acceptodds
Under review as a conference paper at ICLR 2027

Same Identity, Relaxed Evidence: MarginCell for RF Domain Generalization

Abstract

Radio-frequency fingerprint identification (RFFI) provides lightweight physical-layer authentication solution in zero-trust wireless networks. In practical deployments, however, its recognition performance degrades sharply under unseen environments. Existing robust RFFI schemes follow a suppress-and-stabilize paradigm that seek to eliminate environment-induced variation, leaving residual-shift tolerance unmodeled. To tackle this, we reformulate source-only RF domain generalization via decision-evidence geometry. We prove that true-vs-rival margins form decision-complete evidence coordinates, under which each identity occupies a relaxed region of decision-equivalent states rather than a single point. This geometry yields a maximal reserve-preserving evidence cell and an exact bottleneck robustness radius, while revealing cross-entropy’s intrinsic contraction toward point evidence. Building upon this insight, we propose MarginCell, a source-only regularizer that combines a bounded dispersion reward for releasing decision-safe variation with a distance-to-cell barrier against reserve-eroding deviations. Experiments show consistent improvements in cross-environment recognition and complementary gains with diverse methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Domain Generalization, Radio Frequency Fingerprint Identification, Robust Classification, Margin-Based Regularization.

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_hEFMOpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript addresses the generalization challenge of radio frequency fingerprint recognition in practical applications, with a particular focus on the significant drop in recognition performance in unseen environments. To tackle this issue, the authors reformulate the source-only RF domain generalization problem through decision-evidence geometry and demonstrate that the margins between true and rival classes constitute decision-complete evidence coordinates. Within these coordinates, each identity occupies a loose region composed of decision-equivalent states, rather than a single point. Ultimately, the authors introduce a MarginCell regularizer, which constructs relative evidence coordinates rikr_{ik} using the class mean spacing within a mini-batch and combines Dispersion Reward and Boundary Barrier through a unified potential function. Experiments conducted on three public datasets, WiSig, UCLA-CORES, and LoRa, demonstrate that this method achieves significant performance improvements over existing approaches.

Critical strengths

Unlike conventional patchwork papers, this manuscript is grounded in the geometric completeness of pairwise intervals at the classifier output; it analyzes the anisotropic curvature contraction effect of the cross-entropy Hessian matrix, offering a strong theoretical foundation. Furthermore, the manuscript presents an extensive range of experiments.

Critical weaknesses

Why does intentionally introducing a negative relative shift (rik<0r_{ik} < 0) in the source domain actually enhance the reserves for the target domain?

Questions for authors

(1) When applying symmetric dispersion rewards under the zero-mean constraint, will it cause half of the samples to shift in the negative direction and reduce their bottleneck decision-making reserves?

(2) The theorem in the method section assumes that the class mean interval MykM_{yk} is a fixed constant. However, in Sections 4.1 and 4.3, the author specifically emphasizes that Notably, the minibatch reference MykM_{yk} is computed dynamically and remains differentiable.

(3) Classification errors only occur when rik≤−1r_{ik} ≤ -1, and a positive offset rik>0r_{ik} > 0 represents a stronger classification margin than the class mean. How will performance change if an asymmetric potential function is adopted?

(4) The author can consider the performance changes under different batch sizes |B|, that is, the sensitivity of the number of samples per class nyn_y within the small batch and the class scalability.

5 comments

Dear Reviewer hEFM,

Thank you for your constructive review. We are sincerely grateful for your recognition of the theoretical grounding of our analysis, including the curvature contraction of the cross-entropy Hessian, and the extensiveness of our experiments. We provide a point-by-point response to each point below; revisions in the manuscript are highlighted in blue.

W1. Release Mechanism and Target Reserve

We appreciate this insightful question and are glad to clarify the underlying intuition. Because coordinates are centered at the class reference (∑irik=0\sum_i r_{ik}=0), releasing evidence around the mean naturally produces negative deviations as the counterpart to positive ones, while the gain in target reserve comes from the bounded release as a whole rather than from the negative offsets alone. Crucially, the reserve cost on the negative side is controlled: reserve depends only on the worst rival coordinate, s=1+min⁡kriks=1+\min_k r_{ik}, which the barrier penalizes below −τ-\tau, so every in-cell sample retains s≥1−τs\ge1-\tau, well away from the decision boundary at r=−1r=-1. The barrier guards precisely the low tail that governs risk, namely the samples whose reserve approaches zero and whose decisions become fragile.

