acceptodds
Under review as a conference paper at ICLR 2027

C2F-DiffAttack: Coarse-to-fine diffusion for local black-box attack on time series classifiers

Abstract

Time series classifiers remain vulnerable to adversarial perturbations that can alter model predictions while preserving much of the original temporal structure. Local black-box attacks must determine both where to perturb and what perturbation to generate, while restricting perturbations to a small subset of timestamps. Existing methods often separate temporal support selection from perturbation generation, despite both decisions depending on broader temporal context. We propose C2FDA, the first diffusion-based framework for local black-box adversarial attacks on time series classifiers. C2FDA couples the where and what decisions through shared coarse-to-fine temporal representations. Coarse attention captures region-level relevance, while fine attention refines it into timestamp-level evidence. For where, region-level and timestamp-level evidences are combined to determine an exact-budget Top- temporal support. For what, coarse relevance provides regional context for fine refinement, and the resulting timestamp-level evidence modulates diffusion features before noise prediction to shape perturbation content. A shared hybrid boundary objective encourages boundary crossing on the final masked perturbation, jointly optimizing temporal support and perturbation content. C2FDA performs black-box attacks through cross-model transfer. Experiments across diverse time series datasets and classifier architectures show state-of-the-art attack success rates while maintaining low perturbation distortion and high temporal similarity.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Time Series Classification, Adversarial Attacks, Black-Box Attacks, Local Perturbations, Diffusion Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_kS1vOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes C2FDA, a diffusion-based framework for local black-box adversarial attacks on time series classifiers. The method employs a coarse-to-fine attention mechanism to jointly determine perturbation locations and generate perturbation content, together with a hybrid boundary objective to improve attack effectiveness. Experiments on demonstrate promising attack success rates with limited temporal coverage and low perturbation distortion.

Critical strengths
  1. The paper addresses the problem by jointly considering perturbation locations and content in local time series attacks.
  2. The proposed coarse to fine attention mechanism provides a coherent framework for coordinating temporal localization and perturbation generation.
  3. Extensive experiments across 14 datasets and five target models demonstrate strong attack performance with limited temporal coverage.
Critical weaknesses
  1. The generalizability of the attack remains uncertain. The evaluation relies on a single surrogate model and relatively lightweight target classifiers, most of which are based on conventional architectures. The lack of experiments on modern time series classification models, such as Transformer-based models, limits the evidence for the broader applicability of the proposed method.
  2. Insufficient validation of the joint where–what learning. A central claim of the paper is that jointly optimizing perturbation locations and content is more effective than treating them separately. However, the ablation studies mainly evaluate individual attention components and do not directly compare joint optimization with a decoupled alternative using the same backbone. Such a comparison would better substantiate the claimed advantage of the proposed coupling mechanism.
  3. The overall presentation needs further refinement. The writing would benefit from additional language polishing. Figures 1 and 2 could be improved in terms of visual clarity and consistency, while Tables 2, 3, and 4 are overly dense and could benefit from better formatting and readability.
Questions for authors

See Weaknesses

9 comments

We thank the reviewer for the constructive review. Each point below opens with a one-sentence answer and then follows the reviewer's sub-points in order. Table and figure numbers refer to the revised PDF, with submitted numbers in parentheses where they differ. In the revised PDF, revisions for this review are in orange and revisions shared with other reviews are in purple.

WeaknessAnswer
W1Against a new held-out Transformer target, C2FDA has the highest ASR of 13 attacks on all three datasets (62.5–72.5%); with CNN as target, it also has higher 14-dataset average ASR and lower average MSE than SGAN for each surrogate.
W2Two same-backbone controls, one without WHERE→WHAT coupling (Simple Joint) and one trained WHERE-then-WHAT (Sequential), trail the full model by up to 22.2 and 33.3 ASR points, with higher MSE and DTW on every dataset.
W3We polished the language, redrew Figures 1–2, and reformatted Tables 2–6, while preserving the reported values for unchanged comparisons.

W1. "The evaluation relies on a single surrogate model and relatively lightweight target classifiers, most of which are based on conventional architectures."

Against a held-out Transformer target [1], C2FDA has the highest ASR of 13 attacks on all three multivariate datasets. Across three surrogate architectures with CNN as target, C2FDA also has higher 14-dataset average ASR and lower average MSE than SGAN for each surrogate.

Transformer target. For the three multivariate datasets, we evaluate transfer to a held-out Transformer target [1]. C2FDA reaches 62.5%, 66.7%, and 72.5% ASR, versus 37.5% (SGAN), 50.0% (TSadv), and 57.5% (AutoZOOM) for the strongest other attack on each dataset, and has the lowest DTW on BasicMotions and HandMovementDirection (Table 2): (Details: Sec. 4.2, Table 2.)

