acceptodds
Under review as a conference paper at ICLR 2027

Fragility-Aware Noise Training for Robust Split Computing over Noisy Channels

Abstract

Split computing (SC) partitions deep neural networks (DNNs) between edge and cloud devices, reducing edge-side computation while leveraging powerful cloud resources. However, in practice, intermediate features must traverse non-ideal communication links, where diverse perturbations can corrupt task-relevant representations and substantially degrade inference performance. To address this challenge, we develop a biologically inspired SC framework that progresses hierarchically from peripheral perception through neural transmission to central cognition, and further propose Fragility-Aware Noise Training (FNT). FNT estimates a relative fragility field over intermediate features, converts it into a bounded local SNR field, and allocates stronger but controlled perturbations to fragile locations under a min–max objective, encouraging the split model to learn more robust intermediate representations. Extensive experiments across diverse tasks, architectures, and split points show that FNT consistently improves robustness, achieving an average accuracy gain of 8.75% over the strongest baseline and a gain of 36.27% under the strongest additive perturbations. Moreover, under cross-noise evaluation, FNT retains strong cross-noise robustness, improving average accuracy by 10.75% and the performance under the strongest unseen perturbations by 28.19%. These results demonstrate that FNT improves the robustness of intermediate-feature transmission under diverse and unmatched channel perturbations.

Primary area:
optimization
Keywords:
Robustness, Split Computing, Feature Transmission, Adversarial Optimization, Noisy Channels

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

The paper studies split computing (SC) in which a neural network is divided into two sub-networks to be executed by the edge device and the cloud server. An intermediate feature z is therefore computed by the edge sub-network and is sent to the cloud sub-network over a noisy link, modeled as a perturbation of z. The paper proposes Fragility-Aware Noise Training (FNT) to make the intermediate features robust. To this end, a lightweight ScoreNet finds the sensitive regions of z by mapping it to per-location scores, which are turned into local SNRs used for robust training. Specifically, during training, heteroscedastic Gaussian noise at these SNRs is added to z. The ScoreNet is later discarded at deployment. The pipeline is framed as a bio-inspired perception–transmission–cognition framework. The framework is evaluated on nine configurations and is compared with Vanilla training, COMSPLIT-style two-level noise training (COM), feature-space FGM, and an adapted DeepJSCC under AWGN. The reported mean gains over the strongest baseline are 8.75% (five-SNR average) and 36.27% (at −10 dB), and 10.75/28.19% in the cross-noise setting.

Critical strengths
  1. FNT changes only the training procedure and leaves no inference overhead in deployment time. At −10 dB AWGN, it almost maintains its clean accuracy in all nine configurations, while the strongest baseline at that SNR retains only 8% according to Table E.3. The advantage is also clear in packet erasure and colored noise on the two configurations tested based on Table E.2. The testbed covers CNNs and a vision–language transformer on natural, medical, and multimodal data.
  2. The appendix provides per-SNR means and standard deviations over 10 channel realizations for every method and configuration, baseline formulations and hyperparameters, sensitivity analyses, allocation statistics, and an intervention test. The authors also acknowledge that baselines are comparable or even better under mild noise.
Critical weaknesses
  1. The main baseline comparison does not study the contribution of fragility-aware noise allocation alone. Specifically, Appendix A states that FNT trains with local SNRs between -14 and -3 dB, while COM uses a two-level SNR schedule, FGM uses a fixed feature-space L2 radius of 0.1, and DeepJSCC uses a 10 dB training SNR with a 1/6 channel ratio. More importantly, FNT minimizes clean plus noisy loss (Equation (12)), while COM, FGM, and DeepJSCC objectives use noisy loss only. Even though the stronger-noise test in Table A.1 is useful, changing each baseline's perturbation strength separately is not considered a matched-budget, matched-loss comparison. Therefore, the large performance gaps cannot be attributed to learned fragility allocation in particular.
  2. The necessity of input-dependent fragility learning and its difference from existing adaptive feature-noise methods is unclear. For example, the "without ScoreNet" ablation uses a feature-magnitude-based substitute, instead of uniform, randomized, or fixed-mask allocations matched for total noise power. The corresponding results in Table E.4 reports very high cross-sample correlations between learned allocation maps. This means that it is possible that a largely static allocation could have caused much of the benefit. Even though Table E.5 supports the fragility ranking on one configuration, it does not report paired confidence intervals for the top-versus-random differences.
Questions for authors
  1. Can you report a controlled comparison in which a uniform/random-noise baseline uses FNT's sampled SNR range, matched expected perturbation power, and the same clean-plus-noisy loss, preferably on CIFAR-10 split6 and one larger task? I am curious to see how much of the reported gain remains in this case?
  2. Does FNT outperform a fixed dataset-average allocation map or a shuffled ScoreNet map when the local-SNR histogram and total noise budget are held constant? Can you also report a paired bootstrap interval for the top-10%-versus-random intervention in Table E.5? These results can clarify whether sample-specific learned fragility is necessary.
3 comments
AuthorsOpenReview ↗