To see how this overall release improves target reserve, since sT≥sS−D∞s_T\ge s_S-D_\infty (Theorem 3.4), once the low tail is bounded, target reserve is determined by how far the shift moves the evidence. Intuitively, compressing each identity's evidence to a single point forces the classifier to over-rely on the feature directions that happen to be most stable in the source; once an unseen environment perturbs those directions, the evidence moves sharply. Allowing bounded variation spreads this reliance across more directions, so that no single shifted direction moves the evidence as much. Theoretically, we have added Proposition 3.5 in Section 3.5, with the complete proof in Appendix A.5, to formalize how bounded dispersion lowers target error. It shows that source dispersion is stationary at the minimum-dispersion state; therefore, whenever unseen shifts misalign with source variation, a bounded mean-preserving release increases source dispersion only at second order while reducing shift-induced variance at first order.

Comparing MarginCell with PointAlign, which compresses evidence toward r=0r=0 without any release, confirms this (UCLA-CORES, Target Pr⁡[s>0]\Pr[s>0] is the target accuracy):

MethodSource Pr⁡[s≤0.5]\Pr[s\le0.5]Target Pr⁡[s>0]\Pr[s>0]Target Pr⁡[s>0.5]\Pr[s>0.5]
ERM2.6765.3738.81
PointAlign1.0070.6055.57
MarginCell0.9676.7463.01

Both keep the source low tail equally small, so the difference lies in the release: MarginCell retains a higher target reserve at both thresholds, e.g., 63.01% vs. 55.57% of samples with s>0.5s>0.5. Bounding the worst case preserves the source reserve; releasing the rest carries that reserve to the target domain.

Q1. Bottleneck Reserve under Symmetric Release

Building on W1, we are glad to clarify how the bottleneck reserve is protected: because coordinates are referenced to the mean, some samples inevitably shift in the negative direction. This shift, however, is bounded: it consumes only surplus reserve, never brings the bottleneck reserve to a dangerous level, and in fact leaves fewer samples at risk. The reward operates only inside the cell and saturates at its boundary, so it pushes down samples near the class reference, whose reserve is close to 1, without pushing them out of the cell. For samples that have already fallen below −τ-\tau, the reward exerts no further downward pull, and the barrier instead pulls them back. Methods without such a floor have no mechanism to keep weak samples from dropping further. Consequently, the source reserve distribution exhibits a "lowered middle, raised floor" profile. On UCLA-CORES, the fraction of samples with reserve below 0.75 rises from 14.26% under ERM to 23.08%, whereas the fraction below 0.5 drops from 2.67% to 0.96%, and the 5% quantile of ss rises from 0.604 to 0.704. More samples spread mildly, while the severely weakened samples, which actually govern the bottleneck reserve, become fewer.

Q2. Class-Level vs. Minibatch Reference

We appreciate this careful reading and are glad to clarify how the theoretical analysis in Section 3 connects to the dynamic implementation in Section 4. Rather than requiring MykM_{yk} to stay constant across training epochs, the phrase "at fixed class-mean margins" in Section 3 means holding the class-mean level fixed at a given classifier state to study mean-preserving deviations around it (mi=M+DMrim_i=M+D_Mr_i with ∑iri=0\sum_i r_i=0). As noted in the Remark of Appendix A.3, Theorem 3.3 characterizes the local decision geometry around the current reference state; since the classifier changes at every step, training tracks this reference online through the minibatch mean. The minibatch coordinates satisfy ∑i∈Byrik=0\sum_{i\in B_y}r_{ik}=0 exactly, so Theorem 3.3 holds at the current reference at every step. Theorems 3.2 and 3.4 use MykM_{yk} only as a positive scale: Theorem 3.2 holds for any positive reference, and Theorem 3.4, which concerns the trained model, requires only that source and target share the same positive reference. In our experiments, this reference is computed once as the final model's class-mean margins over the entire source training set.

Differentiability of the reference does not alter this analysis either. Moving margins along a mean-preserving direction leaves the minibatch mean unchanged, so backpropagating through the reference introduces no first- or second-order term along these directions. The local analysis, including the release threshold αc\alpha_c, is therefore identical to the fixed-reference case (full proof added to Appendix A.3 of the revised manuscript). We keep the reference differentiable because rikr_{ik} is invariant to a uniform rescaling of a class's margins (mik→c mikm_{ik}\to c\,m_{ik}, up to ϵ\epsilon). With a differentiable reference, the regularizer has zero gradient along this rescaling direction, so it shapes only the relative evidence within a class and leaves the margin scale to cross-entropy. A detached reference would ignore this dependence: the regularizer would then produce a spurious gradient that rescales the margins without reducing the regularizer itself. Q4 reports the accuracy of the online estimate: the reference error is small at every batch size, and performance remains robust to it.

