. 7–8; App. B.4; Apps. D.5, E.1–E.4).

  10. Multimodal and computational evidence. Our responses also cite modality-specific attacks and robustness–efficiency comparisons, including measured training, inference, memory, and adversarial-evaluation costs (Table 8; Figs. 7–9; Table 9; Apps. I.2–I.3 and I.5, pp. 45–48).

The requested baseline comparisons and joint-component results are therefore explicitly reported. If these or the accompanying explanations remain insufficient, we would be grateful for a specific indication of the unresolved point and the relevant evidence or assumption. This would allow us to provide a focused clarification and make the remaining scientific disagreement clear.

Thank you for your time and consideration.

With kind regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.