acceptodds
Under review as a conference paper at ICLR 2027

Learning Normal Diffusion Dynamics for Backdoor Defense in Text-to-Image Models

Abstract

Backdoor attacks pose a serious threat to the secure deployment of text-to-image (T2I) diffusion models. Existing defenses typically detect backdoors from specific abnormal patterns in internal representations, which may limit their generalizability with the emergence of increasingly diverse attack mechanisms. In this paper, we study backdoor defense of T2I diffusion models from a transition-dynamics perspective. We observe that benign diffusion trajectories exhibit structured and timestep-dependent transition patterns from cross-attention, latent and noise spaces, whereas backdoor attacks tend to induce deviations from such normal evolution. Motivated by these observations, we propose Normal Diffusion Dynamics Learning (NDDL), a novel backdoor defense framework that learns the normal transition dynamics of diffusion trajectories utilizing only benign samples. NDDL constructs compact multi-space trajectory representations and trains a timestep-conditioned dynamics model to predict the diffusion evolution. In the inference phase, deviations between the observed and predicted transitions are exploited to quantify dynamics inconsistency for backdoor detection. NDDL further enables trigger localization without any prior knowledge of the embedded backdoor by performing substitution with low-semantic words. Extensive experiments for diverse backdoor attacks demonstrate the effectiveness and generalizability of our proposed NDDL.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
T2I diffusion models, Backdoor detection, Trigger localization, Diffusion dynamics, Normal dynamics learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_twcEOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes NDDL, a framework for backdoor prompt detection and trigger localization in text-to-image diffusion models. It extracts temporal representations and trains a lightweight, timestep-conditioned network on benign trajectories to predict next-step feature increments. During inference, prediction errors serve as anomaly scores for detection, while low-semantic token substitutions enable trigger localization. Experiments evaluate the framework across multiple diffusion architectures, attack methods, and trigger types.

Critical strengths
  • The paper is clearly written and easy to follow.
  • The evaluation covers a broad range of attacks, diffusion architectures, and trigger types, providing substantial empirical support for the method’s effectiveness in the evaluated settings.
Critical weaknesses
  • Since the dynamics predictor is trained on benign trajectories, large prediction errors may reflect a shift in the benign input distribution rather than backdoor activation. The paper should evaluate benign OOD inputs: 1) training on benign DiffusionDB captions and testing on benign COCO captions, 2) training on benign prompts longer than 15 tokens and testing on prompts shorter than 7 tokens.
  • Although a threshold formula is provided, the parameter values like λ,m\lambda, m are unclear, limiting reproducibility and practical deployment.
Questions for authors

The claim of “a new transition-dynamics perspective” appears overstated, as [1] already examines dynamic features for backdoor detection in T2I diffusion models. The authors should discuss this work and clarify the distinction between the two approaches, at least in the related work section.

[1] Wang et al. Dynamic Attention Analysis for Backdoor Detection in Text-to-Image Diffusion Models. TPAMI’25.

7 comments
AuthorsOpenReview ↗

Response to Reviewer (part 1 of 3)

1. Response to Robustness to Benign Distribution Shifts

We thank the reviewer for the constructive comments and positive assessment of our work. We appreciate the concern that a normal dynamics model trained exclusively on benign samples may mistake distributional variations in legitimate prompts for backdoor-induced anomalies. To address this issue, we have conducted additional experiments under two types of benign prompt distribution shifts:

  • Cross-Dataset Distribution Shift: We train the normal dynamics predictor using benign prompts from DiffusionDB and evaluate it on prompts from COCO, without retraining the predictor.
  • Prompt-Length Distribution Shift: We train the predictor using long prompts (more than 15 tokens) and evaluate it on short prompts (fewer than 7 tokens), thereby introducing a substantial change in prompt length.

We report the false positive rate (FPR) and true positive rate (TPR) to evaluate whether these shifts lead to false alarms or degraded backdoor detection.

Table R1: Robustness to benign prompt distribution shifts on SD1.5 (%).

TrainingInferenceFPR ↓TPR ↑
DiffusionDBDiffusionDB1.6698.00
DiffusionDBCOCO2.1297.74
Long promptsLong prompts1.9097.96
Long promptsShort prompts1.4497.70

Results are averaged across five backdoor attacks: BadT2I, EvilEdit, MasqLoRA, Rickrolling, and STEBA.