Examines the contribution of fragility-aware allocation through uniform, random, fixed, and shared allocation controls, supported by paired statistical analysis, while acknowledging differences among the original baselines.

We thank the reviewer for the careful assessment of experimental controls and performance attribution, particularly the suggestions to match training objectives and noise budgets and to examine shared allocation maps. All supplementary results are available in our Anonymous Repository.

To address W1, W2, Q1, and Q2, we conduct three noise-allocation ablations: uniform, randomized, and fixed-mask allocation. W2 asks for more comprehensive ablations; W1 questions whether fragility-aware allocation itself explains the gains; Q1 requests matched-objective, matched-budget tests on two datasets; and Q2 asks whether sample-specific allocation outperforms fixed or shared maps.

W2 (allocation ablations). Uniform Allocation distributes noise evenly while matching FNT's expected total perturbation power. Randomized Allocation preserves each sample's local-SNR distribution but randomly permutes allocation locations. Fixed-Mask Allocation uses an allocation map generated from a fixed reference sample. Table R8 (r8.pdf) and Table R9 (r9.pdf) show that FNT achieves higher average AWGN accuracy than these controls, with larger gains under severe noise. The results support the contribution of learned fragility-aware allocation beyond the effect of noise strength alone.

W1 (isolating the contribution of allocation). By matching perturbation budgets or preserving local-SNR distributions while using the same training objective, we isolate the contribution of fragility-aware location selection. Table R8 shows that FNT still outperforms uniform and randomized allocation on both CIFAR-10 and ImageNet. Thus, its gains cannot be attributed solely to stronger training noise or the additional clean-loss term; learned allocation contributes independently.

Q1 (strictly controlled comparisons). Following the reviewer's request, we evaluate uniform and randomized allocation on CIFAR-10 and ImageNet, using FNT's training-SNR sampling range, Clean + Noisy Loss, and other training settings. In Table R8 (r8.pdf), FNT improves average AWGN accuracy over Uniform and Randomized Allocation by 4.19 and 3.54 percentage points (pp) on CIFAR-10, and 1.41 and 1.75 pp on ImageNet, respectively. At −10-10 dB on ImageNet, the corresponding gains are 3.88 and 4.44 pp. These results demonstrate that fragility-aware allocation provides additional robustness benefits under matched experimental conditions.

Q2 (sample specificity). Since Table E.4 shows high cross-sample correlations between allocation maps, we further compare FNT with fixed-mask and batch-shared allocation on ImageNet. Table R9 (r9.pdf) shows that FNT improves average AWGN accuracy over these two controls by 1.97 and 3.83 pp, respectively, with gains of 4.53 and 8.76 pp at −10-10 dB. Although allocation maps share a common structure, sample-specific learning captures useful differences that fixed or shared allocation does not fully exploit, particularly under severe noise.

We will incorporate these experiments and analyses into Sec. 4.3, Appendix A.1, and Appendix E.3, clarifying the independent contribution and sample specificity of FNT's allocation mechanism.

