- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper investigates targeted poisoning attacks against TTA, where poisoned samples in a test batch can manipulate model predictions on benign co-batched samples. The authors propose PCAC, a poisoning framework that eliminates the need for target-side access during attack crafting by leveraging distribution-shifted auxiliary data. Specifically, PCAC enhances proxy diversity through salient-region augmentation and improves transferability through target-confident effect induction and smooth surrogate-side optimization. Experiments across diverse TTA settings demonstrate that PCAC achieves much higher ASR of prior methods, highlighting its effectiveness in black-box poisoning scenarios.
Critical strengths
- The paper investigates targeted poisoning attacks against TTA, where poisoned sample. Compared with existing works, it further considers a more challenging black-box setting without access to the target model or deployed TTA algorithms during attack construction.
- PCAC combines salient-region augmentation, target-confident effect induction, and smooth optimization to enhance proxy diversity and attack transferability, offering a novel framework to black-box TTA poisoning.
- Experiments across diverse TTA settings demonstrate that PCAC achieves 2.47× the attack success rate of prior methods, providing empirical evidence of its effectiveness under limited-access conditions.
Critical weaknesses
- The authors introduce an additional random perturbation variable when formulating targeted poisoning attacks, but its purpose and significance remain unclear. Moreover, is neither incorporated into the proxy construction module nor explicitly discussed in the adversarial crafting module. Consequently, it is unclear how this perturbation contributes to the proposed method and affects the final results.
- The paper lacks sufficient justification for the design choices and the underlying methodological rationale. For example, in the proxy construction module, why are both Grad-CAM heatmaps and entropy confidence maps needed? What motivates the use of masks to suppress semantically less salient image regions? Similarly, why is enhancing flatness beneficial in the adversarial crafting module? The authors mainly describe the proposed components and overall framework without adequately explaining the motivations behind these choices, which weakens the overall persuasiveness of the paper.
- Existing TTA methods encompass diverse paradigms, including the use of additional adapter modules, fine-tuning with pseudo-labels, and prediction logit correction based on distributional information. It remains unclear how the proposed attack framework can generalize to these diverse TTA approaches. Furthermore, the experimental evaluation should include representative baselines from different TTA paradigms to provide a more comprehensive assessment.
- The authors should extend the experimental evaluation to include additional OOD benchmarks, such as ImageNet-A and ImageNet-V2, as TTA methods are commonly applied in such scenarios. The authors should also evaluate the proposed framework using auxiliary datasets that exhibit substantial distribution shifts relative to both the test samples and the source dataset, thereby better assessing its effectiveness under more challenging distributional conditions.
- The paper uses overly complex mathematical formulations and notations, many of which lack equation labels and numbers. This makes cross-referencing inconvenient and reduces the overall readability of the manuscript. The authors should simplify them.
Questions for authors
See the weaknesses above.
8 comments
To Reviewer odqk (Part II)
W2. Methodological Rationale (continued)
(2) Rationale for Masking. The masks translate the joint saliency responses into spatially selective enhancement, retaining regions supported by class-discriminative evidence and prediction confidence , while attenuating weaker responses. Since different regions contribute unequally to these characteristics, retaining all regions may dilute the two predictive cues. We therefore use joint-saliency masks to preserve regions supported by both characteristics while suppressing weaker responses. Moreover, multiple masking levels provide diverse proxy contexts with varying degrees of saliency emphasis, further enabling to optimize perturbations across diverse proxy contexts.
(3) Reason for enhancing flatness beneficial in the adversarial crafting module. Since the optimization of is conducted using constructed proxy co-batches in place of the unavailable real benign co-batches , and a surrogate model in place of the unavailable target model , the reason enhancing flatness is beneficial is that it reduces the sensitivity of optimization to the proxy-benign context and surrogate-target state discrepancies. Without flatness enhancement, these discrepancies cause optimization at a single perturbation point and under a single surrogate state to produce overly specific updates that are sensitive to these variations, thereby limiting their effectiveness when is co-batched with the actual benign inputs during target-model adaptation under .