As shown in Table R1, NDDL maintains relatively stable detection performance under both distribution shifts:

  • Transferring from DiffusionDB to COCO increases FPR by only 0.46 percentage points (1.66% to 2.12%), while TPR decreases by 0.26 percentage points (98.00% to 97.74%).
  • Under the prompt-length shift, FPR decreases from 1.90% to 1.44%, while TPR changes only slightly from 97.96% to 97.70%.

We further conduct experiments on SD3.5, a DiT-based diffusion model. The results are summarized below.

Table R2: Robustness to benign prompt distribution shifts on SD3.5 (%).

TrainingInferenceFPR ↓TPR ↑
DiffusionDBDiffusionDB9.193.6
DiffusionDBCOCO8.392.4
Long promptsLong prompts8.893.2
Long promptsShort prompts9.492.8

As shown in Table R2, NDDL still maintains comparable FPR and TPR under both distribution shifts on SD3.5. In particular, the cross-dataset shift decreases TPR by only 1.2 percentage points, while the prompt-length shift decreases TPR by 0.4 percentage points.

Overall, these experiments provide empirical evidence that NDDL can tolerate the evaluated benign distribution shifts across both UNet- and DiT-based diffusion architectures. We have included the detailed results in Table 4 and Appendix D.5 (Table 15) of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer (part 2 of 3)

2. Response to Threshold Selection and Reproducibility

We thank the reviewer for pointing out the need to clarify the threshold configuration. We agree that providing explicit parameter settings is important for reproducibility.

In NDDL, the detection threshold is calibrated using only benign validation trajectories. Specifically, we fit a Gaussian distribution to their anomaly scores and determine the detection threshold as: λ1=μbenign+mσbenign\lambda_1 = \mu_{\mathrm{benign}} + m\sigma_{\mathrm{benign}}

where μbenign\mu_{\mathrm{benign}} and σbenign\sigma_{\mathrm{benign}} are estimated from benign validation scores, and mm is a balance weight (m=1m=1 in our experiments). Similarly, the localization threshold λ2\lambda_2 is determined using the same procedure as λ1\lambda_1.

To improve reproducibility, we have added the detailed model and hyperparameter configurations in Appendix D.1 (Table 11), including the loss weights and detection/localization thresholds across different diffusion backbones. Table R3 presents the hyperparameter configurations of NDDL.

Table R3: Hyperparameter configurations of NDDL across different models.

BackboneλA\lambda_Aλz\lambda_zλϵ\lambda_\epsilonλ1\lambda_1λ2\lambda_2
SD1.51.05.32.50.001790.00081
SDXL1.55.02.50.001540.00076
SD3.51.06.02.80.002130.00092
PixArt-α1.35.72.20.001690.00079
AuthorsOpenReview ↗

Response to Reviewer (part 3 of 3)

3. Response to the Novelty Concern: Distinction from DAA

We thank the reviewer for highlighting the closely related work DAA. We acknowledge that DAA has already investigated the temporal evolution of attention representations for backdoor detection. However, the fundamental distinction between DAA and NDDL lies in what is modeled and how backdoor anomalies are identified.

Methodological Distinction. Specifically, the methodological differences can be summarized as follows:

  • Modeling Objective: DAA characterizes attention evolution patterns to identify anomalous behaviors. In contrast, NDDL explicitly learns normal diffusion transition dynamics from benign trajectories, integrating attention, latent states, and noise predictions into a unified multi-space representation.
  • Detection Criterion: NDDL employs a timestep-conditioned predictor to estimate normal transitions and identifies backdoor anomalies through inconsistencies between observed and predicted transitions, rather than relying solely on characteristics of attention evolution.

Experimental Comparison. To further substantiate this distinction, we have included DAA-I and DAA-S as additional baselines across four diffusion models.

Table R4: Detection comparison with DAA across different diffusion models.

BackboneMethodACC ↑AUROC ↑
SD1.5DAA-I76.075.9
DAA-S78.178.6
NDDL (Ours)98.198.2
SDXLDAA-I77.577.1
DAA-S79.880.4
NDDL (Ours)98.098.0
SD3.5DAA-I71.571.6
DAA-S75.875.7
NDDL (Ours)91.292.7
PixArt-αDAA-I76.477.0
DAA-S86.585.6
NDDL (Ours)92.391.7

Results on SD1.5 and SDXL are averaged across five attacks; those on SD3.5 and PixArt-α correspond to Rickrolling.