W1 (fairness of the original baselines). We acknowledge that COM, FGM, and DeepJSCC differ from FNT in training-noise strength, perturbation mechanisms, and loss functions. The original comparisons assess complete methods with their respective mechanisms, rather than strictly matched noise budgets and objectives. We therefore do not attribute the entire baseline gap to fragility-aware allocation; instead, the new controlled experiments isolate its contribution. We will clarify these comparisons in Appendix A.1.

W2 and Q2 (allocation consistency and paired statistics). Table R10 (r10.pdf) reports cross-sample correlations of 0.9012–0.9997, but only 55.39%–72.53% overlap among Top-10% fragile locations. Thus, substantial shared structure coexists with differences in the most fragile locations. To address the request for paired uncertainty in the original Table E.5, we use 256 images, 10 interventions per image, and 50,000 image-level paired bootstrap resamples. Table R11 (r11.pdf) reports 95% confidence intervals for the Top-10%-minus-random loss difference of [0.001799,0.028263][0.001799, 0.028263] at 00 dB and [0.021137,0.075428][0.021137, 0.075428] at −5-5 dB. Both intervals exclude zero, supporting ScoreNet's ability to identify more fragile feature locations. We will add these results and statistical details to Appendix E.3.

We again thank the reviewer for the constructive comments on fairness, allocation mechanisms, and sample specificity. We will incorporate the new experiments and statistical analyses to improve the rigor and clarity of the revised manuscript.

Thank you for the additional experiments. While the rebuttal establishes a benefit over the tested controls, given the modest average gains on ImageNet and the remaining uncertainty about the added value of per-input allocation over a stronger static baseline, I will maintain my rating.

AuthorsOpenReview ↗

Clarification of FNT's robustness gains under severe noise and the additional benefits of per-input fragility-aware allocation over shared strategies.

We sincerely thank the reviewer for carefully considering our additional experiments, recognizing FNT's gains over the controlled baselines, and raising a valuable remaining question about stronger static allocation. We appreciate this constructive feedback and would like to clarify two points.

ImageNet performance. First, FNT targets robustness under severe channel noise, rather than maximizing the mean over five SNR levels. In Table 1 of the manuscript (ImageNet / ResNet50 / split3), FNT improves average AWGN accuracy over COM from 73.12% to 74.26% (+1.14 pp), but at −10-10 dB the gain is much larger: 59.00% to 68.76% (+9.76 pp). Table E.1 further reports 89.23% vs 83.30% Top-5 accuracy at −10-10 dB. Thus, averaging across SNRs understates the benefit in the severe-noise regime central to our work. Second, ImageNet-1K is a challenging 1,000-class benchmark, where 1–2 pp can be meaningful in context. For perspective, AutoAugment (Cubuk et al., CVPR 2019) reports a 0.4-pp ImageNet Top-1 gain, and RandAugment (Cubuk et al., NeurIPS 2020) reports 1.0 pp with EfficientNet-B7. These results are not directly comparable to FNT, but illustrate why gains on large-scale classification merit contextual evaluation. We will describe the dataset and provide a more detailed analysis of average versus low-SNR performance in the revision.

Per-input allocation. All results are in our anonymous repository. Table R8 (r8.pdf) matches training objectives and perturbation budgets for Uniform, Randomized, and FNT. On CIFAR-10, FNT gains +4.19/+3.54 pp in mean AWGN accuracy over Uniform/Randomized; on ImageNet, it gains +1.41/+1.75 pp on average and +3.88/+4.44 pp at −10-10 dB. These controlled comparisons show that learned fragility-aware allocation matters beyond overall noise strength. Crucially, Table R9 (r9.pdf) tests two shared strategies that already reuse ScoreNet's task-loss-learned allocation patterns: Fixed-Sample Shared uses a fixed reference sample's map, while Batch-Shared assigns one sample's map to the rest of its training batch. Their retained robustness is itself consistent with FNT's central insight that learning where to allocate noise is useful; however, both break the correspondence between each input and its own fragility map. FNT exceeds the fixed and batch-shared controls by +1.97/+3.83 pp on average and +4.53/+8.76 pp at −10-10 dB, respectively. Thus, per-input matching provides measurable benefits beyond learning a broadly useful shared pattern.

