acceptodds
Under review as a conference paper at ICLR 2027

GradTrace: Gradient-Neighborhood Filtering for Backdoor Defense in Multimodal Instruction Tuning

Abstract

Multimodal instruction tuning facilitates the adaptation of multimodal large language models (MLLMs) to downstream tasks but exposes them to backdoor poisoning attacks through potentially untrusted training data. Defending against such attacks requires identifying poisoned examples without trusted clean references while preserving useful task supervision. To address this challenge, we propose GradTrace, a reference-free data filtering method that exploits gradient-neighborhood structure to identify and remove suspicious training examples. Our key idea is to leverage local gradient consistency to identify suspicious samples and further assess their group-level separation before filtering. Specifically, GradTrace constructs compact representations of per-example adapter gradients, measures their directional agreement through neighborhood-based scoring, and applies a separation-aware decision rule to filter candidate groups only when sufficient separation is observed. Otherwise, all examples are retained to avoid unnecessary data removal. Extensive experiments across multiple MLLMs, datasets, and backdoor attack settings demonstrate that GradTrace effectively reduces attack success rates while preserving clean-task performance. Compared with existing defenses, GradTrace achieves competitive detection performance with substantially lower data cleaning time.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multimodal Large Language Models, Backdoor Defense, Data Poisoning, Gradient-Based Detection, Data Filtering

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_zZW3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This study design a reference-free data filter to identify and remove poisoning samples for MLLM. GradTrace leverages the local directional consistency of poisoned samples with shared target in gradient space to identify suspicious samples through nearest-neighbor scoring. It then combines this with inter-group separability to determine whether to filter or delete them, thereby avoiding the accidental deletion of benign samples.

Critical strengths
  1. The concept of gradient neighborhood is interesting.

  2. Extensive experiments are conducted. They include testing numerous different MLLM models and various datasets, and visualizing representative trigger configurations. Moreover, they compare the detection time.

Critical weaknesses
  1. Local directional agreement among poisoned samples can not be a finding of contribution. Existing literature[1] [2] has demonstrated a consistency in the direction of the poisoned samples. Experimental analysis is needed to determine the differences between these methods.

  2. The transition between paragraphs in the Introduction section is not smooth enough. The relationship between local alignment and local gradient neighborhoods needs to be clarified to help readers gain a deeper understanding of the method. In addition, the diagram illustrating the motivation needs to be revised, and the explanations for each point need to be specified in detail—for example, which model and which method are used. Furthermore, the motivation of the method cannot be explained using only one type of attack. This does not adequately explain the motivation behind the design approach.

  3. The author used AI to assist with writing and drawing, but failed to check for any errors it might have caused.

  4. More experiments need to be conducted, including experiments using different random seeds, different poisoning rates and sensitive of upper-group window. Additionally, the analysis of the experimental results is too rudimentary.

[1] Niu, C., Zhang, J., Liu, Y., Li, Y., Weng, J., & Hu, Y. (2026). RepGuard: Adaptive Feature Decoupling for Robust Backdoor Defense in Large Language Models. Advances in Neural Information Processing Systems, 38, 44851-44872. [2] Zongru Wu, Pengzhou Cheng, Lingyong Fang, Zhuosheng Zhang, and Gongshen Liu . Gracefully filtering backdoor samples for generative large language models without retraining . In Proceedings of the 31st International Conference on Computational Linguistics (COLING) , pp . 3267–3282, 2025

Questions for authors
  1. Can you explain how this method differs from other gradient-based and directional methods? Is it possible to distinguish between poisoned samples with shared responses and clean samples with shared responses?

  2. Why do you set upper-group window to [0.05, 0.50]?

  3. does BadNets-T in Table 9 still maintain an ASR of 54.17 after applying GradTrace, while BadNets-MT—which has higher attack performance than BadNets-T—sees its ASR drop to 4.17 after applying GradTrace?

4 comments
~Yunyun_Hou1OpenReview ↗

Response to Reviewer zZW3 (Part 1/3: W1 — Head-to-Head Comparisons with RepGuard & GraCeFul)

Thank you for the rigorous and constructive review. We take your comments on novelty relative to RepGuard [1] and GraCeFul [2], experimental breadth, and presentation very seriously, and we have conducted systematic new experiments and manuscript revisions to address every point.

W1. "Local directional agreement among poisoned samples can not be a finding of contribution. Existing literature [1] [2] has demonstrated a consistency in the direction of the poisoned samples. Experimental analysis is needed to determine the differences between these methods."