The comparison highlights two aspects of NDDL's effectiveness:

  • Detection Performance: As shown in Table R4, NDDL consistently outperforms both DAA variants in ACC and AUROC across all four evaluated diffusion backbones. Specifically, NDDL improves the average AUROC by 22.32 and 19.62 percentage points on SD1.5, and 20.94 and 17.64 percentage points on SDXL, compared with DAA-I and DAA-S, respectively. These improvements are also observed on the evaluated DiT-based architectures.
  • Trigger Token Localization: Beyond the detection results in Table R4, NDDL additionally supports explicit trigger token localization, with localization performance reported separately in Tables 2, 3, 13 and 14 of the revised manuscript. In contrast, DAA-I and DAA-S do not provide dedicated mechanisms for identifying malicious trigger tokens.

Experimental Evidence for Normal Dynamics Learning. We further investigate whether learning normal transition dynamics contributes to detection performance, beyond directly exploiting temporal differences.

Table R5: Ablation study of normal dynamics modeling on BadT2I (%).

Modeling StrategyACC ↑AUROC ↑
Raw Diff30.532.9
Linear w/o Time42.843.2
Linear w/ Time38.736.8
NDDL w/o Time62.260.5
NDDL (Ours)99.099.2

The ablation results provide two important observations:

  • Normal Dynamics Modeling: Directly using raw transition differences or linear predictors yields limited detection performance. In contrast, learning nonlinear normal transition dynamics substantially improves detection effectiveness, supporting the importance of explicitly modeling benign diffusion evolution.
  • Timestep Conditioning: Incorporating timestep information further improves the effectiveness of normal dynamics learning, increasing AUROC from 60.5% to 99.2%. Similar improvements are observed across the other four attacks (Appendix D.7, Table 17).

Together, these results provide empirical evidence that NDDL's effectiveness stems from learning timestep-conditioned normal diffusion transition dynamics, rather than merely exploiting temporal attention information. We have also expanded the Related Work discussion to clarify the methodological distinction between NDDL and DAA in the revised manuscript.


We sincerely thank the reviewer for the thoughtful comments and valuable suggestions, which have helped us strengthen the evaluation of generalizability and clarify the experimental configurations. We hope that the additional results and revisions adequately address the reviewer's concerns.


Reviewer_o3rdOpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes NDDL, a backdoor defense framework for text-to-image diffusion models. NDDL extracts trajectory representations from the cross-attention, latent, and noise-prediction spaces, and trains a timestep-conditioned MLP for backdoor detection and malicious trigger localization.

Critical strengths

The motivation and perspective of this paper are relatively well-grounded and cutting-edge, and the paper proposes a fairly substantial solution. The reported performance is generally competitive with included baselines.

Critical weaknesses

My primary concern of this paper is that its core idea may not be entirely novel. For example, [1] also examined the temporal evolution of cross-attention for backdoor detection, however, this closely related work is neither discussed nor included in the experimental comparisons. Although this paper introduces some technical designs that differ from [1], such differences may be insufficient to the claim that the proposed method offers a substantial advantage over more related works without direct experimental evidence.

[1] Z. Wang, J. Zhang, S. Shan and X. Chen, "Dynamic Attention Analysis for Backdoor Detection in Text-to-Image Diffusion Models," in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 48, no. 3, pp. 3652-3665, March 2026, doi: 10.1109/TPAMI.2025.3644016.

Questions for authors

Could the authors provide a more rigorous review of the related literature to fully substantiate the novelty of their work, along with theoretical and/or experimental evidence to support the contribution of their method?

8 comments
AuthorsOpenReview ↗

Response to Reviewer (part 1 of 2)

We thank the reviewer for this valuable comment. Accordingly, we have expanded the related work discussion and conducted new experiments, as detailed below.

1. Methodological Distinction from DAA

We acknowledge that DAA investigates the temporal evolution of attention representations for backdoor detection. However, the key distinction between DAA and NDDL lies in what is modeled and how backdoor anomalies are identified.