Mechanistic support and scope. Table R10 (r10.pdf) reports high cross-sample correlations (0.9012–0.9997) but only 55.39%–72.53% overlap among the Top-10% fragile locations, consistent with shared structure plus sample-specific differences. Table R11 (r11.pdf) provides paired bootstrap evidence that ScoreNet-ranked vulnerable locations are more sensitive than random locations (95% loss-difference CIs: [0.001799,0.028263][0.001799,0.028263] at 00 dB and [0.021137,0.075428][0.021137,0.075428] at −5-5 dB). Together, these results support per-input allocation over the evaluated shared strategies. We agree that a reference-sample map is not a fully optimized global static baseline, so we do not claim superiority over every possible static strategy. We will clarify these distinctions in the revision and consider broader datasets, models, and stronger static controls in future work. We thank the reviewer again for the thoughtful assessment.

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

The paper examines the robustness of split neural networks in the case where the intermediate features being transmitted are corrupted. Fragility-Aware Noise Training (FNT) employs a ScoreNet to calculate feature-wise scores, normalises these scores on a per-sample basis, and then converts them into local SNR values which are used to determine the variances of the Gaussian perturbations. While the ScoreNet aims to maximise the noisy task loss, the split model itself seeks to minimise both the clean and the noisy task losses. After training the auxiliary network is discarded, so that both the final architecture and the transmitted representation keep their original dimensions. The study includes broader evaluations as well as further tests involving packet erasure and coloured Gaussian noise. The appendix provides information on component ablations, tests using a stronger noise baseline, analyses of clean accuracy and seed effects, and an equal-budget intervention on the learned scores. The results make a valuable empirical contribution. However, it is not yet fully established to what extent the improvements specifically depend on the allocation of sample-dependent fragility.

Critical strengths
  • A useful training method, supported by extensive evidence of improved robustness against severe noise. The improvements are most notable in cases of severely corrupted data; therefore, the results cannot be interpreted as indicating uniform superiority across the entire noise range.

  • A number of additional experiments support the mechanism that is proposed. When ScoreNet is replaced by feature-magnitude scores, the mean accuracy decreases by more than 4.6 points among the three corruption families in the configuration that was tested. This manuscript shows broad evaluation across various benchmarks and shows good performance enhancements. Although these experiments do not answer all the questions regarding attribution, but the evidence manuscript provide is worth considerable.

Critical weaknesses

W1.Reproducibility is limited by the lack of code and unclear implementation details.

W2. The approach of learning an input-conditioned adversarial noise distribution is already present in ADT, together with its amortised Gaussian-based version. Other relevant examples of learned noise are provided by ANT, NAN, and AdaNI. The possible advance of FNT lies in the standardised, feature-conditioned allocation of SNR for split inference and in the empirical effectiveness hence the paper should make this distinction clear. Because these methods have different objectives and use different perturbation families, clarification is needed to show how they differ from previous baselines.

The results also indicate a trade-off: in all nine configurations FNT lies below the strongest clean baseline (as shown in Table E.3). Since DeepJSCC has a different transmission budget and training SNR and FGM employs an absolute feature-space radius of 0.1, these comparisons not sufficiently prove that FNT has surpases performance under equal resource conditions. The COM approach and the further experiment involving stronger noise still provide more supportive evidence in favour of FNT. Lastly, the validation procedure used to select the settings in Figs. A.1 to A.2 needs to be clarified before the cross-noise results can be interpreted as indicating generalisation to conditions that were not encountered during the entire model development process.

Relevant references: Dong et al., NeurIPS 2020, Rusak et al., ECCV 2020, Ma et al., WACV 2020, Li et al., CVIU 2024.

Questions for authors
  1. Referring to Table E.4, how much of the benefit is due to input-specific scores rather than a shared allocation that has been learned. In Table E.5, can the current per-image results give a paired confidence interval for the difference between the top result and the random one.

  2. Regarding novelty and the scope of comparison: What specific improvement does the feature allocation based on the SNR offer over that in ADT and ANT, and in what way does it differ from the feature-adaptive noise used in NAN and AdaNI? Explain the actual perturbation-to-feature norm ratio for FGM and the intended meaning of DeepJSCC's different transmission budget.

