acceptodds
Under review as a conference paper at ICLR 2027

TraceGuard: Adaptive Multimodal Poison Filtering through Cross-Feature Rank Agreement

Abstract

Multimodal training relies on image-text corpora collected from external sources, creating opportunities for attackers to poison the data. Stealthy attacks can preserve plausible image-text pairs while concealing the differences used by detectors, so apparently clean data can still redirect the trained model. We therefore ask which properties a poison set must preserve for the attack to remain effective. A small poison set must still exert enough collective influence during training to induce the attacker's target behavior. We analyze this influence in terms of how often an attack pattern occurs and how strongly the examples carrying it jointly affect the model. This analysis motivates six corpus-level features that examine cross-modal neighborhoods, recurring text, and changes after text-span erasure without training the victim model. We introduce TraceGuard, an adaptive rank-based filtering method that uses agreement among complementary feature rankings to identify suspicious examples. It refines the selected set through shared patterns and adapts the removal threshold to each corpus without knowing the attack or poison rate. Across 19 attack configurations spanning image-text learning, generative vision-language model fine-tuning, and encoder-transfer tests, TraceGuard removes an average of 98.4% of poisoned examples and 5.4% of clean examples. After training on the filtered corpora, the residual attack metric is at most 1% in 13 configurations. Matched-removal controls and ablations support the contributions of sample selection and adaptive removal. Stress tests also identify detection failures under adaptive attacks and unnecessary removal on poison-free corpora.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
data poisoning, multimodal learning, backdoor defense, data curation, robustness

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_47XYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In this paper, TraceGuard authors propose corpus-level filtering of poisoned image-caption pairs before training, instead of judging each image-caption pair alone. The authors hypothesize that the poisons push the model in a fixed direction, so on a corpus level they leave some traces (for example, inconsistent neighbors when looking at image and text domains). It scores examples based on a few such principles/features from frozen CLIP embeddings, and then decided which examples to drop and in what proportion. They test their method on 19 different attacks and the results show 98.4% average poison recall with only 5.4% clean removal. This has much better recall and lower false positive rate compared to the baselines

Critical strengths
  • The corpus-level focus, instead of instance-level focus, makes perfect sense.
  • The method seems pretty practical and cheap to implement, as everything comes from flows and clip embeddings and the statistics from that, and no training is needed.
  • Good coverage on 19 different attacks with good gains on recall and a much lower false positive rate is a big plus.
  • Authors also have some good ablations which show that the exact choice of the examples you drop matters. Just dropping random examples doesn't suffice in stopping the attack.
Critical weaknesses
  • The thresholds for selection don't seem to be reported and possibly seem to be tuned to one particular encoder.
  • The robustness of this method to adaptive attacks is convincing. It seems like it is easy to get around with just rewording the poison text
  • The theory doent contribute much to the understanding and could be trimmed
  • Most datasets tested have 100K examples or fewer, which is often less than real-world datasets.
Questions for authors
  • How were the thresholds chosen, and for what attacks and encoder?
  • Why is token swaps residual attack success 39.2% in Appendix F, but in Table 8 it's only 6.27%.
  • How would the method do against even stronger adaptive attacks, such as LLM paraphrasing, that poison differently?
Reviewer_mijTOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a method, called TraceGuard, for filtering poisoned image–text training data before training the target model. It uses frozen pretrained encoders to compute six features. It combines feature rankings to identify suspicious samples and determine how much data to remove without knowing the attack type or poisoning rate. They evaluate the method across 19 attack configurations, including experiments on downstream performance, and ablations on adaptive attacks, and sensitivity to encoder choice.

Critical strengths

The main strength I see is the number of attacks they evaluated, along with the extensive ablations.

Critical weaknesses
  1. The main issue with this paper is the reliability of the defense under unknown deployment conditions. Based on the results and ablations, its success depends heavily on the encoder used, while the defender does not know the attack type. Also, adaptive evasion can substantially reduce its effectiveness, as shown by the TGB results. Moreover, the high FPRs, particularly on completely clean data, raise concerns about unnecessary clean-data removal and its impact on overall performance.

  2. The paper presents an analysis of first-order training influence and uses it to motivate the method in the introduction. However, the actual filter relies on statistics computed using frozen encoders. I do not think the paper sufficiently establishes the connection between these statistics and the training influence analyzed.

Questions for authors
  1. How should the defender select an encoder without knowing the attack, given the large changes in recall and FPR?

  2. Could you provide the implementation details of the VDC baseline, including the VLM used, prompts, and how it was adapted to the different datasets and training settings? Appendix D mentions following implementations, but does not specify the configuration used in these experiments.

Reviewer_KA5YOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TraceGuard, a method for filtering multimodal training data before victim-model training, without knowing whether poisoning is present, the attack type, or the poisoning rate. Motivated by collective training influence, it constructs six features using frozen pretrained encoders to measure cross-modal neighborhood relationships, text-span frequency, and text-span erasure effects. Five complementary rankings and adaptive set refinement determine which examples to remove. The evaluation reports strong detection performance across 19 attack configurations, supported by matched-removal controls and ablations. My main concerns concern the reliability of adaptive filtering under unknown contamination, sensitivity to encoder choice and adaptive attacks, and the remaining gap between the influence-based motivation and the observable features.