Q3. Performance under Asymmetric Potentials

Following this constructive suggestion, we evaluated asymmetric variants on UCLA-CORES, changing only the potential: the reward is either symmetric or positive-only, and the barrier either two-sided or on the negative side only. All variants underperform the symmetric design, especially on hard identities:

Reward / barrierAccWEnvTIA@20Source coords r<−0.25r < -0.25 (%)Source coords r>+0.25r > +0.25 (%)
Symmetric / two-sided (MarginCell)76.74±2.8675.17±1.2153.44±3.2114.857.38
Symmetric / negative side only74.23±5.1572.20±4.8343.45±9.8725.4430.78
Positive-only / two-sided74.68±2.1472.70±2.2045.45±10.993.814.60
Positive-only / negative side only74.85±5.0073.72±4.5043.26±11.867.9924.51

The reviewer is right that classification errors occur only when r≤−1r\le-1, and the upper barrier does not treat positive offsets as errors. Yet both bounding and release need to act on both sides, and each of the first two variants isolates one of them.

  • For bounding, coordinates are measured from the class mean, so positive and negative deviations must balance: unbounded positive deviations must be offset by deeper negative ones, pushing the negative tail toward the decision boundary, while the extra positive margin adds no reserve, since reserve is governed only by the weakest rival, s=1+min⁡kriks=1+\min_k r_{ik}. Removing the upper barrier confirms this: positive excursions rise from 7.38% to 30.78%, negative excursions follow from 14.85% to 25.44%, and TIA@20 drops by 10.0 points.
  • For release, CE contracts evidence toward a single point from every direction, and Theorem 3.3 requires the reward to counteract this contraction in every direction. With both barriers kept, a positive-only reward leaves the negative side contracted; since ∑irik=0\sum_i r_{ik}=0 couples positive and negative offsets, this stalls the release on both sides (fewest deviations: 3.81% and 4.60%). The evidence thus stays nearly point-like and, as in W1, relies on a few source-stable directions that unseen shifts can perturb, dropping mean TIA@20 by 8.0 points, so the loss comes from unreleased freedom rather than from unsafe deviations.

Bounding both sides and releasing both sides are therefore each necessary, which is exactly what the symmetric cell of Theorem 3.2 provides.

Q4. Batch Sensitivity

We appreciate this practical suggestion. To test the sensitivity to ∣B∣|B| and to the per-class count nyn_y, we swept ∣B∣|B| from 512 to 32 on WiSig (2→2 d.), reducing the median nyn_y from 85 to 5 (each cell reports ERM / MarginCell):

∣B∣\lvert B\rvertMedian nyn_yReference deviation (%)AccWEnvTIA@20
3259.6588.08/94.3287.17/91.6777.75/83.35
64116.6784.58/94.7082.50/92.8070.25/84.55
128214.5883.25/98.1783.17/96.5050.00/94.70
200333.6286.83/97.0786.33/95.3061.00/91.50
256433.1883.50/95.1583.00/91.3351.00/85.70
400672.5285.58/93.9882.67/90.5357.50/82.80
512852.2182.75/91.8382.17/89.1349.00/75.85

As nyn_y shrinks, the deviation of the minibatch reference from the full-source reference grows from 2.21% to 9.65%, yet MarginCell consistently outperforms ERM in Acc, WEnv, and TIA@20 at every batch size (by 6.2–14.9, 4.5–13.3, and 5.6–44.7 points), maintaining 94.32% accuracy even at ny=5n_y=5. This is comparable to the per-class regime of the 58-class UCLA-CORES at ∣B∣=256|B|=256 (ny≈4n_y\approx4, 8.70% reference deviation), where MarginCell still exceeds ERM by 11.4 Acc points. Complementing the analysis in Q2, this shows empirically that the gain of MarginCell is robust to the batch size and the resulting reference error. We have added the full results and analysis to Appendix C.8 in the revised manuscript.


Thanks so much for your support and insightful suggestions for improving our work! We are eager to address any remaining concerns you might have.

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

The paper proposes MarginCell to address the substantial performance degradation of RF fingerprint identification (RFFI) in unseen environments. Its key insight is that classification evidence for the same device need not be identical across environments; instead, bounded variation is allowed while maintaining a sufficient advantage of the true class over rival classes. The main contribution is a training regularizer that can be added to existing training objectives without introducing an additional inference branch, aiming to improve RFFI performance in unseen environments. The evaluation covers three public RFFI benchmarks and includes comparisons with existing methods, plug-in experiments that add the MarginCell regularizer to other methods, and component ablations. The results provide evidence for cross-environment performance gains and the contributions of the Reward and Barrier components. Overall, the evaluation is relatively comprehensive.

Critical strengths
  1. A key strength of the paper is its “release-and-bound” perspective on cross-environment robustness. Instead of following the conventional “suppress-and-stabilize” paradigm that seeks to eliminate environment-induced variation, the paper argues that decision-safe evidence variation should be preserved while deviations that erode the decision margin should be constrained. This provides an interesting new perspective on domain generalization for RFFI.