DatasetMethodASR (%) ↑MSE ↓DTW ↓ACF ↓
BasicMotionsFGSM6.20.0900170.420.29
PGD0.0———
MI-FGSM6.20.0813163.390.30
NES0.0———
SPSA25.00.0530116.770.14
AutoZOOM31.20.032742.46‾\underline{42.46}0.07
GATN0.0———
TSadv31.20.1023112.140.26
AdvGAN0.0———
Diff-PGD6.20.3019318.980.50
GE-AdvGAN0.0———
SGAN37.5‾\underline{37.5}0.1244186.290.28
C2FDA62.50.0473‾\underline{0.0473}34.910.09‾\underline{0.09}
HandMovementDirectionFGSM0.0———
PGD0.0———
MI-FGSM0.0———
NES0.0———
SPSA5.60.0559‾\underline{0.0559}794.590.10
AutoZOOM11.10.0221383.630.03
GATN5.60.1416505.390.24
TSadv50.0‾\underline{50.0}0.0981110.93‾\underline{110.93}0.19
AdvGAN0.0———
Diff-PGD5.60.1836883.370.29
GE-AdvGAN11.10.1983804.590.31
SGAN22.20.1139322.350.20
C2FDA66.70.0676108.720.06‾\underline{0.06}
UWaveGestureLibraryFGSM6.20.0900239.430.28
PGD7.50.0474126.530.19
MI-FGSM11.20.0598129.490.23
NES8.80.0900264.270.31
SPSA42.50.0601162.260.15
AutoZOOM57.5‾\underline{57.5}0.0208‾\underline{0.0208}98.280.03
GATN2.50.1279337.860.36
TSadv6.20.003932.910.07‾\underline{0.07}
AdvGAN17.50.1544310.490.43
Diff-PGD16.20.1420323.470.39
GE-AdvGAN12.50.1291249.720.36
SGAN25.00.0778131.980.23
C2FDA72.50.087260.77‾\underline{60.77}0.16

Other targets and classes. We evaluate transfer to recurrent (LSTM, BiRNN) and convolutional (CNN, TCN, DynamicConv) targets across all 14 datasets, where C2FDA has the best average ASR rank (1.64), ahead of AutoZOOM (2.29), SGAN (5.25), and TSadv (7.86) (Fig. 3). On the same UWaveGestureLibrary class-7/8 test subset, C2FDA reaches 81.0% ASR when its generator excludes those classes (85.7% with all classes), whereas SGAN reaches 28.6% and 71.4%, respectively (Table 13). (Details: Sec. 4.2, Fig. 3; App. D, Table 13.)

References [1] Zerveas et al., A Transformer-based Framework for Multivariate Time Series Representation Learning, KDD 2021.

W1 (continued).

Surrogates. With CNN as the target, Table 6 compares C2FDA and SGAN over 14 datasets. C2FDA's average ASR / MSE is 73.79% / 0.0379 with MLP, 82.72% / 0.0334 with ResNet, and 86.85% / 0.0322 with a Transformer, versus SGAN's 46.2% / 0.0749, 53.7% / 0.0686, and 59.2% / 0.0757, respectively; C2FDA reaches 86.85% average ASR with the Transformer surrogate, compared with 82.72% with the default ResNet surrogate. (Details: Sec. 4.5, Table 6.)

