acceptodds
Withdrawn from ICLR 2027

Tracing How Backdoors Are Learned: Poisoned Sample Detection via Information-Trajectory Distortion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Deep neural networks (DNNs) are vulnerable to backdoor attacks, in which adversaries inject training samples containing specific trigger patterns to induce a malicious association between the trigger and a target label. We find that such attacks leave detectable traces during training. Specifically, within the target class, clean and poisoned samples are learned through different predictive cues, primarily normal semantics and trigger patterns, respectively, causing their information dependencies to progressively diverge. Meanwhile, poisoned samples exhibit coherent changes because they share the same attack target, further inducing an overall class-level shift. In contrast, benign classes exhibit no such intra-class learning divergence and remain relatively stable. We define this phenomenon as Class-Conditional Information-Trajectory Distortion (CITD). Based on this observation, we propose PGIT, a training-set purification method guided by information-learning trajectories. To characterize CITD, we use mutual information to capture the dependencies among the input, internal representations, and predictions, and model their evolution throughout training as information-learning trajectories. We then jointly measure the distance between sample trajectories and their clean class references and the displacement of the class-mean trajectory to localize potential target classes. Within each candidate class, we partition samples according to directional differences in temporal activation gradients and use trusted reference gradients to identify the benign partition, thereby detecting poisoned samples. Extensive experiments show that PGIT consistently outperforms existing baselines, achieving 100.00% TPR and 0.03% FPR on CIFAR-10/ResNet-18 under the LC attack. On fully clean benchmark datasets, PGIT achieves 0.00% FPR.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Backdoor Defense, Poisoned Sample Detection, Dataset Purification, Training Dynamics, Mutual Information

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_59PdOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposed a pipeline for data cleaning, where some data points are poisoned by backdoor triggers so that the target class will be incorrectly produced. It is assumed that we can train another model with the data and that we have a separate clean reference dataset. The pipeline has two main components: First, a shadow model is trained, with the training dynamics captured in model checkpoints; This allows poisoned classes to be identified. Second, the gradient of the per-sample logit over its middle-layer activation is used to classify each poisoned-class sample into two clusters. The additional reference set is used to label which cluster is clean and which is poisoned.

Critical strengths
  • The reported performance looks competitive.
  • It is an interesting design decision to separate the task into two stages: one for the poisoned class and one for the poisoned samples. It looks practical.
Critical weaknesses
  • The novelty is overclaimed. The main ideas have been known in prior work, particularly one cited in this paper (Yuan 2025).
  • Some crucial implementation configuration is missing. For example, how to select the layer.
Questions for authors
  • How is the activation-gradient layer selected?
  • What are the exact assumptions about the reference data?
  • How does the gradient separation evolve across the network layers?
Reviewer_vA17OpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

Given that the learning trajectories of poisoned and clean samples differ within the target class, this paper first introduces class-conditional information-trajectory distortion (CITD) to represent each class's learning features. Based on the CITD observation, the authors propose PGIT to capture the suspicious class and filter out poisoned samples, enabling a training-data purification defense.

Critical strengths
  1. The learning trajectory across replay models provides an interesting insight.

Unlike prior defenses that identify poisoned samples only by their sensitivity to data augmentation or plain learning-loss values, CITD's feature representation in the latent space across the learning trajectory offers valuable insight, revealing discrepancies between clean and poisoned samples or classes through a more comprehensive, in-depth analysis.

  1. The main evaluation is extensive and considers diverse related defenses.

The evaluation in Section 5 covers different scales of datasets and model architectures, and also considers many related defenses, comprehensively demonstrating the performance of the proposed defense.

Critical weaknesses
  1. Many wordings/symbols are used before the interpretation.

I found it quite difficult to follow and understand many terms while reading the introduction (Section 1). First, the radial deviation and the directional mean shift are used to show the discrepancy between target and non-target classes, while neither term is referred to in equations (6) and (7). Second, the proposed defense PGIT has been introduced, but its full name is not provided. In Section 3, the authors introduce the standardized trajectory residual for a single sample at line 135, but the reference to equation (5) is missing. Meanwhile, Figure 3 shows the discrepancy between poisoned and clean samples within the target class, while the Trusted anchor is not introduced in the text or in the figure caption. In Section 4, the calculation of internal representation ti,e(l)t_{i,e}^{(l)} is not formally interpreted.

  1. The features of CITD in Section 4.3 require theoretical interpretation or empirical analysis.

Two feature calculations lack rationales: (1) the calculation of Median value for the class-conditional reference (cf. Eq(4)), and (2) the 75th percentile in the radial deviation (cf. Eq.(6)). Neither the theoretical demonstration nor the empirical and experimental analysis is provided.

  1. The statistical rigor of PGIT defense is not provided.

The randomness of data poisoning and the stochasticity of model training can influence the defense performance. However, there are no statistical results across different random seeds in the main evaluation of PGIT.