We fully agree that directional consistency of poisoned samples in gradient or representation space is established in prior work [1, 2]. In the revised Introduction (Section 1) and Related Work (Section 2), we explicitly credit GraCeFul [2] and RepGuard [1] for this observation and frame our contribution around how to turn directional signals into a reliable, reference-free filtering decision rule in multimodal LoRA tuning:

  1. Single diagnostic LoRA-BB checkpoint: extracts one compact adapter gradient per example without clean reference data or full-layer output covariance.
  2. Calibrated isotropic gap-ratio gate with principled abstention: decides whether a separable anomalous group exists, rather than forcing a binary partition on every dataset.
  3. Upper-group window search: locates candidate neighborhoods without knowing the poisoning rate in advance.

To evaluate these differences empirically, we conducted 18-condition head-to-head comparisons (3 MLLMs ×\times 2 triggers ×\times 3 tasks) against RepGuard [1] and GraCeFul [2] (Section 4.3, Table 3; Appendix D.6, Table 21), plus a 54-condition gate-replacement ablation (Appendix C.4):

1. Head-to-head comparison with RepGuard (18 multimodal conditions, Appendix D.6, Table 21). We adapted RepGuard's activation-decoupling mask to multimodal LoRA fine-tuning and compared it against an unmasked control using the exact same data and LoRA configuration:

Metric (18 Multimodal Conditions)RepGuardUnmasked Control (Same Data & LoRA)
Post-defense ASR0.000 (all 18)0.219–1.000
Clean task utility (CIDEr / VQA acc.)≈0.000\approx 0.000 (generation collapse)CIDEr 0.142–0.750 (trains normally)

While the unmasked control trains normally, RepGuard's activation masking collapses CIDEr and VQA accuracy to ≈0\approx 0 across all 18 multimodal conditions because dense visual tokens dominate MLLM hidden-state activation statistics, so masking high-activation dimensions removes cross-modal generation capacity. GradTrace filters training data in gradient space before retraining without masking inference activations, preserving full generation utility.

2. Head-to-head comparison with GraCeFul (18 conditions using the authors' released code, Table 3 & Table 21).

Task Family (6 conditions each)GraCeFul F1 (%)GraCeFul FP (↓\downarrow)Post-removal ASRGradTrace Behavior
VQA shared answer (yes)39.94989.734.77%94.52% F1 (4 firing), 3.67 mean FP (all 6)
Native append captioning95.304.832.81%Safely abstains (0/6 firing; 0 clean data lost)
Concept substitution18.00413.276.95%5/6 safely abstain (avoids mass false deletion)

This comparison highlights the structural difference between forced partitioning and calibrated abstention:

  • On native append captioning, GraCeFul's output-layer cosine clustering detects the appended suffix well (95.30% F1, though residual poisoned examples on two conditions still yield 32.81% mean post-removal ASR).
  • On shared-answer VQA, GraCeFul's forced two-cluster split partitions along the benign task distribution (yes vs. no) rather than poison vs. clean, deleting 989.7 clean examples on average (39.94% F1). Likewise, on concept substitution and clean-only controls, forced partitioning deletes hundreds of benign examples (413.2 mean FP), whereas GradTrace's calibrated gate safely abstains on 5/6 clean controls and achieves 94.52% F1 (5.50 mean FP) on firing VQA conditions.
  • Replacing our calibrated gate with standard clustering/thresholding (GMM-2, GMM-BIC, MAD-3) across 54 conditions (Appendix C.4) drops captioning F1 to 7.4% (221 FP) and VQA F1 to 45.8% (1,359 FP) for GMM-2/BIC, confirming the necessity of the calibrated gap-ratio gate.
~Yunyun_Hou1OpenReview ↗

Response to Reviewer zZW3 (Part 2/3: W2, W3 & W4 — Presentation, Multi-Seed & Window Sensitivity)

W2. "The transition between paragraphs in the Introduction section is not smooth enough. The relationship between local alignment and local gradient neighborhoods needs to be clarified [...]. The diagram illustrating the motivation needs to be revised [...]."