DatasetMLP C2FDA (ASR (%) / MSE)MLP SGAN (ASR (%) / MSE)ResNet C2FDA (ASR (%) / MSE)ResNet SGAN (ASR (%) / MSE)Transformer C2FDA (ASR (%) / MSE)Transformer SGAN (ASR (%) / MSE)
BasicMotions43.8‾\underline{43.8} / 0.072450.0 / 0.1201‾\underline{0.1201}50.0‾\underline{50.0} / 0.060656.2 / 0.1045‾\underline{0.1045}68.8 / 0.057162.5‾\underline{62.5} / 0.1373‾\underline{0.1373}
HandMovementDirection77.8 / 0.041166.7‾\underline{66.7} / 0.1208‾\underline{0.1208}94.4 / 0.034277.8‾\underline{77.8} / 0.1123‾\underline{0.1123}88.9 / 0.033883.3‾\underline{83.3} / 0.1257‾\underline{0.1257}
UWaveGestureLibrary89.6 / 0.0936‾\underline{0.0936}64.9‾\underline{64.9} / 0.092697.4 / 0.084576.6‾\underline{76.6} / 0.0896‾\underline{0.0896}93.5 / 0.081383.1‾\underline{83.1} / 0.0872‾\underline{0.0872}
Car80.0 / 0.014860.0‾\underline{60.0} / 0.0327‾\underline{0.0327}86.7 / 0.012666.7‾\underline{66.7} / 0.0286‾\underline{0.0286}93.3 / 0.012473.3‾\underline{73.3} / 0.0261‾\underline{0.0261}
CinCECGTorso64.0‾\underline{64.0} / 0.012170.1 / 0.0425‾\underline{0.0425}72.3‾\underline{72.3} / 0.010479.9 / 0.0368‾\underline{0.0368}79.9‾\underline{79.9} / 0.009884.2 / 0.0349‾\underline{0.0349}
ECG500089.2 / 0.024366.5‾\underline{66.5} / 0.0831‾\underline{0.0831}100.0 / 0.022173.7‾\underline{73.7} / 0.0783‾\underline{0.0783}98.7 / 0.021978.1‾\underline{78.1} / 0.0747‾\underline{0.0747}
ECGFiveDays93.2 / 0.020729.4‾\underline{29.4} / 0.0221‾\underline{0.0221}100.0 / 0.0177‾\underline{0.0177}35.6‾\underline{35.6} / 0.017298.3 / 0.018140.1‾\underline{40.1} / 0.0204‾\underline{0.0204}
Earthquakes83.1 / 0.10310.0 / —94.8 / 0.09440.0 / —96.1 / 0.08960.0 / —
FordA79.3 / 0.015662.5‾\underline{62.5} / 0.1013‾\underline{0.1013}87.3 / 0.012771.6‾\underline{71.6} / 0.0935‾\underline{0.0935}92.5 / 0.011879.5‾\underline{79.5} / 0.1036‾\underline{0.1036}
FordB82.6 / 0.014276.4‾\underline{76.4} / 0.0751‾\underline{0.0751}90.3 / 0.012784.1‾\underline{84.1} / 0.0693‾\underline{0.0693}89.0 / 0.013988.9‾\underline{88.9} / 0.0677‾\underline{0.0677}
Trace62.5 / 0.010950.0‾\underline{50.0} / 0.0459‾\underline{0.0459}72.5 / 0.009260.0‾\underline{60.0} / 0.0417‾\underline{0.0417}82.5 / 0.008667.5‾\underline{67.5} / 0.0407‾\underline{0.0407}
ItalyPowerDemand78.4 / 0.044514.1‾\underline{14.1} / 0.0934‾\underline{0.0934}87.3 / 0.040320.2‾\underline{20.2} / 0.0787‾\underline{0.0787}91.1 / 0.038626.3‾\underline{26.3} / 0.1076‾\underline{0.1076}
SonyAIBORobotSurface162.9 / 0.028716.1‾\underline{16.1} / 0.0511‾\underline{0.0511}71.8 / 0.025022.6‾\underline{22.6} / 0.0625‾\underline{0.0625}78.2 / 0.023729.0‾\underline{29.0} / 0.0543‾\underline{0.0543}
SonyAIBORobotSurface246.7 / 0.034820.0‾\underline{20.0} / 0.0936‾\underline{0.0936}53.3 / 0.031926.7‾\underline{26.7} / 0.0787‾\underline{0.0787}65.1 / 0.030732.3‾\underline{32.3} / 0.1038‾\underline{0.1038}
Average (14 datasets)73.79 / 0.037946.2‾\underline{46.2} / 0.0749‾\underline{0.0749}82.72 / 0.033453.7‾\underline{53.7} / 0.0686‾\underline{0.0686}86.85 / 0.032259.2‾\underline{59.2} / 0.0757‾\underline{0.0757}

W2. "Insufficient validation of the joint where–what learning."

We added two decoupled alternatives with the same backbone and C2F modules, and the full C2FDA beats both on every dataset and metric: Simple Joint still updates the WHERE and WHAT heads simultaneously with the same Top-k mask and hybrid loss but removes the WHERE→WHAT conditioning, and Sequential keeps the full architecture but learns WHERE first, fixes the mask, and then learns WHAT.

Joint versus decoupled. Table 3 reports the CNN-target ablations:

DatasetMetricFull C2FDAw/o Coarsew/o Finew/o C2F AttentionRandom IntervalsExact Margin OnlySimple JointSequential
BasicMotionsASR(%)↑50.0‾\underline{50.0}50.0‾\underline{50.0}31.256.231.243.843.837.5
MSE↓0.0606‾\underline{0.0606}0.08970.05040.19150.09790.08720.08140.0768
DTW↓71.7497.8376.46‾\underline{76.46}239.27102.6194.3896.1789.21
HandMovementDirectionASR(%)↑94.477.866.788.9‾\underline{88.9}50.083.372.261.1
MSE↓0.0342‾\underline{0.0342}0.06080.03020.20460.05730.04150.04270.0684
DTW↓91.48‾\underline{91.48}103.7290.34504.85143.7297.63103.57116.92
UWaveGestureLibraryASR(%)↑97.489.6‾\underline{89.6}76.684.449.487.084.480.5
MSE↓0.0845‾\underline{0.0845}0.10240.08420.14900.09620.11830.09670.0971
DTW↓48.2157.6753.20‾\underline{53.20}248.4156.3861.8560.4263.84

Full C2FDA reaches 50.0%, 94.4%, and 97.4% ASR, versus 43.8%, 72.2%, and 84.4% for Simple Joint and 37.5%, 61.1%, and 80.5% for Sequential, with lower MSE and DTW in every case. Since each control changes one factor, both the coupling and joint optimization contribute; even without coupling, simultaneous optimization (Simple Joint) exceeds Sequential in ASR on all three datasets. (Details: Sec. 4.3 "Effect of joint where–what learning", Table 3.)

Attention ablations. As the reviewer notes, w/o Coarse, w/o Fine, and w/o C2F Attention evaluate components. Removing C2F attention also changes feature extraction and conditioning paths, so we report it as a module ablation; Simple Joint and Sequential provide the direct coupling tests while retaining the same backbone and C2F modules. (Details: Sec. 4.3; App. D, Table 11 (9).)