Specifically, the methodological differences can be summarized as follows:

  • Modeling Objective: DAA characterizes attention evolution patterns, whereas NDDL explicitly learns normal diffusion transition dynamics from benign trajectories, using a timestep-conditioned predictor and multi-space representations derived from attention, latent states, and noise predictions.
  • Detection Criterion: DAA identifies anomalies through characteristics of attention dynamics, whereas NDDL detects backdoor activation by measuring inconsistencies between observed and predicted transitions.

Thus, our contribution is not the use of temporal information itself, but the formulation of backdoor detection through normal diffusion dynamics learning and transition prediction inconsistency.

2. Direct Experimental Comparison with DAA

To substantiate this distinction, we have included DAA-I and DAA-S as additional baselines across four diffusion backbones.

Table R1: Detection comparison with DAA across different diffusion models.

BackboneMethodACC ↑AUROC ↑
SD1.5DAA-I76.075.9
DAA-S78.178.6
NDDL (Ours)98.198.2
SDXLDAA-I77.577.1
DAA-S79.880.4
NDDL (Ours)98.098.0
SD3.5DAA-I71.571.6
DAA-S75.875.7
NDDL (Ours)91.292.7
PixArt-αDAA-I76.477.0
DAA-S86.585.6
NDDL (Ours)92.391.7

Results on SD1.5 and SDXL are averaged across five attacks; those on SD3.5 and PixArt-α correspond to Rickrolling.

The comparison highlights two aspects of NDDL's effectiveness:

  • Detection Performance: As shown in Table R1, NDDL consistently outperforms both DAA variants in ACC and AUROC across all four evaluated diffusion backbones. Specifically, NDDL improves the average AUROC by 22.32 and 19.62 percentage points on SD1.5, and 20.94 and 17.64 percentage points on SDXL, compared with DAA-I and DAA-S, respectively. These improvements are also observed on the evaluated DiT-based architectures.
  • Trigger Token Localization: Beyond the detection results in Table R1, NDDL additionally supports explicit trigger token localization, with localization performance reported separately in Tables 2, 3, 13, and 14 of the revised manuscript. In contrast, DAA-I and DAA-S do not provide dedicated mechanisms for identifying malicious trigger tokens.

Overall, these experimental results demonstrate that NDDL consistently outperforms DAA-I and DAA-S in backdoor detection, while additionally offering trigger token localization capabilities that are not provided by either DAA variant.

AuthorsOpenReview ↗

Response to Reviewer (part 2 of 2)

3. Experimental Evidence for Normal Dynamics Learning

We further conduct ablation studies to investigate the contribution of normal dynamics modeling.

Table R2: Ablation study of different modeling strategies on BadT2I.

Modeling StrategyACC ↑AUROC ↑
Raw Diff30.532.9
Linear w/o Time42.843.2
Linear w/ Time38.736.8
NDDL w/o Time62.260.5
NDDL (Ours)99.099.2

As shown in Table R2, we observe that:

  • Raw transition differences and linear predictors exhibit limited detection performance.
  • Nonlinear dynamics learning improves detection effectiveness, particularly with timestep conditioning.

Similar improvements are observed across the other four attacks (Appendix D.7, Table 17). Additionally, the multi-space ablation (Table 6) demonstrates the complementary contributions of attention, latent and noise representations.

These results support the importance of multi-space representations and timestep-conditioned normal dynamics learning for effective backdoor detection.

4. Theoretical Justification

We also provide a theoretical analysis in Section 3:

  • Proposition 1 characterizes the propagation of backdoor-induced perturbations along diffusion trajectories under a local smoothness assumption.
  • Proposition 2 establishes that, under a bounded prediction error assumption, benign transition residuals are upper-bounded by the predictor approximation error, whereas backdoor residuals are lower-bounded by the trigger-induced perturbation magnitude minus this error.

This analysis provides a theoretical motivation for detecting backdoors through deviations from learned normal diffusion dynamics.

We believe these revisions further clarify the methodological contribution of NDDL and provide additional theoretical and experimental evidence supporting its effectiveness.


We sincerely appreciate the reviewer's insightful comments and constructive feedback, which have helped us better clarify the methodological novelty and strengthen the experimental justification of NDDL. We hope that our detailed explanations and additional results satisfactorily address the reviewer's concerns.


Dear Reviewer o3rd,

Thank you again for your thoughtful review. As the discussion phase approaches its end, we would like to check whether our response has addressed your main concerns. If any point remains unclear or if there are additional questions we could clarify, we would be happy to provide further clarification.

