acceptodds
Under review as a conference paper at ICLR 2027

Reliable Posterior Support for Long-Tailed Adversarial Robustness

Abstract

Adversarial training under long-tailed distributions suffers from severe class-wise robustness degradation, especially on tail and worst-performing classes. Existing methods improve long-tailed robustness through rebalancing, distillation, geometric regularization, or posterior-aware supervision, but they largely overlook whether the posterior signal is supported by a reliable local neighborhood. We argue that robust supervision should depend not only on posterior confidence at a training sample, but also on the reliability of the posterior support explored by adversarial perturbations. This issue is most severe for tail classes, whose support is estimated from scarce data and is therefore noisy, high-variance, and easily distorted. We propose a reliable posterior support framework, FEPT(Free Energy-guided Posterior Transport), which constructs free-energy posterior support to down-weight posterior mass in off-support, low-density, or unstable regions, and transports reliable low-free-energy support from related head/medium classes to tail classes. Our analysis formalizes why local margin stability and variance-reducing posterior transport are useful for reliable robust supervision, while extensive experiments show consistent improvements in tail and worst-class performance, balanced robustness, and posterior-support stability.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Long-tailed adversarial robustness, Adversarial training, Posterior support reliability, Free-energy modeling, Posterior transport, Class-imbalanced learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_x2rROpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies adversarial training under long-tailed distributions and argues that tail-class robustness degrades because posterior supervision becomes unreliable in unstable regions of representation space. The proposed method, FEPT, reweights an EMA teacher posterior using a score that combines support distance, local density, and posterior sensitivity, and supplements tail classes with support transferred from related head classes. Experiments on CIFAR10-LT, CIFAR100-LT, and TinyImageNet-LT report improved tail and worst-class robustness over existing methods.

Critical strengths

None

Critical weaknesses

W1. Motivation is not established. The paper claims that posterior support unreliability causes tail robustness degradation but never demonstrates it before introducing the method. The diagnostics that appear later (Fig. 2, Table 10) measure the same quantities FEPT optimizes, so they cannot show that this phenomenon causes the problem. Beyond performance numbers, the paper also gives little insight into how the method works: it never shows which samples are down-weighted, what is transferred from head classes, or which failures of prior methods FEPT fixes.

W2. Terminology obscures a simple method. The method combines a nearest-neighbor distance under a local-PCA-based anisotropic metric, a kernel density estimate, and the norm of the posterior gradient into a reweighting score, then mixes in prototype-shifted support from head classes. Terms such as free energy, Riemannian metric, and transport add nothing beyond these operations and make the method harder to follow. The paper also does not explain why these particular factors are needed or why they are combined in this form, so the score reads as a combination of heuristics rather than a justified design.

W3. The role of the theory is unclear. Theorem 1 is a standard Lipschitz margin argument whose tail-specific conclusion follows directly from the assumed form of L_y, and Theorem 2 is a generic bias–variance decomposition that holds for any weighted average. It is unclear what these results contribute to the paper. What specific insight about FEPT do the authors intend them to provide? If they are meant only as intuition, it would be better to state that intuition briefly or remove the theorems altogether. (In my view, formalizing an intuition is not the same as theoretical analysis. A theoretical result is valuable when it provides guidance that the intuition alone does not, which is not the case here.)

W4. Baseline results are inconsistent with prior work. The PAT entries, including standard deviations, are copied directly from the original PAT paper, even though the experimental settings of the two papers appear to differ in several respects. This contradicts the statement that all results follow the same protocol. Moreover, all other baselines are substantially higher than the values reported in the PAT paper. For example, AT-BSL PGD accuracy with ResNet-18 is 23.84 vs. 10.98 on TinyImageNet-LT. As a result, PAT appears much weaker than it actually is.

W5. A possible evaluation flaw may explain the discrepancy in W4. Mixing model.train() and model.eval() is a common mistake in adversarial training, so I raise this as a possible concern. In the supplementary code, pgd() calls model.train() before returning. As a result, adversarial examples are generated in eval mode but classified in train mode, and test data updates the BatchNorm statistics saved in the checkpoint. If this code produced the reported results, the PGD numbers for FEPT and the re-run baselines may be inflated, while the copied PAT numbers are not. This would account for why only PAT appears much weaker than in prior work.

Questions for authors

None

Reviewer_7bKAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates adversarial robustness under long-tailed distributions, identifying the unreliability of posterior support as a potential cause for the diminished robustness of tail-class samples. To address this problem, the authors propose Free Energy-guided Posterior Transport (FEPT). FEPT evaluates the reliability of posterior distributions using semantic support distance, local density, and posterior sensitivity. It then employs a prototype-alignment-based transport to transfer reliable support information from head-to-tail classes. The paper provides a theoretical analysis of local posterior stability and the bias-variance trade-off inherent in the posterior transport process. Experiments on CIFAR10-LT, CIFAR100-LT, and TinyImageNet-LT with different architectures report improvements in tail-class and worst-class robustness under multiple adversarial attacks.

Critical strengths

An insight into perspective on long-tailed adversarial robustness. The paper identifies posterior support unreliability as an underexplored issue in long-tailed adversarial training, highlighting that improving point-wise posterior predictions alone may not be sufficient, as they can still be unreliable in the neighborhoods of tail-class samples.