W3. "The overall presentation needs further refinement."

We polished the language throughout, revised Figures 1 and 2, and reformatted the dense tables, while preserving the reported numerical values in unchanged comparisons.

Language. We proofread the main text and appendix, fixing grammar and improving sentence structure, making comparisons explicit, and unifying terminology.

Figures 1 and 2. Each perturbed segment now starts and ends on the original series, so the perturbed series is continuous. In Figure 1, the original and perturbed series boxes share one size and position across panels, panel (b) uses the same split arrows as panel (a), and titles and label colors are consistent. Figure 2 follows the same convention (black clean series, red perturbed support joined to it); its binary mask marks the same two windows, the Product operator is a readable block like Top-k, "Exact budget" matches the selection in Sec. 3.3, and a one-off green arrow now uses the WHAT color. (Details: Fig. 1, p. 2; Fig. 2, p. 4.)

Tables. Table 2 keeps all attacks, targets, and metrics in one table for direct comparison and is now easier to read: ASR cells drop the repeated percent sign, dataset names are rotated, target groups are separated, and C2FDA rows are shaded, which together allow a larger font at the same height. Table 3 uses a fixed 9-pt font, Tables 4 and 6 give each metric its own column instead of slash-separated cells, and Table 5 uses full dataset names. (Details: Tables 2–6, pp. 7–10.)

We hope these answers address the reviewer's concerns and would be glad to clarify any remaining point.

Reviewer_4VcsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes C2FDA, the first diffusion-based framework for local black-box adversarial attacks on time series classifiers. C2FDA couples the where and what decisions through shared coarse-to-fine temporal representations, aiming to achieve joint WHERE-WHAT learning.

Critical strengths

The authors propose that "WHERE" and "WHAT" should be learned jointly, a well-motivated idea, given that the optimal attack location and the content of the perturbation are interdependent.

The use of a coarse-to-fine approach to improve local localization, leveraging "coarse" attention to capture regional information and "fine" attention to pinpoint specific temporal moments, is also logically sound.

Critical weaknesses

The necessity of using a diffusion-based model for perturbation generation is not sufficiently justified. While the authors compare their method (C2FDA) against approaches like Diff-PGD, AdvGAN, and SGAN, these baselines differ from C2FDA in multiple aspects, including attack location selection, generator architecture, and training objectives. Consequently, these comparisons fail to directly address the question: If the proposed coarse-to-fine support selection and hybrid boundary objective were retained, could a simpler generator achieve comparable performance without relying on diffusion?

While the authors claim that their joint architecture outperforms simple joint optimization, supporting this claim through overall baseline comparisons and modular ablation studies. A direct, controlled comparison specifically supporting this architectural choice is missing.

Questions for authors

If the proposed coarse-to-fine support selection and hybrid boundary objective were retained, could a simpler generator achieve comparable performance without relying on diffusion?

The paper includes an experiment labeled "w/o C2F Attention", but this is not equivalent to "w/o WHERE–WHAT coupling". Removing C2F attention simultaneously alters feature extraction capabilities, temporal context modeling, and the sources of information for the two branches. My understanding is the observed changes in performance cannot be attributed solely to joint WHERE–WHAT learning.

3 comments

We thank the reviewer for the precise questions. Q1 repeats the question that closes W1, and Q2 explains why the modular ablation cited in W2 does not isolate the coupling, so we answer each pair together. Each answer opens with one sentence and then follows the reviewer's sub-points in order. Table and figure numbers refer to the revised PDF, with submitted numbers in parentheses where they differ. In the revised PDF, revisions for this review are in blue and revisions shared with other reviews are in purple.

PointAnswer
W1 & Q1With C2F support selection and the hybrid objective retained, replacing the diffusion generator with either a GAN or a VAE does not achieve a comparable ASR–distortion trade-off in the controlled comparison: ASR falls by 11.1–33.3 points, while MSE and DTW increase on all three datasets.
W2 & Q2Simple Joint removes only the WHERE→WHAT conditioning and keeps C2F attention; Simple Joint and Sequential both trail the full model on every dataset and metric, by up to 22.2 and 33.3 ASR points, and w/o C2F Attention is reported as an attention-module ablation.

W1 & Q1. "If the proposed coarse-to-fine support selection and hybrid boundary objective were retained, could a simpler generator achieve comparable performance without relying on diffusion?"

With coarse-to-fine support selection and the hybrid boundary objective retained, replacing the diffusion generator with either a GAN or a VAE does not achieve a comparable ASR–distortion trade-off in the controlled comparison: both alternatives have lower ASR and higher MSE and DTW than the diffusion generator on all three datasets.