  2. The proposed MarginCell regularizer translates this insight into a simple and practical solution. It can be easily added to existing training objectives without introducing an additional inference branch or changing the deployment pipeline. Comparisons with existing methods and the plug-in experiments show promising and consistent improvements in unseen environments. This combination of empirical performance, ease of integration, and practical applicability is a notable strength.

Critical weaknesses
  1. I’m wondering about the impact of using different source-target splits and model architectures for experiments, since each dataset in the paper uses only one fixed source-target split, and all experiments use the same ResNet-18 backbone. Such experiments would help better support the paper’s general claim about the benefit of “release-and-bound” over “suppress-and-stabilize”.

  2. I'd like to know how well the perturbation model reflects real RF domain shifts, given that the paper's theory claims to provide robustness guarantees for bounded perturbations in the proposed evidence space. I recognize that MarginCell creates bounded evidence variation and that this correlates with better target robustness, but this does not directly validate the perturbation model used in the theory.

  3. There is an unexplained numerical discrepancy between the main results table and the component ablation figure; see Question 1 for the authors. This raises concerns about the reliability of the experimental results and the credibility of the conclusions. If the authors cannot clarify this discrepancy, it would be a significant reason to recommend rejection.

Questions for authors

Please refer to the above weaknesses for what need to be addressed, with particular attention to the following question.

  1. Appendices B.1 and B.4 specify a unified data-splitting and training protocol. However, in Figure 9(b), both Target Acc and Target Worst-Env for the full MarginCell model on UCLA-CORES appear to exceed 90%, substantially differing from the 76.74% Acc and 75.17% WEnv reported in Table 1. Please clarify whether Figure 9 follows this default protocol, specify the random seeds used, explain how the results were aggregated, and define the error bars. If different settings were used, please explicitly list the differences.
5 comments

Dear Reviewer rQbA,

Thank you for your time and expertise in reviewing our submission. We are grateful for your recognition of release-and-bound as "an interesting new perspective" on DG, the practical value of MarginCell, and the comprehensiveness of our evaluation. We provide responses to your concerns below, with revisions highlighted in blue.

W1. Additional Splits and Backbones

Following this constructive suggestion, we have added a second source–target split on all three datasets and two additional backbones, each comparing against all 12 baselines in every setting (Appendix C.4, Tables 6–8 in the revised manuscript). MarginCell achieves the best or tied-best Acc in every new setting:

SettingStrongest baselineMarginCell
Single-target, WiSig 3→1100.00 (LS and others)100.00
Single-target, UCLA-CORES 4→189.06 (LS)91.17
Single-target, LoRa 7→199.20 (LS)99.38
Transformer, WiSig 2→293.64 (TFPS)94.04
SepCNN, WiSig 2→293.31 (CIS)94.56

For domain splits, since RF fingerprints drift over time, we preserve the chronological deployment order (training on earlier days, testing on later days) and vary the split ratio: the single-target protocol uses more source domains and holds out one target domain (WiSig 3→1 days, UCLA-CORES 4→1 days, LoRa 7→1 environments). The gain is largest on the hardest UCLA-CORES benchmark. Under a single target, WEnv coincides with Acc, which is why our main results adopt two-target splits.

For model architectures, we replace ResNet-18 with a Transformer and a depthwise-separable CNN (SepCNN), both commonly used in RFFI. MarginCell achieves the best Acc and TIA@20 under both backbones (e.g., +4.05 points in TIA@20 over the strongest baseline under Transformer), trailing CIS by only 1.56 points in WEnv on SepCNN. Notably, the strongest baseline changes across backbones, and the strongest robust RFFI baseline in Table 1, CLPS (94.33% on WiSig), drops to 80.67% under Transformer, whereas MarginCell consistently leads.

Thus, the advantage of release-and-bound over suppress-and-stabilize does not depend on the specific split or backbone.

W2. Perturbation Model under Real RF Shifts

We appreciate this insightful question on bridging our perturbation model with real RF shifts. The budget ∥δ∥∞≤ρ\Vert\delta\Vert_\infty \le \rho in Theorems 3.2 and 3.4 is a worst-case uncertainty set in evidence space, not a condition that physical RF shifts must satisfy. Since which rival an unseen shift will push evidence toward is unknown in advance and varies across environments, the budget treats all rivals alike in relative terms (a raw-margin change of ρMyk\rho M_{yk}). Moreover, because a decision fails only when some rival margin crosses zero, keeping the largest relative margin change below the reserve, D∞:=∥δ∥∞<s(x,y)D_\infty := \Vert\delta\Vert_\infty < s(x,y), guarantees correctness in any direction. Theorem 3.4 also allows shifts larger than ρ\rho: they simply enter the risk bound through the exceedance rate η(ρ)=Pr⁡(D∞>ρ)\eta(\rho) = \Pr(D_\infty > \rho) in RTbal≤Pr⁡Sbal[s≤ρ]+η(ρ)R_T^{\mathrm{bal}} \le \Pr_S^{\mathrm{bal}}[s \le \rho] + \eta(\rho).

