acceptodds
Withdrawn from ICLR 2027

DIFD: Pre-Training Poisoned Sample Screening via Timestep-Calibrated Diffusion Residuals

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Backdoor attacks embed covert triggers into training data to induce attacker-specified results during inference. To prevent these malicious shortcuts from being encoded into model parameters, screening poisoned samples prior to model training is essential. However, pre-training screening faces two key hurdles: defenders lack access to a trained classifier or trigger priors, and backdoor triggers exhibit vast heterogeneity across spatial and spectral domains, making fixed-rule detectors hard to generalize. To address these challenges, we propose DIFD, a pre-training poisoned sample screening framework via timestep-calibrated diffusion residuals. We observe that intermediate reconstruction residuals of poisoned and benign samples exhibit timestep-dependent separability, while inadequately reconstructed trigger components remain in the residual as spatial structural or spectral anomalies. Based on these observations, DIFD first performs timestep calibration to locate an intermediate reconstruction window where this separability is maximized. It then employs spatial-spectral residual scoring to jointly measure spatial structural deviations and spectral distribution shifts, producing a unified sample-level risk score. DIFD relies solely on a fixed diffusion model pretrained on clean data, requires no auxiliary detector training or access to a downstream classifier, and avoids full reverse diffusion sampling. Experiments across multiple datasets, image resolutions, and attack settings demonstrate that DIFD achieves stable cross-attack detection performance and efficient poisoned sample screening.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Backdoor Detection, Poisoned Sample Screening, Diffusion Models, Timestep Calibration, Spatial-Spectral Residual Scoring

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_mC3SOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DIFD, a pre-training poisoned-sample screening framework based on timestep-aware diffusion reconstruction residuals. By adaptively locating intermediate diffusion stages with high residual separability and jointly modeling spatial and spectral residual anomalies, DIFD enables unified detection of diverse backdoor trigger patterns.

Critical strengths

The datasets involved in experiments are comprehensive.

Critical weaknesses

The defense comparison is limited in scope: only four methods are included, all from before 2024. Since the proposed method primarily employs a diffusion model as the architecture of a detection model rather than being restricted to diffusion-model-based tasks, the evaluation should include mainstream detection methods beyond diffusion models. Suitable baselines may be sourced from BackdoorBench.

The empirical effectiveness and reporting require careful verification. In Table 1, two metrics fluctuate greatly, with ranges of 0.8135–1.0000 and 0.0400–1.0000, respectively. There are also data annotation errors: in the CelebA experiment, the authors marked 0.0400 as the best, whereas the corresponding value for the TERD method is 0.3700, much higher than DIFD. This phenomenon occurs in more than one place; the authors are expected to check carefully.

The manuscript lacks theoretical basis and theoretical analysis, which makes it difficult to construct a complete understanding of the backdoor mechanism.

Questions for authors

I would appreciate clarification on the following points, which are intended to strengthen the theoretical and empirical evaluation.

1.Could the authors clarify why the trigger causes “intermediate reconstruction residuals of poisoned samples and benign samples to exhibit time-step-dependent separability”? Is there a theoretical basis or causal relationship underlying this phenomenon? Moreover, does this observation contribute to understanding the backdoor mechanism?

2.Do benign samples from different categories also exhibit time-step-dependent separability?

3.The manuscript states that “DIFD first performs time-step calibration to locate an intermediate reconstruction window in which separability is maximized.” How is the window size determined? Is it consistent across different datasets? What is the impact of hyperparameter selection on the effectiveness of the method?

4.The manuscript states that “DIFD relies only on a fixed diffusion model pretrained on clean data.” To what extent does the method depend on clean samples (in terms of number and categories; this relates to its practicality)? Is the generality of the method limited to diffusion models? For tasks not suitable for diffusion models, is adaptation difficult (e.g., can it be transferred to Transformer architectures)?

5.Since the proposed method is built on spatial-spectral residuals, frequency-domain-related triggers should be indispensable test cases. Narcissus can complete an attack with only a 0.05% poisoning rate. Can this method establish separability of residuals in this setting? SIBA triggers are coupled with normal features. Can this method establish separability of residuals in that case? Does it work for backdoor attacks with multi-target settings?

6.The organization of the paper could be improved. In the introduction, the authors claim three contributions, but experimental results should not be presented as a separate contribution. The presentation of illustrations requires improvement. The authors use extremely low-resolution images, which appear blurry. I recommend high-resolution visualizations on Tiny-imagenet involved in the experiments. In addition, the titles do not explain the image content, making the figures difficult to understand.

  1. In addition, several important experiments are missing: (i) sensitivity to the detection threshold; (ii) sensitivity to the window-length hyperparameter; (iii) cross-model-architecture capability (different diffusion model architectures); and (iv) degree of dependence on clean samples.
Reviewer_uS73OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

DIFD screens training images for poisoning using a fixed clean diffusion prior. It calibrates a denoising timestep and combines spatial and spectral reconstruction residuals. Experiments cover several datasets, trigger families, and poison rates.

Critical strengths
  1. The method screens poisoned images before classifier training, using a trusted clean diffusion prior.
  2. Experiments cover several datasets and trigger families, including spatial and frequency-based attacks.
  3. Reporting TPR at low FPR and varying poison prevalence makes the evaluation relevant to practical data screening.
Critical weaknesses
  1. Eqs. 15–17 leave essential parts of the detector undefined, including the saliency operator, VGG features, projection subspace, normalization, and mixing weight. These choices affect the final score and need to be specified for the results to be reproducible.
  2. The adaptation of DisDet, TERD, and CBD to the proposed screening task is unclear. Their original settings require different models or information, so the comparison needs an explicit account of what each method receives and how it produces sample-level scores.
  3. The experiments do not isolate the benefit of timestep calibration. Figure 3 alone cannot establish that the proposed surrogate selects more discriminative timesteps. Comparisons against fixed timesteps and the unweighted surrogate, with a clearly specified calibration set, would address this gap.
Questions for authors
  1. Please specify the operators, checkpoints, normalization, and mixing weight in Eqs. 15–17.
  2. How were DisDet, TERD, and CBD adapted, and what models and information did each receive?
  3. Does calibrated timestep selection outperform fixed timesteps and the unweighted surrogate under the same evaluation protocol?
Reviewer_obLdOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes DIFD, pre-training phase detection method to identify poisoned sample designed to backdoor a model training. DIFD uses diffusion model to denoise the input image and analyze the reconstructed noise/residual noise to identify potential anomaly. DIDF leverages the observation that intermediate reconstruction residuals between poisoned and benign samples show timestep-dependent separability. Evaluation on multiple datasets and attack categories demonstrate the effective of DIFD.

Critical strengths
  • DIFD relies entirely on a fixed diffusion model pre-trained on clean data, avoiding the need for auxiliary detector training or downstream classifiers.
  • By utilizing timestep calibration and spatial-spectral residual scoring, it effectively captures both structural and spectral anomalies across diverse backdoor attacks.
Critical weaknesses
  • It remains unclear how DIFD reliably differentiates between naturally noisy or anomalous benign data and actual backdoor-triggered samples, raising concerns about potential false positives.
  • The reported evaluation results exhibit fluctuating detection rates across different datasets and attack settings, undermining claims of robust generalization.
Questions for authors
  • How does DIFD theoretically and empirically distinguish between naturally occurring anomalies or noisy benign samples and actual backdoor-triggered data?
  • Can the authors clarify the causes behind the fluctuating detection rates across certain datasets and attack types, and how the framework's stability can be improved for more robust cross-attack generalization?

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

Related papers

Discussion (0)

Sign in to comment.