Controlled generator replacement. C2F-AdvGAN and C2F-VAE keep C2FDA's C2F support selection and hybrid objective, use the CNN target, and replace only the diffusion generator (Table 4):

DatasetC2FDA ASR (%) ↑C2FDA MSE ↓C2FDA DTW ↓C2F-AdvGAN ASR (%) ↑C2F-AdvGAN MSE ↓C2F-AdvGAN DTW ↓C2F-VAE ASR (%) ↑C2F-VAE MSE ↓C2F-VAE DTW ↓
BasicMotions50.00.060671.7437.5‾\underline{37.5}0.0825‾\underline{0.0825}83.49‾\underline{83.49}31.20.086586.92
HandMovementDirection94.40.034291.4883.3‾\underline{83.3}0.063796.26‾\underline{96.26}61.10.0567‾\underline{0.0567}106.73
UWaveGestureLibrary97.40.084548.2185.7‾\underline{85.7}0.1013‾\underline{0.1013}62.82‾\underline{62.82}77.90.116566.74

C2FDA raises ASR over C2F-AdvGAN by 12.5, 11.1, and 11.7 points while lowering both distortion metrics. (Details: Sec. 4.3, Table 4.)

Why diffusion helps. The diffusion generator refines perturbation content over 20 state-dependent DDIM steps instead of producing it in one pass. Each updated state is fed back through the C2F modules, so location and content are refined together, while the denoising objective is trained across diffusion noise levels and the attack and hiding losses balance effectiveness against perturbation magnitude. The GAN and VAE replacements in Table 4 do not use this diffusion-specific iterative denoising process, which is consistent with their lower ASR and higher MSE and DTW under the same support selection and hybrid objective. (Details: Sec. 3.3; App. A; Eqs. (4), (9)–(11), (29)–(30).)

W2 & Q2. "A direct, controlled comparison specifically supporting this architectural choice is missing." / "The observed changes in performance cannot be attributed solely to joint WHERE–WHAT learning."

Section 4.3 now includes a direct same-backbone comparison. Simple Joint retains the C2F modules and hybrid loss but removes the WHERE→WHAT conditioning; Sequential retains the full architecture but trains WHERE first, fixes the mask, and then trains WHAT. C2FDA outperforms both controls on all three datasets and metrics, while w/o C2F Attention remains a module ablation because it changes feature extraction and conditioning paths.

Joint versus decoupled. Table 3 reports the CNN-target ablations:

DatasetMetricFull C2FDAw/o Coarsew/o Finew/o C2F AttentionRandom IntervalsExact Margin OnlySimple JointSequential
BasicMotionsASR (%) ↑50.0‾\underline{50.0}50.0‾\underline{50.0}31.256.231.243.843.837.5
MSE ↓0.0606‾\underline{0.0606}0.08970.05040.19150.09790.08720.08140.0768
DTW ↓71.7497.8376.46‾\underline{76.46}239.27102.6194.3896.1789.21
HandMovementDirectionASR (%) ↑94.477.866.788.9‾\underline{88.9}50.083.372.261.1
MSE ↓0.0342‾\underline{0.0342}0.06080.03020.20460.05730.04150.04270.0684
DTW ↓91.48‾\underline{91.48}103.7290.34504.85143.7297.63103.57116.92
UWaveGestureLibraryASR (%) ↑97.489.6‾\underline{89.6}76.684.449.487.084.480.5
MSE ↓0.0845‾\underline{0.0845}0.10240.08420.14900.09620.11830.09670.0971
DTW ↓48.2157.6753.20‾\underline{53.20}248.4156.3861.8560.4263.84

Full C2FDA reaches 50.0%, 94.4%, and 97.4% ASR, versus 43.8%, 72.2%, and 84.4% for Simple Joint and 37.5%, 61.1%, and 80.5% for Sequential, with lower MSE and DTW in every case. (Details: Sec. 4.3 "Effect of joint where–what learning", Table 3.)

Attention ablations. As the reviewer notes, w/o Coarse, w/o Fine, and w/o C2F Attention evaluate components. Removing C2F attention also changes feature extraction and conditioning paths, so we report it as a module ablation; Simple Joint and Sequential provide the direct coupling tests while retaining the same backbone and C2F modules. (Details: Sec. 4.3; App. D, Table 11 (9).)

We hope these answers address the reviewer's concerns and would be glad to clarify any remaining point.

Thank you again for reviewing our paper and for the feedback that helped us improve it. We have addressed your questions in our rebuttal and revised manuscript. Please let us know if you have any remaining questions or if there is anything we can clarify further.

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

This paper proposes C2FDA, a coarse-to-fine diffusion framework for local black-box adversarial attacks on time series classifiers. It jointly learns where to perturb and what perturbation to generate through shared coarse-to-fine temporal representations and a mask-aware hybrid boundary objective. Experiments across 14 datasets and five target architectures show that C2FDA achieves the highest mean attack success rate (68.61%) with the smallest temporal coverage (14.64%), while maintaining low perturbation distortion and high temporal similarity.

Critical strengths
  1. The proposed framework jointly optimizes temporal selection and perturbations, which addresses a limitation of existing local attacks.

  2. Extensive experiments are conducted across various time series datasets and classifier architectures. They demonstrate promising cross-model transferability, high attack success rates, and low perturbation distortion.