To test directly how much of an RF shift falls within this threshold and is covered by the guarantee, we froze the trained model on LoRa and measured the exceedance rate η^(0.5)\hat{\eta}(0.5) and the certified fraction 1−Bound=1−(Pr⁡Sbal[s≤0.5]+η^(0.5))1-\text{Bound} = 1 - (\Pr_S^{\mathrm{bal}}[s \le 0.5] + \hat{\eta}(0.5)) through the full IQ-to-logit pipeline. Physically, a received packet is x=Fy(u)∗h+νx=\mathcal{F}_y(u)*h+\nu, where the fingerprint Fy\mathcal{F}_y of device yy acts on the baseband waveform uu while a new environment mainly changes the multipath channel hh and the noise ν\nu (now stated in Section 4); we therefore evaluated both the real cross-environment shift and paired channel impairments in ν\nu and hh (AWGN noise from 20 to 5 dB SNR; multipath fading from standard line-of-sight Rician channel model with K = 10 dB ~ -5dB to non-line-of-sight Rayleigh channel model).

Quantity (%)MethodReal20 dB15 dB10 dB5 dBK=10K=3K=0K=−5Ray.
η^\hat\etaERM35.690.000.317.2839.492.3915.4123.3532.9241.03
MarginCell19.120.060.383.6625.171.469.1614.1321.0626.95
1−BoundERM55.2290.6190.3083.3351.1288.2275.2067.2657.6949.58
MarginCell78.0196.7796.4593.1871.6795.3787.6882.7075.7769.88
AccuracyERM87.9399.0698.8696.7884.0898.4294.5491.5286.6782.51
MarginCell93.3499.4199.3198.3990.1198.9196.2894.0390.4587.03

Under typical RF shifts (SNR ≥\geq 15 dB, K=10K=10 dB), nearly all displacements stay within the threshold (η^≤1.46%\hat{\eta} \le 1.46\% for MarginCell), and its bound reaches 95–97%, close to the ~99% observed accuracy. Even under 5 dB strong noise and non-line-of-sight Rayleigh fading, about 70% of MarginCell's packets remain certified, far above ERM. Beyond coverage, the evidence-space model tracks the physical shifts: as noise or fading strengthens, η^\hat\eta rises monotonically, and across controlled shifts the bound ranks the settings consistently with the observed accuracy. The real cross-environment shift also falls within this sweep.

MarginCell attains a higher bound than ERM in every setting, as it keeps more displacement within its reserve: under strong noise and fading (SNR ≤\leq 10 dB, K≤K\leq 3 dB), it reduces the exceedance rate η^\hat{\eta} by about a third to a half. We have summarized this finding in Section 5.3 and added the full results and analysis to Appendix C.7 (Table 9).

W3 & Q1. Protocol Clarification for Figure 9(b)

We are sincerely grateful for this careful check and apologize for the confusion caused by the mixed protocols in the original Figure 9(b). We initially evaluated both the single-target (days 1–4 → day 5) and two-target (days 1–3 → days 4–5) protocols; since WEnv coincides with Acc when there is only a single target day, we retained the more challenging two-target protocol for the main results (Table 1), but inadvertently mixed in the single-target UCLA-CORES subplot when assembling the original Figure 9(b). This is directly visible in the original panel, where Target Acc and Target Worst-Env are identical for every ablation variant.

Regarding Q1, this discrepancy stemmed from the single-target versus two-target domain split, while the backbone, preprocessing, training schedule, and hyperparameters all followed Appendices B.1 and B.4 under five random seeds {37, 73, 101, 137, 173} (each bar reports the mean and one standard deviation across seeds). In the revised manuscript, we have updated Figure 9 to the default two-target protocol of Table 1, so the full MarginCell results on UCLA-CORES now match the main table (76.74% Acc, 75.17% WEnv), and Figure 5 has been updated accordingly.

In addition, we retain and fully report the baseline comparisons under the single-target protocol (Table 6) and the component ablations across all three datasets (Figure 10, with 91.17% Acc/WEnv on UCLA-CORES). Under both protocols, the full MarginCell achieves the best or tied-best target performance on every dataset, leaving our ablation conclusions unchanged. We have also verified the protocol of other figures and tables and now state the data splits and statistical aggregation in Appendices B.1, B.4, and C.6.


We sincerely thank the reviewer again for all the time and effort spent on these professional and insightful suggestions, and we hope our responses resolve the concerns you raised.

