1. Limited methodological novelty and insufficient differentiation from prior work.
The central idea of exploiting BatchNorm statistics to disrupt benign predictions is closely related to existing studies on the vulnerability of normalization-based TTA, particularly MedBN [R1]. Feature-level poisoning has also been explored in FCA [R2], while the grey-box threat model and NHE objective are directly inherited from RTTDP [R3]. The main methodological contribution is therefore the specific construction of a standardized mean and log-variance displacement objective, rather than a fundamentally new poisoning framework. Although this objective is potentially useful, the paper does not sufficiently demonstrate why its particular formulation is superior to simpler alternatives, such as feature-distance maximization, mean-only or variance-only displacement, or unweighted moment matching. The current ablations primarily compare moment transport with NHE, leaving the necessity and distinct contribution of the proposed moment metric insufficiently established.
2. The theoretical guarantees rely on strong assumptions whose practical relevance is unclear.
Theorem 1 derives a source-to-adapted moment transfer guarantee under a Jacobian stability assumption and an additional lower-sensitivity condition requiring the source moment displacement to be bounded below by a positive constant times the input perturbation norm. However, the latter assumption can be difficult to satisfy in the high-dimensional perturbation space considered here. In particular, when the dimensionality of the selected moment statistics is smaller than that of the input perturbations, the moment Jacobian necessarily has a nontrivial nullspace, precluding such a uniform lower bound over a full neighborhood of perturbation directions. The theorem remains mathematically conditional, but the paper does not establish whether its assumptions hold for the actual attack space or adapted models. Several other theoretical results also follow from standard mixture identities, Taylor expansions, and smoothness arguments. Consequently, the theoretical analysis provides useful intuition but does not establish a sufficiently strong or practically verified guarantee of attack effectiveness.
3. The relationship between moment displacement and benign risk degradation is insufficiently validated.
A central motivation is that maximizing pre-BN moment displacement leads to distorted benign features and harmful adaptation updates. However, the proposed attack optimizes moment displacement rather than benign risk, and the theoretical analysis does not establish that increasing the former necessarily increases the latter. Theorem 3 decomposes update displacement into a moment-representable component and a residual, but the relative contribution of these components is not empirically quantified. More importantly, Theorem 4 requires harmful alignment between the adaptation displacement and the benign-risk gradient, whereas the experiments in Appendix Tables 14–15 measure alignment between the moment and NHE input gradients instead. These are different quantities. The reported results also show that larger moment displacement does not consistently correspond to greater held-out loss. Thus, the experiments do not directly validate the key theoretical conditions connecting the proposed objective to attack success. More explicit measurements of source-to-target moment transfer, update decomposition, and risk-gradient alignment would substantially strengthen the claimed mechanism.
4. Incomplete comparisons with relevant attack and defense approaches.
The experimental evaluation mainly compares MTP against DIA and two RTTDP variants (NHE and NHE+DR). However, the paper does not include direct comparisons with feature-based poisoning methods such as FCA [R2], nor does it sufficiently investigate simpler feature-statistics objectives that would help establish the contribution of the proposed formulation. Although differences in threat models may prevent a direct comparison with the original FCA protocol, matched or adapted baselines would be informative. In addition, the experiments primarily target BN-dependent adaptation methods, while robustness against stronger normalization defenses such as MedBN [R1] is not examined. This limits the evidence supporting the novelty, comparative effectiveness, and general applicability of MTP beyond the specific normalization mechanisms it targets.
5. Limited and statistically uncertain improvements under constrained attack budgets.
The strongest gains are obtained under the default setting with a relatively large perturbation budget (epsilon = 0.3) and a poison ratio of 0.5. Under the smaller budget of epsilon = 8/255, the five-seed improvements over the strongest baseline decrease to only 1.21 and 0.90 percentage points on CIFAR-10-C and CIFAR-100-C, respectively. On CIFAR-10-C, MTP does not outperform DIA in every seed, and the paired differences provide limited evidence of a robust advantage with only five runs. The CIFAR-100-C results also exhibit substantial seed variability relative to the reported improvement. Furthermore, the paper does not evaluate the combination of a small perturbation budget and a low poison ratio. These limitations make it difficult to assess whether the proposed attack provides a practically meaningful and consistently reproducible advantage under more restrictive threat conditions.
6. Numerical consistency and experimental reliability require clarification.
There appears to be a potentially important inconsistency between the three-seed and five-seed CIFAR-100-C results at epsilon = 0.3. Table 1 reports 49.85% benign error for MTP over three seeds, while Table 2 reports 47.12 ± 2.72% over five seeds. If the original three seeds are included in the five-seed evaluation under the same aggregation protocol, the remaining two seeds would need an average of approximately 43.03%, implying a minimum possible five-seed sample standard deviation of approximately 3.74%, which exceeds the reported 2.72%. These results therefore cannot be reconciled under that interpretation. The discrepancy may reflect different seeds, aggregation procedures, or reporting errors, but the paper does not clearly explain it. Since the main contribution relies heavily on quantitative improvements, the authors should clarify the evaluation protocol and provide the corresponding per-seed results.
[References]
[R1] "MedBN: Robust Test-Time Adaptation against Malicious Test Samples," CVPR, 2024.
[R2] "On the Adversarial Vulnerability of Label-Free Test-Time Adaptation," ICLR, 2025.
[R3] "On the Adversarial Risk of Test Time Adaptation: An Investigation into Realistic Test-Time Data Poisoning," ICLR, 2025.
I also have some minor concerns as follows:
-
The moment-only objective has zero gradient at the clean input, making initialization important. The paper should explicitly describe the initialization protocol and its sensitivity.
-
The effect of the selected BN layers is not sufficiently analyzed. A layer-selection ablation would help determine whether the method depends on particular normalization layers.
-
The evaluation on ImageNet-C covers only four corruption subsets, limiting conclusions about large-scale generalization.
-
The default comparison uses 10 attack steps for MTP and 40 for the baselines. Although additional step-matched results are provided, comprehensive comparisons under matched computational budgets would be more informative.
-
The reported CPU efficiency results are limited to selected corruption settings and do not fully characterize the computational overhead of computing moment statistics and their gradients.
-
The recovery experiments demonstrate persistent degradation over a limited horizon, but do not establish how long the effect remains under continued benign-only adaptation.
-
The confidence intervals based on grouped corruption and seed results may not fully account for dependence among evaluations sharing the same corruption or adaptation trajectory.