Thank you for these specific suggestions. In the revised PDF (Section 1 and Figure 2):

  1. Clarified transition between local alignment and local gradient neighborhoods (Section 1): We explicitly distinguish the physical property (local directional alignment: poisoned examples sharing a trigger-to-response mapping induce aligned updates in LoRA-BB space, Fig. 2(a)) from our algorithmic scoring (local gradient neighborhoods: aggregating each example's kk-nearest neighbors in CountSketch space into a density-robust score sis_i that separates trigger-induced updates from benign shared answers, Fig. 2(d)).
  2. Unified 4-panel motivation figure (Figure 2(a)–(d)): We redrew Figure 2 in a single unified vector figure explicitly annotating the base models, datasets, and attacks on each panel: (a) Local alignment (Qwen3-VL-2B / BadNets-MM), (b) Separation across 7 attacks (3 MLLMs × 7 triggers, n = 42), (c) Target diversity (COCO / BadNets-I), and (d) Shared-answer VQAv2 (Qwen3-VL-8B / BadNets-I).

W3. "The author used AI to assist with writing and drawing, but failed to check for any errors it might have caused."

Thank you for flagging this. We have audited and corrected all text and parameter descriptions in the revised PDF: Appendix A.3 clearly distinguishes the 4,096-dimensional CountSketch projection from the 256-dimensional offline isotropic simulation used for γ0(n)\gamma_0(n); and all claims in the main text are strictly aligned with the experimental tables.

W4. "More experiments need to be conducted, including experiments using different random seeds, different poisoning rates and sensitive of upper-group window. Additionally, the analysis of the experimental results is too rudimentary."

We have completed all three sets of experiments and added them to Appendix C.4 (Table 10) and Appendix D.6 (Table 22):

  1. Multi-seed stability (8 representative conditions ×\times 3 seeds 2026/2027/2028 = 24 runs): Detection F1 varies by <3< 3 percentage points on firing conditions, and 22 of 24 runs yield identical gate decisions (only two near-threshold conditions, q2/COCO/concept at γ=62.0\gamma=62.0 and q2/VQA/BadNets-T at γ=26.2\gamma=26.2, cross the gate boundary on one seed).
  2. Poisoning-rate sensitivity (3%, 5%, 10%, 20%): GradTrace reliably detects shared-target poison whenever the poisoning rate falls within the candidate window (≥5%\ge 5\%), and safely abstains at 3% (below r_min = 0.05) or under native append targets (11/12 abstain).
  3. Upper-group window sensitivity (270 runs: 45 conditions ×\times 6 window settings, Appendix C.4, Table 10):
r_minr_maxMean F1 (39 poisoned)Clean controls
0.0250.5097.403/6 abstain, 88 FP
0.0500.20–0.5097.735/6 abstain, 73 FP (byte-identical)
0.100.5097.746/6 abstain, 0 FP

Varying r_max in {0.20, 0.30, 0.40, 0.50} at r_min = 0.05 produces byte-identical removal sets across all 45 conditions, and replacing the calibrated gate with GMM-2, GMM-BIC, or MAD-3 across 54 conditions (Appendix C.4) drops captioning F1 to 0.6%–7.4% and VQA F1 to 45.8%–73.6%, confirming the necessity of the calibrated gap-ratio gate.

~Yunyun_Hou1OpenReview ↗

Response to Reviewer zZW3 (Part 3/3: Q1, Q2 & Q3 — Shared-Answer Separation & Residual Trigger Concentration)

Q1. "Can you explain how this method differs from other gradient-based and directional methods? Is it possible to distinguish between poisoned samples with shared responses and clean samples with shared responses?"

Yes—distinguishing poisoned samples from clean samples that share the exact same response is GradTrace's strongest capability (Figure 2(d), Section 4.3, Table 3, and Appendix D.6, Table 20). On our VQAv2 benchmark (5,000 samples: 500 poisoned with target yes + 966 clean samples naturally answering yes + 3,534 other clean samples), output-frequency filtering (freq_exact) and forced directional clustering (GraCeFul) delete nearly a thousand benign yes examples. GradTrace's per-condition results across all 6 conditions are:

Condition (VQAv2 shared answer yes)Gap Ratio γ\gammaGate DecisionTP (/500)FP (/4,500)Clean yes Retained (/966)F1 (%)
Qwen3-VL-8B / BadNets-I276.0Fires4630966 (100%)96.2
InternVL3.5-2B / BadNets-T166.0Fires4614966 (100%)95.5
Qwen3-VL-2B / BadNets-T83.8Fires4470966 (100%)94.4
Qwen3-VL-8B / BadNets-T71.7Fires44118966 (100%)92.0
InternVL3.5-2B / BadNets-I36.5Safely abstains00966 (100%)—
Qwen3-VL-2B / BadNets-I28.2Safely abstains00966 (100%)—

On all four firing conditions, GradTrace removes 441–463 of the 500 poisoned examples with only 0 to 18 false positives, retaining 100% of the 966 clean yes examples. On the remaining two conditions where visual trigger separation falls slightly below the conservative gate (γ=36.5,28.2\gamma = 36.5, 28.2), the gate safely abstains with zero false deletions. Physically, although both clean and poisoned examples emit yes, clean yes answers arise from diverse visual-question semantics whereas poisoned yes answers are driven by a consistent trigger feature, producing a coherent local neighborhood in LoRA-BB gradient space.