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

This paper investigates the source-only domain generalization problem in Radio Frequency Fingerprint Identification (RFFI) and proposes a method called MarginCell. The core argument is that classification evidence for a given identity does not need to be strictly consistent across different environments, it suffices for the true-vs-rival margins to maintain an adequate decision buffer. Based on this observation, the authors reparameterize the classifier's output into relative margin coordinates and construct a symmetric margin cell. By employing a dispersion reward alongside a boundary barrier, the method allows for greater flexibility in intra-class evidence while simultaneously constraining the risk of crossing decision boundaries.

Critical strengths
  1. The paper addresses a genuine cross-environment generalization challenge in RFFI, making the task highly relevant to practical applications.
  2. MarginCell is simple to implement, it requires no additional inference branches or extra trainable parameters, making it an effective "plug-in" component.
  3. The paper’s perspective analyzing domain shift within the classifier's output space rather than the intermediate feature space is insightful.
Critical weaknesses
  1. The exact robustness radius discussed in the paper holds only within the output space, not necessarily against actual RF domain shifts.
  2. The claim that increasing source evidence dispersion improves generalization to unseen domains lacks a rigorous theoretical guarantee.
  3. The method might essentially be a complex margin-variance regularizer rather than a novel domain generalization mechanism.
  4. There is a discrepancy between the theoretical class-level geometry: which relies on the class-rival mean margin (MykM_{yk})—and the actual implementation, where the value is computed dynamically for each minibatch as 1ny∑imik\frac{1}{n_y}\sum_i m_{ik}.
Questions for authors
  1. What is the fundamental difference between using true-vs-rival logit differences as coordinates and the standard multiclass margin geometry?
  2. Why should real-world RF domain shifts satisfy an ℓ∞\ell_\infty-perturbation assumption within the normalized evidence space?
5 comments

Dear Reviewer CnUP,

Thank you for your thoughtful and constructive review. We are sincerely grateful for your recognition of the "insightful" output-space perspective of our work, its real-world significance, and its practicality as an "effective plug-in". We respond to your concerns below, with revisions highlighted in blue.

W1 & Q2. Robustness Radius and ℓ∞\ell_\infty Budget under RF Shifts

We appreciate these valuable questions connecting output-space geometry to physical RF shifts:

  • For the robustness radius (W1), s(x,y)s(x,y) indeed lives in the output space, and the radius ρ\rho in Theorem 3.4 is not a hard threshold that shifts must stay within: the theorem bounds the target risk with two terms, RTbal≤Pr⁡Sbal[s≤ρ]+η(ρ)R_T^{\mathrm{bal}}\le\Pr_S^{\mathrm{bal}}[s\le\rho]+\eta(\rho), namely the fraction of source samples with reserve at most ρ\rho, and the fraction η(ρ)\eta(\rho) of samples whose evidence the shift moves by more than ρ\rho (∥δ∥∞>ρ\Vert\delta\Vert_\infty>\rho). Larger shifts are thus allowed and enter through η\eta; we evaluate this under real RF shifts below.
  • For the ℓ∞\ell_\infty budget (Q2), it is how we measure a shift's effect on a decision, not a condition the shift must satisfy. We adopt it because which rival an unseen shift will favor is unknown in advance, so all rivals are treated alike in normalized terms (a raw-margin change of ρMyk\rho M_{yk}), and because a decision fails as soon as any single rival margin crosses zero, so keeping the largest relative margin change below the reserve ∥δ∥∞<s(x,y)\Vert\delta\Vert_\infty<s(x,y) suffices for correctness in any direction, and ss determines the largest radius with this guarantee.

What matters is thus how far actual RF shifts move the evidence. Physically, a received packet is x=Fy(u)∗h+νx=\mathcal{F}_y(u)*h+\nu: the fingerprint Fy\mathcal{F}_y of device yy acts on the baseband waveform uu, while a new environment mainly changes the multipath channel hh and the noise ν\nu (now stated in Section 4). We therefore measured their effect on the evidence through the full IQ-to-logit pipeline on the LoRa dataset with the trained model frozen, under the real cross-environment shift due to time-varying noise and channel fading: noise (SNR from 20 to 5 dB) and multipath fading (from line-of-sight Rician channel model with K = 10 dB~-5dB to fully non-line-of-sight Rayleigh channel model). Since the source term is fixed for a given model (about 9% for ERM and 3% for MarginCell), we report the exceedance rate η^(0.5)\hat\eta(0.5) and the resulting accuracy lower bound 1−Bound=1−Pr⁡Sbal[s≤0.5]−η^(0.5)1-\text{Bound}=1-\Pr_S^{\mathrm{bal}}[s\le0.5]-\hat\eta(0.5):