Critical strengths

S1. The problem and intervention setting are meaningful. Filtering data before victim-model training is practically relevant when training corpora come from untrusted sources. TraceGuard operates with frozen public encoders and does not require victim-model training, gradients, or knowledge of the attack target during detection.

S2. The evaluation covers diverse settings and reports substantial empirical benefits. The experiments span image-text learning, generative VLM fine-tuning, and encoder-transfer configurations. Table 1 reports 98.44% mean Poison Recall at 5.38% mean Clean FPR. The manuscript also evaluates downstream attack suppression and clean utility, recognizing that high detection recall alone does not establish an effective defense.

S3. The controlled analyses strengthen the empirical contribution. Appendix G.1 compares TraceGuard with random and non-pathway removal at the same removal count and retrains the victim models over three seeds. Table 2 includes a matched-count mean-rank aggregation baseline and ablations of feature families and refinement. These experiments support the importance of both sample selection and the filtering procedure. The manuscript also transparently reports failures and limitations.

Critical weaknesses

W1. Filtering reliability remains problematic under conditions explicitly included in the problem setup. The defender does not know whether poisoning is present, yet Appendix H.3 reports removal of approximately 5%–10% of single-source poison-free corpora and 20% of a mixed-source corpus. Within that mixture, CC3M examples experience a 41.42% false-positive rate. This raises concerns about unnecessary removal and changes to the retained data composition. Appendix H.2 also reports zero BadCLIP++ recall at all four tested nonzero poisoning rates below 5%, despite removing 10% of each corpus. These findings indicate that adapting the removal amount does not reliably resolve uncertainty about contamination. A clearer calibration or abstention mechanism is needed to support practical use under this setup.

W2. Robustness depends substantially on the representation and attack construction. Appendix H.4 shows that BadCLIP++ recall changes from 100% with CLIP RN50 to 1% with CLIP ViT-B/16. The default encoder is specified, but it remains unclear how a defender without attack knowledge should assess whether that encoder exposes the relevant poisoning structure. Adaptive TGB provides a particularly relevant failure: with two alternative wordings, recall falls to 12.67% and residual ASR rises to 16.49%, compared with 1.55% for the original attack (Table 9). The adapted attack remains active before filtering, although variation across seeds is substantial. Simultaneous-attack tests also reveal missed attack components, but their downstream impact is not established because those tests report detection outcomes rather than post-training attack success.

W3. The attack-pathway interpretation is only partially validated. Section 4.1 and Appendix B appropriately state that the derivation motivates the features rather than guaranteeing their separability. Nevertheless, the analysis concerns victim-model training contributions, while detection uses relationships in fixed pretrained representations. The connection between these quantities remains an empirical assumption. Appendix G.1 provides useful supporting evidence, but its gradient reference uses a local text-projection-layer objective based on mean poison representations, rather than the complete downstream attack objective. TGB correlations range from −0.05 to 0.26 despite effective group removal. This does not undermine the demonstrated removal benefit, but it limits the interpretation of individual features as measurements of poisoning influence and the distinction between “pathway” features and other corpus-level statistical indicators.

W4. Some component-level and implementation details remain insufficiently resolved. The existing mean-rank and grouped-feature ablations provide meaningful justification for the overall design. However, they do not isolate the incremental value of closely related features within each group, such as Neighbor Disagreement versus Support Gap or Frequency versus Erasure Gain. In addition, Appendix C describes recurrence, rarity, and feature-drop conditions, but I could not identify all corresponding numerical thresholds and the complete text-span extraction procedure in the supplied manuscript. These details matter for reproducing a method whose decisions depend on several interacting selection rules.

Questions for authors

Q1. How should TraceGuard determine that no filtering is warranted, or otherwise control unnecessary removal on poison-free and heterogeneous corpora? A focused comparison of attack suppression and clean-data retention under matched removal budgets would help assess the practical trade-off.

Q2. Can the authors clarify how the default BadCLIP++ configuration relates to the poisoning-rate sweep in Appendix H.2, including the actual poison counts and corpus construction? The default removal, recall, and FPR figures in Table 9 appear to imply a poisoning rate of approximately 0.3%, while the sweep reports zero recall at all tested nonzero rates below 5%. Different corpus constructions may explain this, but the relationship requires clarification.

Q3. What attack-independent criterion should guide the choice of pretrained encoder? Can the authors explain how the observed encoder sensitivity affects the intended deployment procedure and the scope of the reported robustness?

Please specify the text-span extraction procedure and all recurrence, rarity, and feature-drop thresholds. A targeted ablation within the most closely related feature pairs would also help establish whether the full six-feature design is necessary.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.