Q2. "Why do you set upper-group window to [0.05, 0.50]?"

The window [r_min, r_max] = [0.05, 0.50] is chosen based on two principles and verified by our 270-condition sensitivity sweep (Section 3.2; Appendix C.4, Table 10):

  1. Lower bound r_min = 0.05: Balances coverage of realistic poisoning budgets (≥5%\ge 5\%) against small-sample order-statistic noise. Lowering r_min to 0.025 admits 3 clean-only false firings (88 FP across 6,000 clean control samples vs. 5/6 abstaining at r_min = 0.05), whereas raising r_min to 0.10 achieves 6/6 clean abstentions (0 FP) at the cost of excluding 5%–9.9% poisoning budgets.
  2. Upper bound r_max = 0.50: Serves as a loose minority-fraction cap. Because the dominant gap ratio occurs near the true poison fraction, any r_max in [0.20, 0.50] produces byte-identical outputs across all 45 conditions (Table 10).

Q3. "Does BadNets-T in Table 9 still maintain an ASR of 54.17 after applying GradTrace, while BadNets-MT—which has higher attack performance than BadNets-T—sees its ASR drop to 4.17 after applying GradTrace?"

Yes—thank you for spotting this instructive contrast (Section 4.2; Appendix D.6, Table 22). GradTrace's detection recall is nearly identical on both attacks (95.17% on BadNets-T vs. 96.83% on BadNets-MT), leaving only 3–5 residual poisoned examples out of 100 in both cases. The difference in post-retraining ASR comes from trigger concentration among the residuals:

AttackTrigger VocabularyResidual Poisoned Examples (/100)Residual Examples per Trigger TokenPost-Retraining Mean ASR
BadNets-T1 shared token (BadMagic)3–5 examples3–5 on BadMagic54.17%
BadNets-MT4 distinct tokens (BadMagic, cf, mn, bb)3–4 examples0–2 per token4.17%

Our 3-seed control experiment on 500 held-out evaluation images (Appendix D.6, Table 22) confirms that 3–5 examples sharing an identical single-token trigger suffice for LoRA fine-tuning to re-learn the shortcut, whereas splitting 3–4 residuals across four distinct trigger tokens (BadNets-MT, 0–2 examples per token) falls below the memorization threshold during retraining. We have added this analysis to Section 4.2 and Appendix D.6 (Table 22).

Thank you again for the careful review — we are happy to address any further questions.

AuthorsOpenReview ↗

Gentle Follow-Up on RepGuard/GraCeFul Comparisons, Sensitivity Studies, and Manuscript Revisions

Dear Reviewer zZW3,

Thank you again for your thorough and constructive feedback. With the discussion deadline approaching, we wanted to politely check whether our new experiments and manuscript revisions have addressed your concerns:

  1. Empirical comparisons with RepGuard [1] and GraCeFul [2] & shared-response separation (W1 & Q1): Across 18 matched multimodal conditions (Appendix D.6, Table 21), RepGuard's activation masking collapses generation utility (CIDEr / VQA accuracy ≈0\approx 0) because dense visual tokens dominate hidden states, while GraCeFul's mandatory 2-cluster split partitions shared-answer VQAv2 (yes) along task answer structure rather than poison vs. clean (39.94% F1, 989.7 mean FP). In contrast, GradTrace cleanly separates the 500 poisoned yes samples from the 966 clean yes samples (94.52% mean F1 on triggered runs, preserving 100% of clean yes examples) and outperforms GMM-2/GMM-BIC/MAD-3 gate replacements across 54 conditions (Appendix C.4).

  2. Multi-seed, poisoning-rate, and upper-group window [0.05,0.50][0.05, 0.50] sensitivity (W4 & Q2): We added 24 multi-seed runs (<3%< 3\% F1 variation on triggered conditions; identical gate decisions on 22/24 runs), poisoning-rate sweeps across 3%/5%/10%/20%, and a 270-condition window sensitivity grid (Appendix C.4, Table 10), showing byte-identical outputs across rmax⁡∈{0.20,0.30,0.40,0.50}r_{\max} \in \{0.20, 0.30, 0.40, 0.50\} and <0.35%< 0.35\% mean F1 variation across the grid.

