- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper investigates the transferability of untargeted adversarial attacks against skeleton-based human action recognition models. The authors argue that existing attack methods mainly focus on suppressing the ground-truth class while insufficiently exploiting information from non-ground-truth classes. Motivated by the observed cross-model consistency of high-confidence negative classes, the paper proposes Dynamic Negative Target Guidance (DNTG), a plug-and-play framework designed to improve black-box attack transferability.
The main contribution is a negative-class guidance mechanism that dynamically selects the Top-K high-confidence negative classes, assigns soft guidance weights through temperature-scaled softmax, and periodically updates the selected targets during optimization. This guidance is incorporated into existing attack objectives without requiring model retraining.
Critical strengths
-
A simple intervention with improvements across a reasonably broad set of tests.DNTG is tested with six attack baselines, two surrogate choices, and three benchmarks. For example, with CTR-GCN as surrogate on NTU120, DIM improves from 27.41% to 35.15% mean ASR (Table 1). Other gains are smaller, but their breadth makes the method potentially useful as an addition to existing robustness evaluations. The objective-level integration is also straightforward.
-
An empirical motivation that is connected to the design.Figure 2 reports both initial negative-rank concentration and cross-model top-10 overlap. Tables 3 and 4 further examine whether refreshing the negative set helps and whether the selected classes change over the trajectory. These analyses provide a more informative motivation than reporting transfer scores alone, even though the causal interpretation still needs stronger controls.
Critical weaknesses
<1>Although the paper reports improvements in attack transferability, it does not sufficiently establish the added value of its proposed negative-target guidance mechanism over closely related loss formulations. When K=1, the DNTG objective reduces to a selected-negative logit margin. Moreover, with detached soft weights and a fixed negative set, its local gradient matches that of a temperature-scaled Log-Sum-Exp margin. However, the paper lacks direct comparisons with these alternatives. Table 3 also does not separately evaluate confidence-based versus random negative selection or temperature-scaled versus uniform weighting. Consequently, it remains unclear whether the improvements arise from the distinctive design of DNTG or simply from introducing an additional margin-based objective. This uncertainty directly affects the assessment of the paper's main methodological contribution.
<2>The paper does not clarify whether gradients propagate through the soft target weights in Eqs. (11)–(15), which affects the actual optimization objective. Important evaluation details, including sample-selection rules, clean model accuracies, and the precise ASR definition, are also missing. Furthermore, Tables 3–4 and Figure 4 do not clearly specify their experimental settings, and no uncertainty estimates are provided for the reported gains. This is particularly concerning for improvements that are relatively small. Although these omissions do not invalidate the results, they make it difficult to assess their reproducibility, fairness, and practical significance. As a result, the current experimental evidence is not sufficiently convincing to support acceptance.
Questions for authors
<1>Regarding gradient computation and implementation, in Eqs. (11)–(15), is a stop-gradient operation applied to the soft target weight q_t before computing gradients? Please clarify the actual gradient computation, and how, for each baseline (e.g., the momentum in MI-FGSM, the input transformation in DIM, and the specific design of TASAR), the additional objective is combined with the original loss, perturbation constraints, and update rule.
<2>Regarding data splits, sample selection, and ASR computation, please specify the split protocol for each dataset, the number of evaluation samples used, the sample-selection rule, and the exact criterion for computing the attack success rate.
<3>Regarding novelty and the independent contribution of key components, could you add loss-level controlled comparisons under matched conditions against a single highest-scoring negative class (K=1), random negative classes, Top-K uniform weighting, and a plain margin objective, in order to separate the class-selection mechanism, the soft-weighting mechanism, and a generic margin loss?
<4>Regarding the ablation configuration and hyperparameters, please clarify the settings used for Table 3, Table 4, and Figure 4, including the dataset, surrogate model, attack baseline, number of evaluation samples, and other hyperparameter values.