As analyzed in our theoretical decomposition (Section 3.2, pp. 4–5; Appendix E, pp. 21–22), the crafting gap arising from proxy-context and surrogate-model discrepancies is upper-bounded by the following terms:
Here, characterizes local perturbation sensitivity, captures surrogate-state sensitivity, and accounts for the remaining inaccessible target-side discrepancy. PCAC’s flatness-enhanced crafting update (FCU) addresses the two accessible sensitivities through perturbation-neighborhood averaging and surrogate-parameter shifting, respectively, aiming to reduce dependence on specific proxy–surrogate optimization conditions. Consequently, perturbations optimized under the proxy co-batch and surrogate model are less dependent on specific optimization conditions, improving ’s effectiveness when transferred to the actual benign co-batch under the target model .
W3. Mechanistic Explanation and Evaluation Across Diverse TTA Paradigms
Consistent with prior TTA poisoning studies (e.g., DIA (Wu et al., ICML 2023), TePA (Cong et al., IEEE S&P 2024), and FCA (Rifat et al., ICLR 2025)), PCAC exploits the batch-dependent nature of TTA, in which adaptation signals or prediction refinements are derived from co-batched samples, allowing attacker-controlled inputs to influence the underlying adaptation and prediction process. By adding perturbations to attacker-controlled inputs , an attacker can optimize them to shape the data-characteristic of poisoned inputs to induce its desired target, thereby affecting the resulting adaptation states or batch-level predictions and causing target-class misclassifications on co-batched benign inputs .
To Reviewer odqk (Part III)
W3. Mechanistic Explanation and Evaluation Across Diverse TTA Paradigms (continued)
Our original evaluation already covers diverse TTA paradigms with different mechanisms, including normalization-based adaptation (TeBN), entropy-based parameter updates (TENT, EATA, SAR), pseudo-label-based updates (Hard PL, Soft PL, RPL), and feature alignment (TTAC) (Table 1, p. 7). We additionally include two representative methods covering further TTA paradigms: TAST (Jang et al., ICLR 2023), which updates trainable adapter modules using nearest-neighbor-based pseudo-labels, and LAME (Boudiaf et al., CVPR 2022), which performs parameter-free prediction correction through cross-sample affinities. Both methods are evaluated under the same default CIFAR100-C poisoning setting, with ASR (%) results reported below.
| TTA Method / Attack | No-Poison | White-Box | TargetCE | DIA | TePA | FCA | BLE | NHE | Our PCAC |
|---|---|---|---|---|---|---|---|---|---|
| TAST | 0.94$_{\pm 0.07}$ | 82.91$_{\pm 1.21}$ | 33.74$_{\pm 1.36}$ | 45.93$_{\pm 1.10}$ | 50.12$_{\pm 1.33}$ | 51.06$_{\pm 1.35}$ | 55.34$_{\pm 1.24}$ | 57.15$_{\pm 1.45}$ | 64.36$_{\pm 1.19}$ |
| LAME | 1.06$_{\pm 0.09}$ | 69.64$_{\pm 1.48}$ | 21.61$_{\pm 0.83}$ | 29.88$_{\pm 1.26}$ | 35.26$_{\pm 1.31}$ | 33.10$_{\pm 1.42}$ | 41.02$_{\pm 1.17}$ | 38.54$_{\pm 1.29}$ | 50.75$_{\pm 1.40}$ |
The results show that PCAC remains effective across these additional TTA, achieving ASRs of 64.36% and 50.75% on TAST and LAME, respectively.
W4. Additional OOD Benchmarks and Auxiliary Distribution Shifts.
Evaluation on OOD Benchmarks. We extend experiments to ImageNet-V2 (Recht et al., ICML 2019) and ImageNet-A (Hendrycks et al., CVPR 2021). ImageNet-V2 contains independently collected natural images, while ImageNet-A comprises naturally occurring challenging examples. We evaluate eight original TTA methods, together with TAST and LAME, on both benchmarks and retain VGG-19-BN surrogate and DenseNet-121 target on default poisoning setting. For both benchmarks, we also employ 10 labeled auxiliary samples from ImageNet-Sketch, whose sketch-style visual distribution differs from both the natural-image ImageNet-1K source distribution and the natural-image target distributions of ImageNet-V2 and ImageNet-A.