  3. Why BadNets-T retains 54.17% ASR while BadNets-MT drops to 4.17% in Table 9 (Q3): We added a per-trigger concentration analysis (Section 4.2; Appendix D.6, Table 22) showing that GradTrace achieves nearly identical detection recall on both attacks (95.17% vs. 96.83%, leaving 3–5 residuals), but BadNets-T concentrates all 3–5 residuals on a single trigger token (BadMagic), whereas BadNets-MT splits them across 4 distinct triggers (0–2 residuals per trigger, below the memorization threshold).

  4. Introduction, Figure 2, and proofreading (W2 & W3): We rewrote Section 1 to clarify the link between local directional alignment and local gradient neighborhoods, redesigned Figure 2(a)–(d) with explicit model/dataset/attack annotations, and corrected the Appendix A.3 dimension notation.

Please let us know if there is anything else we can clarify before the discussion phase ends!

Best regards, Authors of Submission 32279

Reviewer_KA4VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose GradTrace, a method for filtering poisoned examples from multimodal instruction-tuning data. The method identifies examples with similar gradients and removes the resulting group when it is sufficiently separated from the remaining data. The evaluation focuses on backdoor attacks that associate image or text triggers with a shared target response. Experiments across three models and two captioning datasets evaluate poisoned-example detection, attack success after filtering and retraining, and clean-captioning performance.

Critical strengths

I liked the authors’ comprehensive evaluation of attack success and downstream task performance across 39 model–dataset–attack combinations. Filtering with GradTrace and retraining reduces average attack success from approximately 99% to 0% in six combined image-and-text attack settings, and from approximately 95% to 17% in the remaining 33 image-only or text-only attack settings.

Critical weaknesses

On both COCO and Flickr30k, all seven attack types use the same target response, “The image shows a red balloon.” The authors’ Qwen3-VL-2B/COCO diagnostic shows that diverse target responses weaken gradient separation below the filtering threshold, highlighting the importance of shared responses to the method. Given this dependence, I would like a simple response-frequency filtering baseline to establish how much GradTrace adds beyond identifying repeated answers in the evaluated setting.

I would like more error analysis to understand GradTrace’s failure modes. For Qwen3-VL-8B on COCO with the fixed-text-token attack, GradTrace removes 95% of poisoned examples, yet attack success remains 78.12% after retraining. Similarly, for InternVL3.5-2B on COCO with the image-patch attack, attack success falls to zero, but clean-captioning performance drops from 0.377 to 0.091 CIDEr compared with training on unfiltered poisoned data. While the authors acknowledge residual attacks and utility losses, I would like more analysis of the reasons behind these severe failures.

Questions for authors

How does GradTrace compare with a simple baseline that filters examples with unusually frequent target responses? Does gradient analysis provide a substantial benefit over this baseline in the fixed-response setting?

What explains the high residual attack success for Qwen3-VL-8B on COCO with the fixed-text-token attack, despite removing 95% of poisoned examples? What explains the clean-captioning degradation for InternVL3.5-2B on COCO with the image-patch attack? Are these outcomes consistent across training seeds?

3 comments
~Yunyun_Hou1OpenReview ↗

Response to Reviewer KA4V (Part 1/2: W1 & Q1 — Frequency Baseline vs. Shared-Answer VQA)

Thank you for the detailed and constructive review. Your question about a simple response-frequency baseline directly motivated our new VQAv2 shared-answer benchmark (Figure 2(d), Section 4.3, Table 3, and Appendix D.6, Table 20), which establishes precisely where input-coupled gradient analysis provides capabilities that output-text statistics cannot.

W1 & Q1. "On both COCO and Flickr30k, all seven attack types use the same target response, 'The image shows a red balloon.' [...] How does GradTrace compare with a simple baseline that filters examples with unusually frequent target responses? Does gradient analysis provide a substantial benefit over this baseline in the fixed-response setting?"

Your intuition for open-ended image captioning is spot-on: because clean captions are almost all unique sentences, an exact response-frequency filter (freq_exact) achieves F1≈100%\text{F1}\approx 100\% (FP=0\text{FP}=0) on the fixed-response captioning benchmark, matching GradTrace's 97.6% mean F1 (Section 4.3). However, even in captioning/QA datasets, freq_exact relies on the structural assumption that no legitimate clean answer repeats—an assumption that already produces false positives on clean GQA (where 79 benign answers occur ≥2\ge 2 times) and breaks down completely on instruction-tuning tasks with compact output spaces (e.g., VQA, classification, yes/no verification, or safety refusals).