We greatly appreciate your time and consideration.

Best regards,

The Authors

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

The paper introduces Normal Diffusion Dynamics Learning (NDDL), a defense method for poisoned T2I diffusion models. The paper argues that the previous works focused on specific abnormal behaviors caused by specific backdoor attacks which cannot be generalized. However, NDDL learns normal diffusion trajectories using only benign samples and detects backdoor attacks with features from the denoising process. NDDL achieved the best score compared to the previous baselines against to the attack methods.

Critical strengths
  1. The idea of learning normal diffusion dynamics using only benign samples, rather than focusing on specific backdoor behaviors, is interesting and important.
  2. The proposed method achieves the best performance in most benchmark settings.
Critical weaknesses
  1. The proposed method requires extracting attention, latent, and noise representations across multiple diffusion timesteps, which may be computationally expensive. Table 4 also shows that all these representations are needed to achieve high performance. Although Figure 4 suggests that certain denoising stages are more informative for backdoor detection, the paper does not explore how this observation could be leveraged to reduce computational costs.
  2. The paper argues that this method is generalizable; however, it is not evaluated on different attack methods [1,2] that are not based on prompt injection. Perhaps self-attention maps are more important than the cross-attention maps in these methods.

[1] TrojDiff: Trojan Attacks on Diffusion Models with Diverse Targets (CVPR 2023) [2] How to Backdoor Diffusion Models? (CVPR 2023)

Questions for authors
  1. As Figure 4 suggests, the middle timesteps are the best for detecting the behaviors; have the authors considered using only certain timesteps?
  2. The authors argue that existing methods, such as STEDF, rely on specific abnormal patterns associated with backdoor attacks. However, STEDF also analyzes spatio-temporal features. How is STEDF specifically dependent on these abnormal patterns, and what makes NDDL more generalizable?
  3. Most comparisons are conducted on outdated SD1.5 or SDXL, which is understandable since prior work has mainly focused on these models. Although the authors also evaluate DiT-based models, it is unclear how the attention representations are adapted across different architectures.
8 comments
AuthorsOpenReview ↗

Response to Reviewer (part 1 of 4)

We thank the reviewer for recognizing the importance of learning normal diffusion dynamics from benign samples and the strong detection performance of NDDL. We appreciate the concern regarding its computational overhead and the potential benefits of informative timestep selection.

1. Response to Computational Overhead.

We agree that extracting multi-space representations incurs additional computational costs. To quantitatively address this concern, we have conducted additional inference-time comparisons across four diffusion backbones (Table 5 and Appendix D.6, Table 16 in the revised manuscript).

Table R1: Inference time comparison of different defense methods (seconds).

MethodSD1.5SDXLSD3.5PixArt-α
UFID46.9108.2154.186.4
T2IShield13.826.4——
NaviT2I3.65.118.311.0
STEDF3.75.8——
DAA-I4.729.227.515.1
DAA-S4.830.436.515.9
NDDL6.16.930.516.6

As shown in Table R1, together with the detection results in Tables 1, 3, 12 and 14 of the revised manuscript, we obtain two key observations:

  • NDDL achieves superior backdoor detection performance without incurring excessive inference time across different diffusion backbones.
  • Although NDDL is not the fastest defense, its inference time remains within a reasonable range compared with existing methods.

These results demonstrate that NDDL achieves a reasonable trade-off between detection effectiveness and computational efficiency.

2. Response to Informative Timestep Selection and Computational Efficiency.

We thank the reviewer for this insightful suggestion. We agree that selecting informative denoising timesteps can improve computational efficiency. To address this concern, we provide the following analysis and additional experiments.

Selection of Middle Denoising Stages. As demonstrated by the denoising-stage ablation in Figure 4, the middle stage achieves the best detection performance, indicating that backdoor-induced dynamics deviations are more discriminative during this stage. Motivated by this observation, NDDL focuses on the middle denoising stage across different diffusion backbones rather than extracting representations throughout the entire sampling process. Specifically, we select steps 16–30 out of 50 for SD1.5, steps 11–20 out of 30 for SDXL and SD3.5, and steps 7–12 out of 20 for PixArt-α.

Computational Benefits. To quantify the efficiency improvements, we compare the inference time of full-stage and middle-stage representation extraction across four diffusion backbones.