ImageNet-V2:
| Attack | TeBN | TENT | HardPL | SoftPL | RPL | EATA | TTAC | SAR | TAST | LAME |
|---|---|---|---|---|---|---|---|---|---|---|
| No-Poison | 0.09$_{\pm 0.02}$ | 0.08$_{\pm 0.02}$ | 0.10$_{\pm 0.03}$ | 0.11$_{\pm 0.02}$ | 0.07$_{\pm 0.01}$ | 0.08$_{\pm 0.02}$ | 0.12$_{\pm 0.03}$ | 0.09$_{\pm 0.02}$ | 0.12$_{\pm 0.03}$ | 0.13$_{\pm 0.03}$ |
| White-Box | 91.46$_{\pm 0.75}$ | 90.82$_{\pm 0.93}$ | 89.37$_{\pm 0.89}$ | 85.62$_{\pm 0.98}$ | 90.15$_{\pm 0.82}$ | 89.42$_{\pm 1.14}$ | 82.39$_{\pm 1.26}$ | 85.08$_{\pm 1.08}$ | 79.67$_{\pm 1.35}$ | 67.48$_{\pm 1.48}$ |
| TargetCE | 36.85$_{\pm 1.28}$ | 40.21$_{\pm 1.15}$ | 34.62$_{\pm 1.23}$ | 30.76$_{\pm 1.40}$ | 29.38$_{\pm 1.53}$ | 41.49$_{\pm 1.05}$ | 33.17$_{\pm 1.71}$ | 39.42$_{\pm 1.62}$ | 29.48$_{\pm 1.34}$ | 22.17$_{\pm 0.95}$ |
| DIA | 55.31$_{\pm 1.43}$ | 58.14$_{\pm 1.20}$ | 50.83$_{\pm 1.36}$ | 43.95$_{\pm 1.62}$ | 48.56$_{\pm 1.21}$ | 59.42$_{\pm 1.37}$ | 44.18$_{\pm 1.76}$ | 52.39$_{\pm 1.59}$ | 42.57$_{\pm 1.42}$ | 30.48$_{\pm 1.28}$ |
| TePA | 59.85$_{\pm 1.44}$ | 60.92$_{\pm 1.17}$ | 57.73$_{\pm 1.52}$ | 55.14$_{\pm 1.41}$ | 49.86$_{\pm 1.65}$ | 58.46$_{\pm 1.28}$ | 46.09$_{\pm 1.75}$ | 54.15$_{\pm 1.53}$ | 47.83$_{\pm 1.63}$ | 34.69$_{\pm 1.27}$ |
| FCA | 61.04$_{\pm 1.33}$ | 55.73$_{\pm 1.61}$ | 58.62$_{\pm 1.21}$ | 49.77$_{\pm 1.38}$ | 46.42$_{\pm 1.59}$ | 61.30$_{\pm 1.46}$ | 50.74$_{\pm 1.68}$ | 56.32$_{\pm 1.39}$ | 48.41$_{\pm 1.45}$ | 36.82$_{\pm 1.22}$ |
| BLE | 69.94$_{\pm 1.23}$ | 71.65$_{\pm 1.41}$ | 67.18$_{\pm 1.26}$ | 63.29$_{\pm 1.77}$ | 64.52$_{\pm 1.35}$ | 66.78$_{\pm 1.43}$ | 53.64$_{\pm 1.82}$ | 59.47$_{\pm 1.59}$ | 55.83$_{\pm 1.14}$ | 40.54$_{\pm 1.40}$ |
| NHE | 71.43$_{\pm 1.37}$ | 69.84$_{\pm 1.19}$ | 68.61$_{\pm 1.62}$ | 60.94$_{\pm 1.21}$ | 67.35$_{\pm 1.55}$ | 66.14$_{\pm 1.71}$ | 57.28$_{\pm 1.52}$ | 60.97$_{\pm 1.45}$ | 54.67$_{\pm 1.69}$ | 42.19$_{\pm 1.48}$ |
| Our PCAC | 77.86$_{\pm 1.10}$ | 76.92$_{\pm 1.29}$ | 75.61$_{\pm 1.36}$ | 72.90$_{\pm 1.18}$ | 73.35$_{\pm 1.34}$ | 75.84$_{\pm 1.20}$ | 67.34$_{\pm 1.50}$ | 70.15$_{\pm 1.38}$ | 63.28$_{\pm 1.14}$ | 50.62$_{\pm 1.35}$ |
To Reviewer odqk (Part I)
Dear Reviewer odqk,
Thank you for your detailed review! We provide more detailed clarifications below. Please let us know if anything is still unclear and we'd be happy to follow up.