Quantity (%)MethodReal20 dB15 dB10 dB5 dBK=10K=3K=0K=−5Ray.
η^\hat\etaERM35.690.000.317.2839.492.3915.4123.3532.9241.03
MarginCell19.120.060.383.6625.171.469.1614.1321.0626.95
1−BoundERM55.2290.6190.3083.3351.1288.2275.2067.2657.6949.58
MarginCell78.0196.7796.4593.1871.6795.3787.6882.7075.7769.88
AccuracyERM87.9399.0698.8696.7884.0898.4294.5491.5286.6782.51
MarginCell93.3499.4199.3198.3990.1198.9196.2894.0390.4587.03

As noise or fading strengthens, η^\hat\eta rises monotonically, so the bound falls entirely through η^\hat\eta. Under typical RF shifts (SNR ≥\geq 15 dB, K=10K=10 dB), the bound is close to the observed accuracy (over 95% vs. 98.9–99.4% for MarginCell). In the low-SNR or severe-fading regime, more evidence moves beyond ρ\rho, and the bound and accuracy decline accordingly. The real cross-environment shift also falls within this sweep. In every setting, MarginCell certifies more packets than ERM. We have summarized this finding in Section 5.3 and added the full results and analysis to Appendix C.7 (Table 9).

W2. Theoretical Guarantee of Limited Release

We are grateful for this constructive point on formalizing how source dispersion benefits unseen domains: in the revised manuscript, we have added Proposition 3.5 in Section 3.5 and its detailed proof in Appendix A.5 to show how a bounded release provably tightens the target error bound.

Consider a class-rival margin of the classifier head on fixed features. Variance enters here only as an analysis tool: Cantelli's inequality bounds the target pairwise error by VT/(VT+M2)V_T/(V_T+M^2), where VT=VS+VξV_T=V_S+V_\xi is the source plus shift-induced variance. At the minimum-dispersion solution, which cross-entropy favors (exactly so for Gaussian features), VSV_S is stationary. A suitably directed small mean-preserving release therefore raises VSV_S only at second order while lowering VξV_\xi at first order, unless the shift is aligned with the source variation, a non-generic case. VTV_T thus strictly decreases over an explicit range of release sizes, tightening the bound; for Gaussian features, the actual error Φ(−M/VT)\Phi(-M/\sqrt{V_T}) and, for small releases, the pairwise bound of Theorem 3.4 strictly decrease as well.

This benefit holds only for a bounded release: beyond this range the quadratic terms dominate, and excessive dispersion erodes the source reserve, which is why MarginCell pairs the reward with a barrier. Empirically, consistent with Proposition 3.5, MarginCell's evidence is moved less by real RF shifts: it lowers the exceedance rate η^\hat\eta by a third to a half under the real shift and strong impairments (W1).

W3. Relation to Margin-Variance Regularization

Thank you for highlighting this distinction. While MarginCell does encourage in-cell dispersion to release contracted evidence, its core contribution is not a variance target, but a DG formulation that explicitly characterizes and exploits decision-safe evidence freedom.

  • Different in what it models. Margin-distribution learning (e.g. Large/Optimal Margin Distribution Machine, LDM/ODM) adjusts the margin mean and variance for i.i.d. generalization without modeling cross-domain shift, while existing DG methods encourage cross-domain consistency of representations, risks, or gradients without modeling the decision space itself. MarginCell takes the opposite view: the same identity need not produce the same evidence across environments, only decision-equivalent evidence with sufficient decision reserve. We therefore characterize which class-competitive evidence variations can safely tolerate unseen shifts, and actively release this bounded freedom.
  • Different in how it is derived. Decision geometry determines the regularizer, rather than a variance objective determining the geometry. Decision geometry first fixes the admissible range of evidence variation: the maximal symmetric region that preserves the decision and its bottleneck reserve is the cell [−τ,τ]K−1[-\tau,\tau]^{K-1} (Theorem 3.2). This range then guides the optimization design: in-cell variation is safe yet contracted by CE toward a point (Theorem 3.3), so it must be actively released, while out-of-cell variation consumes the reserve, so it must be bounded. The in-cell quadratic reward is only a simple smooth realization of the release. Distributions with equal variance but different excursions thus incur different penalties: dispersion is only the means of release, while decision geometry defines the admissible freedom and its boundary.

If MarginCell were merely a variance regularizer, regulating or matching variance on the same coordinates rykr_{yk} would reproduce its gain, yet every such variant falls short (UCLA-CORES):

Potential on rykr_{yk}AccWEnvTIA@20
ERM65.3760.9431.61
LDM-style69.2667.9437.08
ODM-style70.1566.9234.22
Variance minimization71.9170.3839.29
Variance maximization6.415.450.00
Variance matching70.6968.0943.12
MarginCell76.7475.1753.44

Even with the same source variance, Variance matching leaves 6 times more low-reserve samples (Pr⁡S[s≤0.5]\Pr_S[s\le0.5]: 5.86% vs. 0.96%): regulating how much evidence varies is thus not enough without the cell to bound where it may vary (+6.1 Acc and +10.3 TIA@20 points).