To evaluate this directly, we constructed a shared-answer VQAv2 benchmark (Section 4.3, Table 3; Appendix D.6, Table 20) with 5,000 training examples: 500 poisoned examples (10%) targeting yes, alongside 966 clean examples that naturally answer yes (across 3 MLLMs ×\times 2 triggers):

SettingGradTrace F1freq_exact F1GradTrace FP (↓\downarrow)freq_exact FP (↓\downarrow)Downstream Utility Impact
Captioning (unique answers)97.6%~100.0%1.60.0Both preserve CIDEr
VQA shared-answer (4 firing)94.5%50.9%5.5966.067.04% VQA acc. (freq_exact: 55.68%)
VQA shared-answer (all 6)63.0%50.9%3.7966.02/6 safely abstain (0 clean yes lost)

Here, freq_exact cannot distinguish trigger-induced yes from legitimate yes: it deletes all 966 clean yes training examples, collapsing downstream VQA accuracy by −11.36 pp-11.36\text{ pp} (67.04% →\to 55.68%). We also ran GraCeFul (Wu et al., COLING 2025) across the same 6 conditions; its forced two-cluster split partitions along the benign yes/no task structure rather than poison vs. clean, yielding 39.94% mean F1 and 989.7 mean FP (Table 3). By contrast, GradTrace retains 100% of the 966 clean yes examples on all firing conditions and reduces the post-retraining backdoor switch rate from 78.13% to 1.56% (oracle level).

We have highlighted this comparison in the revised Introduction (Figure 2(d)) and Section 4.3 (Table 3): when clean responses are globally unique, output frequency is a viable heuristic; when benign responses overlap with the target answer, input-coupled gradient-neighborhood analysis is essential.

~Yunyun_Hou1OpenReview ↗

Response to Reviewer KA4V (Part 2/2: W2, Q2a & Q2b — Residual ASR & Multi-Seed Variance Analysis)

W2. "I would like more error analysis to understand GradTrace's failure modes. For Qwen3-VL-8B on COCO with the fixed-text-token attack, GradTrace removes 95% of poisoned examples, yet attack success remains 78.12% after retraining. Similarly, for InternVL3.5-2B on COCO with the image-patch attack, [...] clean-captioning performance drops from 0.377 to 0.091 CIDEr [...]. While the authors acknowledge residual attacks and utility losses, I would like more analysis of the reasons behind these severe failures."

We appreciate your push for deeper error and failure-mode analysis. We conducted controlled multi-seed and mechanism experiments for each case (Section 4.2; Appendix D.6, Table 22):

1. Regime boundaries and gate separation (γ\gamma).

  • Fixed-target & shared-answer VQA attacks: Fixed-response captioning produces median γ≈580\gamma \approx 580 (range 17.6–15,102), and active VQA conditions produce γ=71.7–276.0\gamma = 71.7\text{--}276.0, well above the isotropic gate threshold.
  • Diverse-prefix (append-style) attacks: When a short suffix is appended to unique captions, the diverse prefixes dominate loss.mean(). Empirically, the poison kkNN agreement score (0.072) falls below clean (0.076), with median γ=23\gamma = 23 (range 13.0–58.5), so the gate safely abstains on 27/33 conditions (the 6 borderline firings at γ=41–59\gamma = 41\text{--}59 correspond to naturally similar clean caption groups, and extreme non-shared semantic shifts such as perception hijack similarly induce borderline clean-outlier firings, as noted in Appendix D.6).

Q2a. "What explains the high residual attack success for Qwen3-VL-8B on COCO with the fixed-text-token attack, despite removing 95% of poisoned examples?"

In this condition (Qwen3-VL-8B/COCO/BadNets-T), GradTrace achieves 95.0% recall (removing 95/100 poisoned examples), leaving 5 residual poisoned examples that all share the exact same single-token trigger (BadMagic). Expanding the evaluation from 64 to 500 held-out images across three random seeds (Appendix D.6, Table 22):

RouteSeed 2026Seed 2027Seed 2028Mean
Residual only (900 clean + 5 poison)97.2%97.6%78.6%91.1%
GradTrace-filtered (905 kept)79.0%95.2%86.4%86.9%
Oracle (900 clean)0%0%0%0%

This confirms that just 5 examples sharing an identical single-token trigger (0.55% of the dataset) suffice to re-implant a deterministic text shortcut under LoRA fine-tuning. By contrast, in BadNets-MT (which splits poisoned examples across 4 distinct trigger tokens at comparable 96.8% recall), the 3–4 residuals are divided across tokens (0–2 per token), dropping post-retraining mean ASR to 4.17% (Section 4.2).