W1. Role and Effect of δ
Consistent with prior poisoning formulations (Wu et al., ICML 2023; Cong et al., IEEE S&P 2024), the perturbations δ={δi} are introduced and applied to the attacker-controlled inputs Ba, i.e., Baδ={xi′+δi∣xi′∈Ba}. Here, δ denotes the learnable adversarial perturbation optimized during crafting, rather than an independently sampled random variable. The purpose of these perturbations is to shape the attacker-controlled inputs Baδ to induce the attacker-desired target-class effect during optimization, such that, when co-batched with benign inputs Bb in the target model's batch-level adaptation process, Baδ can cause the adapted target model to misclassify benign co-batched inputs as the attacker-desired class.
The Proxy Construction (PC) module does not involve the optimization of δ; instead, it constructs an influential proxy co-batch set S for the Adversarial Crafting (AC) module since the real benign co-batch Bb is unavailable. Adversarial Crafting (AC) module then explicitly optimizes δ under the mixed batch Baδ∪S. Specifically, AC iteratively updates δ by jointly optimizing the target-confidence term, which encourages target-class predictions on the perturbed attacker-controlled inputs Baδ, and the attacker-to-proxy guidance term, which uses the δ-dependent batch-averaged predictions of Baδ to induce target-oriented proxy predictions despite S’s emphasized saliencies thereby enhancing the intended adversarial effects when Baδ is co-batched with unavailable benign samples Bb during target-model adaptation. Both terms depend on δ through Baδ (Section 3.2, pp. 4–5; Algorithm 1, p. 5)
Consequently, δ contributes to PCAC by encouraging target-confident predictions on Baδ, thereby enabling the poisoned inputs to influence batch-level adaptation and bias the adapted target model’s predictions on benign co-batched inputs Bb toward the attacker-desired class.
W2. Methodological Rationale
(1) Rationale for Combining Grad-CAM and Entropy-Confidence Maps. Since the real benign co-batch Bb is unavailable during crafting, our Proxy Construction (PC) module constructs influential proxy co-batches S by saliency-enhancing auxiliary samples, allowing our Adversarial Crafting (AC) module to optimize δ on Ba against these proxies. This encourages Baδ to induce target-oriented proxy predictions despite their emphasized saliencies, aiming to strengthen its adversarial effects during target-side adaptation.
The reason for using Grad-CAM and entropy-confidence maps is that they can provide both class-discriminative evidence Acls and entropy-based prediction confidence Cent, which serve as saliency-guided enhancement cues for auxiliary samples. Grad-CAM maps highlight spatial regions contributing to the ground-truth class logit, while entropy-confidence maps highlight regions contributing to higher prediction confidence (lower entropy). Their normalized element-wise multiplication emphasizes regions supported by both characteristics. This design is consistent with prior TTA studies highlighting sample-wise reliability and individual sample influence on adaptation (Niu et al., ICML 2022; Niu et al., ICLR 2023; Li et al., ICML 2026), where feature activations contribute to batch-conditioned statistics (Nado et al., ICML 2020 Workshop), while predictive distributions provide signals for entropy minimization (Wang et al., ICLR 2021; Zhang et al., NeurIPS 2022) and pseudo-label updates (Goyal et al., NeurIPS 2022).
Moreover, we conducted a preliminary analysis under normal TTA to characterize the adopted two characteristics: class-discriminative evidence Acls and entropy-based prediction confidence Cent. Specifically, using DenseNet-121 on CIFAR10-C, CIFAR100-C, and ImageNet-C at corruption severity 3 across eight TTA methods, we compared initially correctly classified samples that retained correct predictions after adaptation with those that became misclassified. Samples retaining correct predictions tended to exhibit higher scores on both Acls and Cent (average pairwise ranking probabilities: 0.67 and 0.74, respectively, compared with 0.56 for maximum softmax probability and 0.51 for top-2 probability margin). These findings further support the need of both Grad-CAM and entropy-confidence as complementary cues for constructing influential proxy co-batch contexts.