Critical weaknesses
  1. The paper evaluates perturbation performance mainly through distance metrics like MSE, DTW, temporal coverage. However, low distortion and high temporal similarity do not necessarily guarantee that the perturbed time series remain physically plausible or semantically meaningful. In particular, perturbations restricted to a small subset of timestamps may violate temporal coherence/autocorrelation or domain-specific constraints. Further investigation is needed to establish whether the generated adversarial examples represent realisticness.

  2. The paper does not investigate adversarially trained classifiers or adaptive defenses specifically designed to counter localized perturbations. This limits conclusions about the effectiveness of C2FDA against stronger defenses.

  3. All transfer attacks are trained using a single frozen ResNet surrogate. What about its performance on OOD data or tasks? Especially, the effectiveness of C2FDA depends on the OOD performance of the pre-trained surrogate model. The authors may want to investigate the proposed approach's performance under different surrogate configurations or architectures.

  4. Although the paper analyzes the sensitivity of proposed approach to perturbation budget, it still lacks systematic sensitivity analysis in terms of key hyperparameters including attack-loss weight, and perturbation-regularization weight. Besides, several important choices, including the temporal coverage ratio, coarse segment number, Taylor blending coefficient, and DDIM reverse steps are also not analyzed.

  5. The theoretical time and memory complexity of the coarse-to-fine attention and iterative diffusion sampling is not analyzed, limiting the assessment of scalability to longer or higher-dimensional time series, given concerns that diffusion-based approaches are used to be time and memory-consuming.

Questions for authors

Please address the main concerns in the weakness.

8 comments

We thank the reviewer for the detailed review. Each point below opens with a one-sentence answer and then follows the reviewer's sub-points in order. Table and figure numbers refer to the revised PDF, with submitted numbers in parentheses where they differ. In the revised PDF, revisions for this review are in red and revisions shared with other reviews are in purple.

WeaknessAnswer
W1On a new autocorrelation-function (ACF) distance, which measures changes in temporal dependence, C2FDA has the lowest or tied-lowest reported ACF among target-query-free attacks and the highest ASR of all 13 attacks on all three multivariate datasets.
W2Across the full defense comparison, C2FDA has the highest residual ASR among attacks with nonzero pre-defense ASR on UWaveGestureLibrary under all three defenses (about 87.0%, 65.0%, and 61.1%) and on HandMovementDirection under PGD-AT and CD-GAN; within the query-free local comparison, its DSR is lower than SGAN's on all three datasets.
W3With each of the MLP, ResNet, and Transformer surrogates, C2FDA has higher average ASR and lower average MSE than SGAN over 14 datasets, and its generator keeps 81.0% ASR on classes excluded from generator training, where SGAN drops to 28.6%.
W4Appendix C evaluates all six requested hyperparameters; the defaults of KK, γ\gamma, and DDIM steps give the highest ASR in their sweeps, and the two loss weights balance ASR against MSE.
W5For fixed KK, time and memory are linear in sequence length; a 16× longer input raises per-sample time only 1.42–2.13×, and a complete test set is generated in 1.58 s.

W1. "Low distortion and high temporal similarity do not necessarily guarantee that the perturbed time series remain physically plausible or semantically meaningful."

To test the temporal coherence raised by the reviewer directly, we added the autocorrelation-function (ACF) distance alongside MSE and DTW. On all three multivariate datasets, C2FDA has the highest ASR of all 13 attacks and the lowest or tied-lowest reported ACF among target-query-free attacks.

Temporal coherence (ACF). ACF compares the autocorrelation of the clean and adversarial series over the full sequence, so it also captures dependencies between perturbed and unperturbed timestamps. Although C2FDA perturbs the fewest timestamps of all attacks (14.64% on average over 14 datasets), it lowers ACF relative to the query-free local attack SGAN from 0.27 to 0.10 (BasicMotions), 0.18 to 0.07 (HandMovementDirection), and 0.23 to 0.17 (UWaveGestureLibrary).

Target-averaged ASR (%) and ACF from Table 2 ("--": zero ASR on every target): (Details: Sec. 4.1–4.2, Table 2, App. B Eqs. (37)–(38); Sec. 4.7, Fig. 4 (6).)

AttackBasicMotions ASRBasicMotions ACFHandMovementDirection ASRHandMovementDirection ACFUWaveGestureLibrary ASRUWaveGestureLibrary ACF
FGSM4.30.290.0--14.00.28
PGD1.00.220.0--12.00.17
MI-FGSM2.10.290.0--14.10.21
NES0.0--0.70.1710.30.32
SPSA17.40.136.80.0952.70.16
AutoZOOM20.70.067.90.0358.1‾\underline{58.1}0.02
GATN5.10.422.90.245.70.34
TSadv34.00.3753.1‾\underline{53.1}0.2016.50.06‾\underline{0.06}
AdvGAN4.10.442.80.2136.70.44
Diff-PGD24.10.493.40.3035.20.37
GE-AdvGAN10.20.473.90.3028.00.36
SGAN42.8‾\underline{42.8}0.2725.10.1847.60.23
C2FDA68.40.10‾\underline{0.10}67.90.07‾\underline{0.07}69.70.17

Constraints on the perturbed values. Every adversarial value stays within ϵ\epsilon of the recorded value at the same timestamp, exactly kk timestamps change, and the temporal mask is shared across channels, so all channels keep their recorded values outside the support (Eq. (1)). (Details: Sec. 3.1, Eq. (1).)