  3. Is the selection of the default hyperparameters and checkpoints based on an independent validation set prior to the final testing? Please clarify.

2 comments
AuthorsOpenReview ↗

Addresses novelty, reproducibility, and baseline fairness by providing core code, discussing related methods and experimental settings, and presenting evidence for sample-specific allocation.

We thank the reviewer for the insightful comments on methodological novelty, baseline fairness, and sample-specific fragility learning. These suggestions have helped us clarify FNT's technical contributions and better understand the sources of its robustness gains. Supplementary results and core code are available in our Anonymous Repository.

W1 (reproducibility). We originally planned to release the implementation after the review discussion and code cleanup. To facilitate inspection and reproducibility during review, we have now made the core implementation available in the anonymous repository.

W2 and Q2 (novelty). We acknowledge that ADT, ANT, NAN, and AdaNI have explored adaptive or learnable noise, mainly for adversarial robustness or visual corruptions. FNT instead addresses intermediate-feature transmission impairments in split inference. Unlike NAN, which generates adversarial feature-space perturbations, or AdaNI, which guides noise using feature responses, FNT uses ScoreNet to learn task-loss-guided, sample-dependent, location-wise fragility. Per-sample standardization, bounded local-SNR mapping, and feature-power scaling then enable controlled, fragility-targeted noise training, with no additional inference overhead. To isolate the allocation mechanism, Table R3 (r3.pdf) reports controls that either preserve each sample's local-SNR distribution while shuffling locations or match total expected perturbation power with uniform allocation. Under the same training objective, FNT improves average AWGN accuracy by 3.54 and 4.19 percentage points (pp), respectively, while also outperforming both controls under other noise types. We will clarify these distinctions in Related Work, Method, and Ablation Studies.

W2 and Q2 (fairness and transmission budget). We keep the backbone and test conditions consistent while retaining each baseline's original mechanism, but acknowledge differences in training objectives, perturbation strengths, and transmission budgets. On CIFAR-10/MobileNetV2/split6, FGM uses a feature-space perturbation radius of ϵ=0.1\epsilon=0.1, with an average perturbation-to-feature norm ratio of 0.1780.178%; DeepJSCC compresses 2048 real-valued features into 192 complex channel symbols (384 real values). FNT aims to improve robustness to transmission impairments, not to reduce communication cost, a limitation already noted in the manuscript. Relative to Vanilla, FNT sacrifices only 0.46 pp in clean accuracy but gains 56.13 pp at −10-10 dB. Thus, the original baseline comparisons assess complete methods rather than strictly matched-resource settings; our new controlled experiments isolate the allocation contribution. We will clarify this in Appendix A.1.

Q1 (sample specificity and statistical reliability). Although Table E.4 reports high correlations between allocation maps across samples, sample-specific differences may still matter. In Table R6 (r6.pdf), we add an ImageNet control in which one randomly selected sample's allocation map is shared across each training mini-batch, with other settings unchanged. FNT improves average AWGN accuracy by 3.83 pp and accuracy at −10-10 dB by 8.76 pp, supporting the value of sample-specific allocation. For Table E.5, we additionally compute 50,000 paired bootstrap resamples at the image level, using 256 images with 10 interventions each. Table R7 (r7.pdf) reports the 95% paired confidence intervals for the Top-10%-versus-random loss difference: [0.001799,0.028263][0.001799,0.028263] at 00 dB and [0.021137,0.075428][0.021137,0.075428] at −5-5 dB. Both exclude zero, supporting ScoreNet's ability to identify fragile locations. We will add these analyses to Appendix E.3.

W2 and Q3 (hyperparameters and checkpoints). We use separate training, validation, and test sets. The defaults κ=1\kappa=1, ΔS=4\Delta S=4 dB, and reference SNR range [−10,−3][-10,-3] dB were determined through AWGN sensitivity analyses on the validation set and fixed before final testing. The sensitivity trends for κ\kappa are also consistent across the three tested noise types, although FNT is trained only with AWGN. FNT uses the checkpoint from the final epoch of a predetermined training schedule rather than selecting a checkpoint based on test performance. We will clarify this protocol and the scope of cross-noise evaluation in Appendix A.2.