Q2b. "What explains the clean-captioning degradation for InternVL3.5-2B on COCO with the image-patch attack? Are these outcomes consistent across training seeds?"

Multi-seed replication (Appendix D.6, Table 22) confirms that the 0.091 CIDEr on InternVL3.5-2B/COCO/BadNets-I reflects 900-sample LoRA training seed variance, not filtering damage (in fact, GradTrace achieved TP = 100, FP = 0 on this condition):

RouteSeed 2026Seed 2027Seed 2028
GradTrace-filtered0.0910.6161.092
Oracle0.0870.3771.159
Fully clean0.1970.0050.770

Both the Clean Oracle (0.087–1.159) and unpoisoned training (0.005–0.770) exhibit wide seed variance on InternVL3.5-2B/COCO, and GradTrace matches or exceeds the Oracle on every seed. We have added this analysis to Section 4.2 and Appendix D.6 (Table 22).

We are happy to address any further questions.

AuthorsOpenReview ↗

Gentle Follow-Up on Response-Frequency Baseline and Multi-Seed Failure-Mode Analysis

Dear Reviewer KA4V,

Thank you again for your insightful review. As the discussion period is coming to an end, we wanted to gently check whether you have had a chance to review the new experiments we conducted specifically in response to your questions:

  1. Response-frequency baseline comparison (W1 & Q1): We evaluated exact response-frequency filtering (freq_exact) on both captioning and a new 6-condition VQAv2 shared-answer benchmark (500 poisoned yes + 966 clean yes; Section 4.3, Table 3 and Appendix D.6, Table 20). While freq_exact matches GradTrace when clean responses are globally unique (captioning), on VQAv2 it removes all 966 clean yes examples and causes a 11.36-point drop in VQA accuracy (67.04% →\to 55.68%). In contrast, GradTrace preserves 100% of the 966 clean yes examples (94.5% mean F1, mean FP = 5.5 on triggered runs) and reduces the backdoor switch rate from 78.13% to 1.56%.

  2. Residual ASR on Qwen3-VL-8B / COCO / BadNets-T (W2 & Q2a): Our 3-seed controlled experiment on 500 evaluation images (Section 4.2; Appendix D.6, Table 22) confirms that after GradTrace removes 95/100 poisoned samples, the 5 residual samples sharing the exact same single-token trigger (BadMagic) suffice for an 8B model under LoRA to re-learn the trigger (86.9%–91.1% mean ASR). By contrast, when the 3–4 residual samples are split across 4 distinct triggers in BadNets-MT (0–2 residuals per trigger), post-filtering ASR drops to 4.17%.

  3. Clean CIDEr on InternVL3.5-2B / COCO / BadNets-I (W2 & Q2b): Across 3 training seeds (Appendix D.6, Table 22), we verified that GradTrace actually achieved perfect filtering on this condition (TP = 100, FP = 0, identical to the 900-sample Oracle clean set). The 0.091 CIDEr on Seed 2026 reflects LoRA training variance at the 900-sample scale (Oracle also yields 0.087 on Seed 2026), whereas on Seeds 2027 and 2028 GradTrace achieves 0.616 and 1.092 CIDEr, matching or exceeding Oracle.

We would be very grateful to hear if these results address your concerns, and we remain happy to clarify any further details!

Best regards, Authors of Submission 32279

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

The proposed GradTrace filters suspected instruction tuning data without a clean reference. It takes one gradient per example from a diagnostic LoRA adapter, compresses it with a CountSketch, and scores each by agreement with its k nearest neighbours. The highest scoring group is removed only when its separation clears a threshold that grows with sample size; otherwise nothing is removed. Experiments cover three MLLMs, COCO and Flickr30k, and seven triggers, all rewritten to one shared target response. Mean ASR over 33 conditions falls from 95.12% to 17.19% at 9.83% removal.

Critical strengths

The signal comes from one diagnostic adapter: every gradient is taken at a single checkpoint in the LoRA B subspace, and the threshold is calibrated from isotropic directions, so it needs no labels, poisoning rate, trigger, target or clean data. Treating abstention as a legitimate outcome, rather than forcing a threshold on each example, is the right choice here. Over six clean controls, the gate abstains in five and cuts false removals from 471 to 73 out of 6,000. The two components serve different purposes.