Critical weaknesses

Inconsistencies in the reported experimental results. There appear to be some inconsistencies in the results presented in Table 1. For instance, on the CIFAR10-LT dataset, FEPT-EMA achieves an AA All of 40.18% and an AA Tail of 34.85%. This implies that the average accuracy for the remaining classes should be 61.50%. However, this figure does not seem to align with the head-class accuracy (43.22%) or the Worst-AA (26.18%) reported in the paper. The authors should clarify the specific methods used to calculate these metrics and provide detailed per-class results, as such data are essential for supporting the paper's core arguments.

Questions for authors
  1. Could the authors clarify the category grouping method used to calculate the head-tail gap in Table 1? Additionally, could the AutoAttack accuracy for each category on the CIFAR10-LT dataset be provided? This is because some of the reported results appear inconsistent with the evaluation protocol stated in the paper.
  2. Theorem 1 focuses on the stability of the posterior margin, while Equation 4 concerns the posterior sensitivity for each class. Could the authors clarify how reducing posterior sensitivity contributes to enhancing the margin stability required by Theorem 1?
  3. Does the memory bank exclude the current sample when calculating the semantic support distance? Since Algorithm 1 updates the memory bank before calculating the free energy, does including the current sample cause the self-distance to become zero, affecting our estimate of the support?
  4. How is the posterior probability of being at the prototype-translated feature locations computed in Eq. 7? Are the shifted features directly fed into the classifier head of EMA Teacher? If so, how to deal with the problem of consistency between the Student model and EMA feature Spaces?
Reviewer_PH9WOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The work addresses the degradation that tail classes suffer when training adversarial model under class imbalances. The author propose to use a class-imbalance aware reliability measure to reweight the posterior distribution. Since the estimation of this statistical quantities can be unreliable in underrepresented classes, the authors also propose head-tail information transfer, combining information from better represented classes.

Critical strengths

S1) The proposed energy-free term is a potentially useful and novel heuristic. Although each individual component are not novel, the combination of support distance, local density, and posterior sensitivity a a single reliability score is intereseting. Similarly, borrowing information frome head classes to better estimate lower classes was also previosuly proposed, but incorporating this idea in posterior-support transport for adversarial training is a useful contribution.

S2) The experimental results are meaningful. The method shows consistent improvements of around 1–5 pp in tail-class robustness compared to recent baselines, across multiple datasets. In particular, it shows improvment in the under represented classes which is particularly important.

Critical weaknesses

W1) The provided code raises significant concerns about the implementation of the method and the reproducibility of the results. The experiments show that the method achieves uniformly better results over only a few pp over the baselines.

Although this few pp improvements is meaningful, specially if is uniform across datasets and attacks, they can be also be potentially fruit from an extensive hyperparameter search for the proposed method, without providing the same tuning budget to the baselines. This is particularly concerning as the provided code selects the best checkpoin using a subset of the test set rather than a validation set. This is why being able to reproduce Tables 1 and 2 is important.

However, not only the experiments are not reproducable, but the provided code implements a very different method from the one described in the paper. The fept/loss.py (lines 21–36) implements class reweighting, while lines 42–68 implement logit adjustment and a KL consistency loss between clean and adversarial predictions, using a truncated softmax over selected classes. The adversarial attack in fept/attack.py (lines 4–16) is standard PGD. I couldn't find any implementation of the proposed geometric memory banks, free-energy score, head-to-tail posterior transport, EMA teacher, etc.

W2) The contribution #1 (lines 82–83), which is "identifying posterior-support unreliability as a key factor in long-tailed adversarial robustness", is not established neither empirically nor theoretically. The authors introduces a heuristic which intends to correct unreliable posterior support. However, this does not show that posterior-support unreliability is a significant cause of the performance degradation, or that it occurs more often in underrepresented classes.

I believe the work would be much stronger if the first part focused on validating this hypothesis. For example, showing that for standard training methods many robustnes degradation indeed ocur on high-confidence posterior with unreliable and poorly supported neighborhoods, and that this happens more often for tail underrepresented classes. The second part could then propose the heuristic as a solution, but clearly separating the identification of the issue from a proposed solution.

W3) The theoretical part seems correct but incomplete as a justification of the proposed heuristic. In particular, considering contribution #3 (line 86–88):

“We theoretically justify that robust supervision benefits from local posterior-support stability.”

Although true, posterior stability is only 1 of the 3 components of the free-energy term. Moreover, the claim that the Lipschitz constant grows for tail classes is simply assumed. The authors justify this scaling by referring to empirical observations of increased posterior volatility, but I couldn't find a direct empirical validation of this scaling in the work or as a reference.

“Posterior transport reduces tail-support estimation variance with controllable bias under source alignment.”

I think this is an overstatement. The theorem shows a general bia vs variance tradeoff when combining the tail and head class estimators, but it does not guarantee a reduction in estimation error. In particular, it does not establish that the proposed source selection and weighting control the transport bias. These properties are assumed instead of justified.

Questions for authors

My two main concerns are the W1 and W2, and thus I have the following 2 questions/requests.

Q1) Can the authors explain why the implementation in the code from the supplementary material is different from the method in the work? Were the experiments in the main paper run using this implementation? If not, could the authors provide the actual implementation and evaluation scripts used to reproduce Tables 1 and 2?

Q2) Can the authors provide evidence that posterior-support unreliability is indeed a key factor of adversarial robustness degradation? For example, showing that it often happends that the model gives high confidence predictions in poorly supported neighborhoods, and that when this happens there is a robustness degradation (i.e., the predictions are unstable or become incorrect under small perturbations). Also, can the authors show that this failures occur more often for tail classes and results in their lower robustness?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.