We thank the reviewer again for the constructive suggestions. We will incorporate the additional experiments, clarify implementation and comparison settings, and strengthen the rigor and transparency of the revised manuscript.

We sincerely thank the reviewer again for the thoughtful and constructive feedback. In response to your concerns, we have released our core implementation, clarified the methodological novelty and baseline fairness, and provided additional controlled experiments and statistical analyses. All supplementary materials are available in our Anonymous Repository.

We sincerely hope that our additional experiments and clarifications have adequately addressed your main concerns. We would greatly appreciate your feedback on whether our responses have resolved your concerns or whether any questions remain. We would be more than happy to provide further explanations and engage in additional discussion. Thank you again for your valuable time, careful review, and constructive suggestions.

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

This paper studies prediction errors caused by noise in intermediate features transmitted between the edge and cloud parts of a neural network. It proposes FNT, where ScoreNet learns relative fragility scores and assigns stronger Gaussian noise to high scoring feature locations during training, while the split model learns from both clean and noisy features. Experiments across several datasets, architectures, and split points report improved accuracy, especially under strong noise, and transfer to packet erasure and correlated noise, although clean accuracy can decrease.

Critical strengths

Table E.5 provides useful evidence for the learned ranking: under the same perturbation budget, disturbing high scoring locations causes a larger average loss increase and accuracy drop in the evaluated setting.

FNT shows substantial accuracy gains under strong corruption across several datasets and split points. Tables E.2 and E.3 also report clean and milder noise results, making the accuracy tradeoff clear.

Critical weaknesses

Table A.1 and Figure 7 support FNT, but do not fully separate the effect of where noise is added from how its strength is chosen. Table A.1 uses different training noise distributions, and Figure 7 does not state whether the noise strength distribution is kept the same when ScoreNet is replaced. Keeping the same noise strengths and training objective while randomly assigning the strengths to feature locations would make the contribution of learned location selection clearer.

The training noise does not need to match real noise exactly, but the practical meaning of the test corruptions is unclear. For example, Appendix C groups 256 feature values into each packet and erases packets independently. Using image recognition tasks does not by itself justify this packet layout or loss model. The authors should explain the intended communication setting and whether these losses remain after error correction or retransmission.

Questions for authors

In Figure 7, is the distribution of noise strengths kept the same when ScoreNet is replaced? If not, do you have an existing comparison that changes only where the same noise strengths are applied?

Table E.2 shows transfer from Gaussian training to packet erasure. What explains this transfer, and when might it fail? Please discuss feature redundancy and the choice of 256 feature values per packet in the intended transmission setting.

Were the default settings fixed before the packet erasure and correlated noise evaluations, or were these results used to select them?

How does FNT differ from PNI and Learn2Perturb in its noise allocation and training objective?

Figure A.2 varies training noise strength but does not show clean accuracy for these settings. Please discuss whether weaker training noise could reduce or avoid the clean accuracy loss while retaining useful robustness gains, using any existing results.

2 comments
AuthorsOpenReview ↗

Clarifies the contribution of learned noise allocation and the communication assumptions, while discussing cross-noise transfer, failure cases, and the trade-off between clean accuracy and robustness.

We thank the reviewer for constructive comments on noise allocation, communication modeling, and cross-noise generalization. Supplementary results are in our Anonymous Repository, with Tables R1–R5 provided as r1.pdf–r5.pdf.

W1 and Q1 (noise strength vs. location). We agree that replacing ScoreNet in Fig. 7 may change both perturbation strength and feature allocation location. Table R1 (r1.pdf) compares local-SNR distributions; Table R2 (r2.pdf) quantifies location changes. The distributions are not identical, and the overlap ratios of strongest and weakest 1010% noise locations are only 18.3818.38% and 4.904.90%, respectively. We thus add two controls in Table R3 (r3.pdf): Random Assignment preserves each sample's local-SNR distribution but randomly permutes locations; Matched-Power Uniform distributes noise uniformly at matched expected total perturbation power. Both use the same training objective and settings as FNT. At comparable clean accuracy, FNT improves average AWGN accuracy by 3.54 and 4.19 percentage points (pp) over Random and Uniform, respectively, and also outperforms both under packet erasure and colored Gaussian noise. These results support the independent contribution of learned fragility-aware allocation. We will update Sec. 4.3, the Fig. 7 discussion, and Appendix D.