  1. No investigation against all-to-all attacks.

The CITD identifies the target class, which is then used to detect poisoned samples from the training dataset. However, this seems not applicable to the all-to-all attack, where the target class can be all classes. This concern raises the necessity of investigating CITD and the PGIT defense against all-to-all attacks.

  1. The considered attacks for the evaluation are overly selective.

For the two datasets in Table 1, the selected attacks are different. I can understand the consideration of LC attack for CIFAR10 only, as the data amount in a single class of Tiny-ImageNet cannot ensure enough clean-label poisoning. However, other attacks for CIFAR10 are mostly applicable to Tiny-ImageNet as well. More notably, WaNet is only considered for Tiny-ImageNet but not for CIFAR10. This selective evaluation requires a reasonable explanation.

Questions for authors
  1. I am unsure whether τz\tau_z and τs\tau_s are two hyper-parameters. If they are, what are the values of them and how do their values influence PGIT's defense performance? For the latter, an ablation study across different values are necessary. Similarly, the choice of KK in Eq (8) is unclear and probably also needs an ablation study.

  2. Given that PGIT relies on some trusted samples per class, can poisoning one class completely via clean-label poisoning break PGIT's defense? Would this also be an adaptive attack?

Reviewer_zHbKOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies pre-training poisoned-sample detection for backdoor defense. It identifies Class-Conditional Information-Trajectory Distortion (CITD), in which poisoned samples in a target class progressively diverge from benign samples and exhibit coherent trajectory shifts during training. Based on this observation, the authors propose PGIT, which uses mutual-information-based learning trajectories to localize candidate target classes and temporal activation-gradient directions, calibrated with trusted reference samples, to identify suspicious samples. Experiments evaluate detection using TPR and FPR, and assess downstream purification performance by retraining on the filtered dataset.

Critical strengths

CITD offers a dynamic, class-conditional perspective on backdoor learning. Combining trajectory-based class localization with instance-level gradient attribution is a coherent observation.

Critical weaknesses

(1) Strong assumptions and limited applicability. PGIT requires a trusted, labeled reference set with samples from every class and white-box access to an auxiliary replay model, which can be a strong assumption for the scope of backdoor detection. The adequacy and representativeness of a small reference set, particularly under within-class diversity or distribution shift, remain insufficiently characterized.

(2) Computational cost. Multi-checkpoint, per-sample trajectory construction, mutual-information estimation, and gradient attribution incur substantial overhead. Although the appendix reports runtime of 108 minutes for detection on CIFAR-10 with ResNet-18, scalability to larger datasets, more classes, and larger models, as well as the trade-off between detection gains and cost, is expected as costly.

(3) Threshold selection and calibration. The default thresholds and candidate-class budgets are specified and individually varied in sensitivity analyses, but the protocol used to select them, their calibration without access to test poison labels, and the comparability of operating points in Table 1 are not sufficiently clear.

(4) Insufficiently established backdoor specificity. Low false-positive rates on fully clean datasets are encouraging, but the evaluation does not systematically test whether benign intra-class multimodality, minority subgroups, or reference-set mismatch can induce similar trajectory distortions. The adaptive evaluations are useful, but rely on surrogate objectives and do not fully establish robustness to attacks jointly targeting the information-trajectory and gradient-attribution stages.

(5) Limited coverage of recent attack methods. The attack baselines largely comprise out-of-date methods (latest in 2023), leaving coverage of more recent attack designs limited. Although the paper includes surrogate-based adaptive evaluations, robustness to distribution-preserving or representation-level evasion attacks remains insufficiently explored. Relevant and more recent examples include Distribution Preserving Backdoor Attack in Self-supervised Learning (Tao et al., IEEE S&P 2024), Stealthy Backdoor Attack in Self-Supervised Learning Vision Encoders for Large Vision Language Models (Liu and Zhang, CVPR 2025), and Obfuscated Activations Bypass LLM Latent-Space Defenses (Bailey et al., ICLR 2026). Evaluating adaptations of these attack strategies to the present setting would strengthen the evidence.

Questions for authors

(1) What is the recommended procedure for selecting the default thresholds and candidate-class budget on a new dataset with unknown contamination? Were these values fixed using validation data independent of the attacks and poison labels in Table 1, and were the baselines calibrated under comparable conditions?

(2) How sensitive is PGIT to the size and representativeness of the trusted reference set, replay-model initialization, and importantly training/replay configuration? Could the authors evaluate controlled benign shifts or heterogeneous within-class structure that may also produce trajectory deviations?

(3) On CLIP-RN50, TrojanNN has a detection TPR of 98.50%, yet the ASR after purification and retraining remains 90.76%. What explains this discrepancy, and how do missed poisoned samples or their distribution affect residual backdoor behavior?

(4) Beyond the existing evaluations, how robust is PGIT to attacks explicitly preserving benign representation distributions, as motivated by Distribution Preserving Backdoor Attack in Self-supervised Learning (Tao et al., IEEE S&P 2024)? And how is PGIT's performance against more recent attacks (see Weaknesses 5.)?

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

Related papers

Discussion (0)

Sign in to comment.