Generated content. Within the support, each perturbed value moves to the denoiser's clean-state estimate of the input, clipped to within ϵ\epsilon of the recorded value, under DDIM sampling anchored on the clean input. The examples in Figs. 16–18 (5, 12, 13) show small deviations near waveform transitions, with the waveform preserved elsewhere. (Details: Sec. 3.3; App. A; Figs. 16–18 (5, 12, 13).)

Response to Reviewer 1 (2/7): W2

W2. "The paper does not investigate adversarially trained classifiers or adaptive defenses specifically designed to counter localized perturbations."

We evaluate all 13 attacks against two adversarially trained CNNs, PGD-AT [1] and frequency-selective F-SAT [2], and against collaborative defense-GAN (CD-GAN) [3] purification (Table 5). Across the full comparison, C2FDA has the highest residual ASR among attacks with nonzero pre-defense ASR on UWaveGestureLibrary under all three defenses and on HandMovementDirection under CD-GAN and PGD-AT. Within the query-free local comparison, C2FDA has lower DSR than SGAN on all three datasets; its UWaveGestureLibrary residual ASRs are about 87.0%, 65.0%, and 61.1% under CD-GAN, PGD-AT, and F-SAT, respectively.

Adversarial training and purification. Pre-defense ASR and DSR (%) under each defense (Table 5; Basic = BasicMotions, HMD = HandMovementDirection, UWave = UWaveGestureLibrary):

MethodBasic ASRBasic CD-GAN DSRBasic PGD-AT DSRBasic F-SAT DSRHMD ASRHMD CD-GAN DSRHMD PGD-AT DSRHMD F-SAT DSRUWave ASRUWave CD-GAN DSRUWave PGD-AT DSRUWave F-SAT DSR
FGSM12.550.4100.0100.00.0———28.672.781.886.4
PGD6.20.0100.0100.00.0———24.778.984.284.2
MI-FGSM6.20.0100.0100.00.0———26.070.080.085.0
NES0.0———0.0———18.264.378.678.6
SPSA43.842.942.957.15.60.0100.0100.089.666.663.866.7
AutoZOOM43.842.928.628.611.149.550.050.094.8‾\underline{94.8}53.457.560.3
GATN12.550.4100.0100.05.6100.0100.0100.09.1100.0100.0100.0
TSadv56.222.1‾\underline{22.1}33.3‾\underline{33.3}44.4100.027.838.9‾\underline{38.9}38.9‾\underline{38.9}57.118.0‾\underline{18.0}40.945.5
AdvGAN6.2100.0100.0100.00.0———49.494.784.286.8
Diff-PGD56.266.566.766.70.0———59.776.078.380.4
GE-AdvGAN25.050.075.075.00.0———48.1100.086.586.5
SGAN56.233.344.455.677.850.042.950.076.627.237.3‾\underline{37.3}42.4‾\underline{42.4}
C2FDA50.0‾\underline{50.0}25.037.537.5‾\underline{37.5}94.4‾\underline{94.4}17.6‾\underline{17.6}29.435.397.410.733.337.3

On UWaveGestureLibrary, SGAN, the strongest other query-free local attack after each defense, retains about 55.8%, 48.0%, and 44.1% ASR. (Details: Sec. 4.4, Table 5.)

Purification on the other 11 datasets. CD-GAN purification is evaluated on all 14 datasets. On the other 11 (Table 12 (10)), C2FDA keeps higher post-purification ASR than AutoZOOM and SGAN on all 11 and than TSadv on nine, e.g., 88.7% on ECGFiveDays vs. 23.7% (AutoZOOM), 46.3% (TSadv), and 18.6% (SGAN). Its DSR is lower than AutoZOOM's on all 11 and than SGAN's on all 10 where SGAN's initial ASR is nonzero. (Details: Sec. 4.4, Table 12 (10).)

References

[1] Madry et al., Towards Deep Learning Models Resistant to Adversarial Attacks, ICLR 2018.

[2] Zhang et al., I Can Hear You: Selective Robust Training for Deepfake Audio Detection, ICLR 2025.

[3] Laykaviriyakul and Phaisangittisagul, Collaborative Defense-GAN for Protecting Adversarial Attacks on Classification System, Expert Systems with Applications 2023.

W3. "All transfer attacks are trained using a single frozen ResNet surrogate. What about its performance on OOD data or tasks?"

With each of three surrogate architectures (MLP, ResNet, and Transformer), C2FDA has higher average ASR and lower average MSE than SGAN over all 14 datasets. When classes 7 and 8 are excluded from generator training, C2FDA keeps 81.0% ASR on their test subset, whereas SGAN drops to 28.6% under the same evaluation.