Therefore, MarginCell is a DG mechanism because it controls exactly what determines unseen-domain risk. A decision can fail on an unseen domain only if the source sample lacks sufficient reserve or the cross-domain evidence displacement exceeds a threshold (Theorem 3.4); the barrier controls the former, and the bounded release makes decisions less sensitive to unseen shifts (Proposition 3.5). Consistently, MarginCell preserves more decisions across all tested RF shifts (W1) and boosts existing DG methods as an environment-label-free plug-in (Table 2).

W4. Class-Level vs. Minibatch Reference

We appreciate this careful observation regarding the reference construction. As noted in the Remark of Appendix A.3, the class-rival mean MykM_{yk} in Section 3 serves as the reference of the decision geometry at a given classifier state; since the classifier changes at every training step, the minibatch mean tracks this reference online. Crucially, none of our theoretical results requires the exact full-source mean, as their conditions hold directly at the tracked reference: Theorems 3.2 and 3.4 use MykM_{yk} only as a positive scale, so they hold for any positive reference; for Theorem 3.3, the minibatch coordinates satisfy the mean-preserving condition ∑i∈Byrik=0\sum_{i\in B_y} r_{ik}=0 exactly, and moving margins along mean-preserving directions leaves the minibatch reference unchanged (with its differentiable denominator adding no curvature at r=0r=0 since ψ′(0)=0\psi'(0)=0), leaving the release threshold αc\alpha_c unchanged. We have clarified the relation between the two references in Section 4.1 and added to Appendix A.3 a rigorous proof that these properties, including the release threshold of Theorem 3.3, hold exactly under the minibatch construction.

We further verified that this estimate is accurate and its deviation does not compromise the gain. With the trained model frozen, the minibatch reference deviates from the full-source reference by only 8.70%, 3.18%, and 3.24% on UCLA-CORES, WiSig, and LoRa, far below the cell half-width τ=0.5\tau=0.5. Expanding ∣B∣|B| from 32 to 512 on WiSig (median nyn_y from 5 to 85) progressively narrows this deviation from 9.65% to 2.21% toward the full-source mean, while MarginCell consistently outperforms ERM on all metrics across the entire range. We have added the detailed batch-size sweep results to Appendix C.8 in the revised version, confirming this implementation discrepancy is practically benign.

Q1. Relation to Standard Margin

The fundamental difference lies in the reference against which evidence is measured. Standard multiclass margin geometry uses a single reference, the decision boundary myk=0m_{yk}=0, and thus only tells how far a sample is from error, typically via the worst-rival margin min⁡kmyk\min_k m_{yk}. Our coordinates add a second, class-level reference, the class-rival mean MykM_{yk}, via ryk=(myk−Myk)/Mykr_{yk}=(m_{yk}-M_{yk})/M_{yk}, which places the class reference at r=0r=0 and every rival boundary at r=−1r=-1, without losing any decision information (Proposition 3.1). This yields two capabilities absent from standard margin geometry: measuring from the class reference reveals how evidence can vary within a class while the decision stays unchanged, the decision-equivalent freedom that MarginCell regulates (W3); and aligning all boundaries at −1-1 lets a single cell width τ\tau grant the same reserve fraction against near and far rivals. Standard margin geometry is retained as a special case: the worst-rival margin reappears as the bottleneck reserve s=min⁡kmyk/Myks=\min_k m_{yk}/M_{yk} (Theorem 3.4).


We are truly grateful for your constructive suggestions for improving the paper. We hope that your concerns have been addressed and are happy to engage in further discussions.

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

We are sincerely grateful to the Area Chair for the time devoted to our submission, and to all reviewers for their thoughtful and constructive suggestions. As the discussion comes to a close, we provide a brief summary below.

All three reviewers expressed consistent support and positive assessments of our paper, with all raised concerns fully addressed. MarginCell introduces a new release-and-bound perspective for RF domain generalization. Each reviewer recognized the novelty and insight of our work, describing it as “an interesting new perspective” (rQbA), noting its “insightful” output-space analysis (CnUP), and distinguishing it from “conventional patchwork papers” (hEFM). They also highlighted its solid theoretical grounding in decision-evidence geometry (hEFM), practical value as an effective plug-in (CnUP, rQbA), and extensive experimental validation (rQbA, hEFM).

Following the reviewers’ suggestions, we strengthened the revised manuscript (highlighted in blue), including additional theoretical analysis (Proposition 3.5), validation under real RF shifts, and expanded splits and backbones.

Thank you so much again for your time and dedication!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.