Table R2: Inference time comparison between full-stage and middle-stage representation extraction (seconds).

BackboneFull-stageMiddle-stageReduction (%) ↑
SD1.56.66.17.58
SDXL8.46.917.86
SD3.556.830.546.30
PixArt-α20.216.617.82

As shown in Table R2, focusing on the middle denoising stage reduces inference time across all four diffusion backbones, with reductions of 7.58%, 17.86%, 46.30%, and 17.82% on SD1.5, SDXL, SD3.5, and PixArt-α, respectively. These results demonstrate that selecting more discriminative denoising stages can reduce the computational overhead of NDDL, particularly on SD3.5.

We have also provided detailed configurations in Appendix D.1 (Table 11) of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer (part 2 of 4)

3. Response to Evaluation on TrojDiff and BadDiffusion.

We thank the reviewer for this valuable suggestion. Our original evaluation primarily focuses on text-triggered backdoor attacks against text-to-image diffusion models, where malicious behaviors are activated by specific textual triggers. In contrast, TrojDiff and BadDiffusion employ noise-based triggers in the diffusion initialization process, introducing different trigger mechanisms and operating without textual conditioning.

To address this concern, we have conducted new experiments against both TrojDiff and BadDiffusion (Appendix D.8, Table 18).

Both attacks are implemented on CIFAR-10 using the diffusion model architecture adopted by TrojDiff. Specifically:

  • TrojDiff: We blend a Hello Kitty pattern with noise as the trigger and use a Mickey Mouse image as the target, with a poisoning rate of 10%. The resulting model achieves a target MSE of 0.0007 and a benign FID of 7.34.
  • BadDiffusion: We inject a 14 × 14 gray patch into noise as the trigger and use a hat image as the target, with a poisoning rate of 20%. The resulting model achieves a target MSE of 0.0009 and a benign FID of 10.32.

Since these unconditional diffusion models do not employ text-conditioned cross-attention, we adapt NDDL by extracting self-attention representations while retaining the normal dynamics learning framework.

Table R3: Detection performance against noise-triggered backdoor attacks.

AttackACC ↑AUROC ↑
TrojDiff98.899.3
BadDiffusion98.698.5

The results in Table R3 show that NDDL achieves ACC and AUROC above 98% on both attacks, demonstrating strong discrimination between benign and triggered diffusion trajectories in these noise-triggered settings.

More importantly, these results extend our original evaluation beyond text-triggered backdoors. These additional experiments provide further empirical evidence supporting NDDL's applicability across different trigger modalities and diffusion architectures.

AuthorsOpenReview ↗

Response to Reviewer (part 3 of 4)

4. Response to How does NDDL differ from STEDF?

We thank the reviewer for this important question. We acknowledge that STEDF also exploits spatio-temporal information. However, the key distinction between STEDF and NDDL lies in what is modeled and how backdoor anomalies are identified, rather than in the use of temporal information itself.

Table R4: Methodological comparison between STEDF and NDDL.

AspectSTEDFNDDL (Ours)
Main perspectiveSpatio-temporal feature analysisNormal transition dynamics learning
CharacterizationWeight enrichment and temporal anisotropyTimestep-conditioned normal state evolution
Detection criterionDiscriminative spatio-temporal characteristicsObserved–predicted transition inconsistency
Reference modelingAttack-related discriminative patternsLearned benign trajectory transitions
RepresentationSpatio-temporal featuresAttention + latent + noise prediction

As summarized in Table R4, the methodological differences between STEDF and NDDL can be highlighted from three aspects:

  • Modeling Perspective: STEDF characterizes backdoor-related spatio-temporal patterns, including weight enrichment and temporal anisotropy. In contrast, NDDL explicitly learns normal diffusion transition dynamics from benign trajectories using a timestep-conditioned predictor.
  • Detection Criterion: STEDF identifies backdoor anomalies based on discriminative spatio-temporal characteristics, whereas NDDL measures inconsistencies between observed and predicted transitions to detect deviations from learned normal dynamics.
  • Representation: STEDF exploits spatio-temporal features, while NDDL constructs compact trajectory representations across attention, latent, and noise prediction spaces to capture diffusion dynamics.

These distinctions highlight that NDDL differs from STEDF in its modeling objective and detection mechanism, rather than merely in the use of temporal information.