Surrogate architectures. With CNN as the target, Table 6 compares C2FDA and SGAN using MLP, ResNet, and Transformer surrogates. C2FDA's average ASR / MSE is 73.79% / 0.0379 (MLP), 82.72% / 0.0334 (ResNet), and 86.85% / 0.0322 (Transformer), versus 46.2% / 0.0749, 53.7% / 0.0686, and 59.2% / 0.0757 for SGAN; its ASR is higher on 12, 12, and 13 of the 14 datasets, respectively. C2FDA reaches 86.85% average ASR with the Transformer surrogate, compared with 82.72% with the default ResNet surrogate. (Details: Sec. 4.5, Table 6.)

DatasetMLP C2FDA (ASR (%) / MSE)MLP SGAN (ASR (%) / MSE)ResNet C2FDA (ASR (%) / MSE)ResNet SGAN (ASR (%) / MSE)Transformer C2FDA (ASR (%) / MSE)Transformer SGAN (ASR (%) / MSE)
BasicMotions43.8‾\underline{43.8}/0.072450.0/0.1201‾\underline{0.1201}50.0‾\underline{50.0}/0.060656.2/0.1045‾\underline{0.1045}68.8/0.057162.5‾\underline{62.5}/0.1373‾\underline{0.1373}
HandMovementDirection77.8/0.041166.7‾\underline{66.7}/0.1208‾\underline{0.1208}94.4/0.034277.8‾\underline{77.8}/0.1123‾\underline{0.1123}88.9/0.033883.3‾\underline{83.3}/0.1257‾\underline{0.1257}
UWaveGestureLibrary89.6/0.0936‾\underline{0.0936}64.9‾\underline{64.9}/0.092697.4/0.084576.6‾\underline{76.6}/0.0896‾\underline{0.0896}93.5/0.081383.1‾\underline{83.1}/0.0872‾\underline{0.0872}
Car80.0/0.014860.0‾\underline{60.0}/0.0327‾\underline{0.0327}86.7/0.012666.7‾\underline{66.7}/0.0286‾\underline{0.0286}93.3/0.012473.3‾\underline{73.3}/0.0261‾\underline{0.0261}
CinCECGTorso64.0‾\underline{64.0}/**0.0121**70.1/0.0425‾\underline{0.0425}72.3‾\underline{72.3}/**0.0104**79.9/0.0368‾\underline{0.0368}79.9‾\underline{79.9}/**0.0098**84.2/0.0349‾\underline{0.0349}
ECG500089.2/0.024366.5‾\underline{66.5}/0.0831‾\underline{0.0831}100.0/0.022173.7‾\underline{73.7}/0.0783‾\underline{0.0783}98.7/0.021978.1‾\underline{78.1}/0.0747‾\underline{0.0747}
ECGFiveDays93.2/0.020729.4‾\underline{29.4}/0.0221‾\underline{0.0221}100.0/0.0177‾\underline{0.0177}35.6‾\underline{35.6}/**0.0172**98.3/0.018140.1‾\underline{40.1}/0.0204‾\underline{0.0204}
Earthquakes83.1/0.10310.0/—94.8/0.09440.0/—96.1/0.08960.0/—
FordA79.3/0.015662.5‾\underline{62.5}/0.1013‾\underline{0.1013}87.3/0.012771.6‾\underline{71.6}/0.0935‾\underline{0.0935}92.5/0.011879.5‾\underline{79.5}/0.1036‾\underline{0.1036}
FordB82.6/0.014276.4‾\underline{76.4}/0.0751‾\underline{0.0751}90.3/0.012784.1‾\underline{84.1}/0.0693‾\underline{0.0693}89.0/0.013988.9‾\underline{88.9}/0.0677‾\underline{0.0677}
Trace62.5/0.010950.0‾\underline{50.0}/0.0459‾\underline{0.0459}72.5/0.009260.0‾\underline{60.0}/0.0417‾\underline{0.0417}82.5/0.008667.5‾\underline{67.5}/0.0407‾\underline{0.0407}
ItalyPowerDemand78.4/0.044514.1‾\underline{14.1}/0.0934‾\underline{0.0934}87.3/0.040320.2‾\underline{20.2}/0.0787‾\underline{0.0787}91.1/0.038626.3‾\underline{26.3}/0.1076‾\underline{0.1076}
SonyAIBORobotSurface162.9/0.028716.1‾\underline{16.1}/0.0511‾\underline{0.0511}71.8/0.025022.6‾\underline{22.6}/0.0625‾\underline{0.0625}78.2/0.023729.0‾\underline{29.0}/0.0543‾\underline{0.0543}
SonyAIBORobotSurface246.7/0.034820.0‾\underline{20.0}/0.0936‾\underline{0.0936}53.3/0.031926.7‾\underline{26.7}/0.0787‾\underline{0.0787}65.1/0.030732.3‾\underline{32.3}/0.1038‾\underline{0.1038}
Average(14datasets)73.79/0.037946.2‾\underline{46.2}/0.0749‾\underline{0.0749}82.72/0.033453.7‾\underline{53.7}/0.0686‾\underline{0.0686}86.85/0.032259.2‾\underline{59.2}/0.0757‾\underline{0.0757}

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.