W2 (communication modeling). FNT targets edge–cloud split inference over wireless or packet-switched networks. AWGN, packet erasure, and colored Gaussian noise represent additive receiver noise, packet loss, and correlated interference. Each packet contains 256 FP32 features (1024 bytes), which fits within the 1500-byte Ethernet MTU (RFC 894) under ordinary IPv4/UDP encapsulation. Itahara et al. (IEEE GLOBECOM 2021) also study packetized split inference over lossy wireless networks. With real-time deadlines, packets may remain unavailable despite error correction or retransmission. Our models assess feature robustness under communication impairments rather than fully simulate physical channels or optimize packet recovery. We will clarify these assumptions in Appendix C.

Q2 (cross-noise transfer and limits). Rusak et al. (ECCV 2020) found that Gaussian-noise training can improve robustness to unseen corruptions. Similarly, although FNT is trained only with AWGN, it improves robustness to packet erasure and colored Gaussian noise in most evaluated settings. One possible explanation is that FNT reduces the model’s reliance on a few sensitive features. Yet Table R4 (r4.pdf) reveals a failure case: on CIFAR-10/MobileNetV2/split3 at a 5050% packet-erasure rate, FNT reaches 48.2648.26%, below Vanilla's 53.1453.14%. Each 256-value packet corresponds to an entire feature channel, so erasure can remove essential information. Cross-noise transfer may therefore depend on feature redundancy and impairment structure. We will add this case to Sec. 4.2 and Appendix C.

Q3 (hyperparameters). FNT is trained only with AWGN. The default κ\kappa, ΔS\Delta S, and reference-SNR range were selected based on AWGN sensitivity experiments; κ\kappa additionally exhibits consistent trends across AWGN, packet erasure, and colored Gaussian noise. We will clarify parameter selection versus cross-noise evaluation in Appendix A.2.

Q4 (PNI and Learn2Perturb). These methods primarily improve adversarial robustness using learned noise parameters or perturbation modules. FNT instead addresses corrupted intermediate features: ScoreNet learns sample-dependent, location-wise fragility and maps it to bounded local SNRs. During training, ScoreNet maximizes the noisy task loss, while the split model minimizes the sum of the clean and noisy task losses. ScoreNet is removed at inference, adding no overhead. We will clarify these distinctions in Related Work.

Q5 (clean–robustness trade-off). Table R5 (r5.pdf) reports additional Chest CT/ResNet18/split2 results. Changing the training reference-SNR range from [−10,−3][-10,-3] dB to [−4,−3][-4,-3] dB increases clean accuracy from 68.2568.25% to 70.4870.48%, but reduces average AWGN accuracy from 69.1169.11% to 62.9362.93%. The [−10,−6][-10,-6] dB setting achieves 75.5675.56% clean accuracy and 72.7972.79% average AWGN accuracy. Thus, training-noise strength affects the clean–robustness trade-off; although some settings incur clean-accuracy losses, robustness gains under severe noise are more pronounced. We will add these results to Appendix A.2.3.

We thank the reviewer again for the helpful feedback and will incorporate these results and clarifications into the revised manuscript.

We sincerely thank the reviewer again for the thoughtful and constructive feedback. In response to your concerns, we have provided controlled experiments isolating the contribution of fragility-aware noise allocation, clarified the practical communication assumptions and packetization model, and further analyzed cross-noise generalization, failure cases, methodological differences, and the clean-accuracy–robustness trade-off. All supplementary experiments and detailed results are available in our Anonymous Repository. We will incorporate these findings and clarifications into the revised manuscript.

We sincerely hope that our additional experiments and explanations have adequately addressed your main concerns. We would greatly appreciate your feedback on whether any questions remain unresolved and would be happy to provide further clarification or engage in additional discussion. Thank you again for your valuable time and careful consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.