Our experimental results further support this distinction, as summarized below:

  • Detection Performance: As shown in Table 1, NDDL outperforms STEDF in both ACC and AUROC on four of the five attacks on SD1.5, while achieving comparable performance on BadT2I. NDDL also demonstrates strong detection performance on SDXL, as reported in Table 12.
  • Trigger Token Localization: Beyond backdoor detection, NDDL enables explicit trigger token localization, whereas STEDF does not provide a dedicated mechanism for identifying malicious trigger tokens. This additional capability further distinguishes NDDL from STEDF.

Table R5: Ablation study of different modeling strategies against BadT2I. (More results of other attacks can be found in Table 17)

Modeling StrategyACC ↑AUROC ↑
Raw Diff30.532.9
Linear w/o Time42.843.2
Linear w/ Time38.736.8
NDDL w/o Time62.260.5
NDDL99.099.2

Furthermore, as shown in Table R5 (Figure 5(a) and Appendix D.7, Table 17), our modeling ablation demonstrates that raw transition differences and linear predictors are substantially less effective than learning nonlinear, timestep-conditioned normal dynamics. This supports the contribution of the proposed learning objective beyond simply analyzing temporal features.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to all reviewers

We sincerely thank all reviewers for their careful evaluations and constructive feedback. We appreciate the recognition of the motivation behind learning normal diffusion dynamics for backdoor defense. The comments have helped us clarify NDDL's methodological contribution, broaden the experimental evaluation, and improve the reproducibility and presentation of the paper.

Key Improvements and Clarifications

1. Expanded comparisons with existing backdoor defenses.

  • Added DAA-I and DAA-S as additional baselines across four diffusion backbones: SD1.5, SDXL, SD3.5, and PixArt-α (Table 1, Table 3, Table 12 in Appendix D.2, Table 14 in Appendix D.4).
  • Expanded the related work discussion to clarify the methodological distinctions between NDDL and existing temporal analysis methods, particularly DAA and STEDF (Section 2).
  • Provided additional experimental evidence to substantiate the effectiveness of normal diffusion dynamics learning (Table 1, Table 3, Table 12 in Appendix D.2, Table 14 in Appendix D.4, Figure 5(a), Table 17 in Appendix D.7)

2. Additional experiments on normal dynamics modeling and generalization.

  • Conducted additional ablation experiments comparing raw transition differences, linear predictors, and nonlinear dynamics models with and without timestep conditioning (Figure 5(a), Table 17 in Appendix D.7).
  • Added evaluations against two noise-triggered backdoor attacks, TrojDiff and BadDiffusion, extending the experimental scope beyond text-triggered backdoors (Table 18 in Appendix D.8).
  • Supplemented the evaluation of NDDL across diffusion architectures, including comparisons with DAA on DiT-based models (Table 3, Table 14 in Appendix D.4).

3. More comprehensive robustness and computational efficiency analysis.

  • Added experiments on benign prompt distribution shifts, including cross-dataset transfer and prompt-length variations, to assess false positives and detection robustness (Table 4, Table 15 in Appendix D.5).
  • Provided additional inference-time comparisons across different diffusion backbones (Table 5, Table 16 in Appendix D.6).
  • Further investigated the computational benefits of selecting middle denoising stages through comparisons with full-stage feature extraction.

4. Improved implementation details and reproducibility.

  • Added detailed model and hyperparameter configurations, including feature weights, detection/localization thresholds, sampling steps, selected timestep intervals, and window sizes (Table 11 in Appendix D.1).
  • Clarified the selection of informative denoising stages and the architecture-specific attention extraction procedures for SD3.5 and PixArt-α (Appendix C.1, Table 11 in Appendix D.1).
  • Provided further explanations of the benign-validation-based threshold calibration procedure.

Please see our detailed responses below each review and the blue-highlighted sections in the updated paper. We appreciate your feedback and welcome any additional questions.

Area_Chair_4aqMOpenReview ↗

Reminder to respond to the authors’ rebuttal

Dear Reviewers,

Thank you for your service as reviewers for ICLR 2027. As the discussion deadline approaches, please read the authors’ rebuttal and respond at your earliest convenience. A brief note indicating whether the authors have addressed your concerns would be helpful.

Thank you for your time.

Best regards,
Area Chair 4aqM

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.