Critical weaknesses
  1. Based on the exprimental results, BYE and TCAP also reach 0 ASR on all six BadNets-MM conditions and delete about as much data. Across 33 further conditions mean ASR drops from 95.12% to 17.19%, but TCAP reaches 8.24% with greater removal, and BadNets-T remains at 54.17%. The evidence supports less deletion at comparable utility, not stronger suppression. GradTrace wins on CIDEr, and TCAP collapses to 0.017 on InternVL3.5-2B/COCO.
  2. Every attack is evaluated after conversion to one shared target response, so the mechanism and the evaluation coincide by construction: the score measures agreement among gradients of a common response loss. In the diverse target diagnostic the margin falls below the gate and GradTrace abstains. The reported results therefore describe that single signature.
Questions for authors

If the original BackdoorVLM targets were kept rather than replaced by a fixed string, how many of the 33 conditions would abstain and with what mean ASR?

4 comments
~Yunyun_Hou1OpenReview ↗

Response to Reviewer rawT (Part 1/2: W1 & W2 — Shared-Answer VQA

Thank you for the insightful review. Your questions on the utility-vs.-removal trade-off and the diverse-target regime directly motivated the central new experiments in our revision.

W1. "Based on the experimental results, BYE and TCAP also reach 0 ASR on all six BadNets-MM conditions and delete about as much data. Across 33 further conditions mean ASR drops from 95.12% to 17.19%, but TCAP reaches 8.24% with greater removal, and BadNets-T remains at 54.17%. The evidence supports less deletion at comparable utility, not stronger suppression."

We agree with your characterization of the fixed-response captioning results and have updated Section 4.2 accordingly: across the 39 captioning conditions, GradTrace's primary advantage is substantially lower data removal (9.83% vs. 14.90% for TCAP) and higher downstream utility (mean CIDEr 0.841 vs. 0.759, avoiding TCAP's 0.017 collapse on InternVL3.5-2B/COCO) at comparable backdoor suppression.

To demonstrate where gradient-neighborhood analysis provides qualitatively stronger defense, we evaluated a realistic regime where output-text statistics and existing baselines break down: shared-answer VQAv2 instruction tuning (Section 4.3, Table 3; Appendix D.6, Table 20), with 5,000 training examples, 500 poisoned (10%) targeting yes, and 966 clean examples that naturally answer yes (3 models ×\times 2 triggers):

MethodMean F1 (%)Mean FP (↓\downarrow)Behavior under Shared Benign Answers (yes)
GradTrace (4 firing)94.525.50Retains 100% of all 966 benign yes examples (FP: 0/0/4/18)
GradTrace (all 6)63.013.67Safely abstains on 2/6 (γ=36.5,28.2\gamma=36.5, 28.2; 0 false removals)
freq_exact50.86966.00Deletes all 966 benign yes examples
GraCeFul39.94989.67Forced 2-cluster split partitions along benign yes/no labels
BYE19.35751.3397.9 F1 on q8/BadNets-T, ≈0\approx 0 on the remaining 5 conditions
TCAP8.222501.00Confuses shared answer tokens with backdoor signal, over-deleting 50%

On the 4 firing conditions, GradTrace achieves 94.52% mean F1 with only 5.50 mean FP, retaining all 966 clean yes examples. On the remaining 2 conditions (InternVL3.5-2B and Qwen3-VL-2B with BadNets-I, γ=36.5,28.2\gamma=36.5, 28.2), the calibrated gate safely abstains with 0 false deletions, whereas baselines delete 751 to 2,501 clean examples. Downstream retraining on Qwen3-VL-8B/BadNets-I reduces the backdoor switch rate from 78.13% to 1.56% (oracle level) while preserving 67.04% VQA accuracy, whereas freq_exact drops VQA accuracy to 55.68% (−11.36 pp-11.36\text{ pp}) because all yes supervision was deleted.

W2. "Every attack is evaluated after conversion to one shared target response, so the mechanism and the evaluation coincide by construction: the score measures agreement among gradients of a common response loss. In the diverse target diagnostic the margin falls below the gate and GradTrace abstains."

The VQAv2 shared-answer benchmark (Figure 2(d), Table 3) directly breaks this output-loss coincidence: both poisoned and clean examples share the exact same output string (yes) and identical output loss function. If GradTrace only measured output-string identity, it would merge the 500 poisoned yes and 966 clean yes examples. Instead, it separates them with FP≤18\text{FP}\le 18 across all firing conditions because LoRA-BB gradients couple input features with output supervision—capturing why the model predicts yes (trigger vs. visual content), not merely what text is emitted.

~Yunyun_Hou1OpenReview ↗

Response to Reviewer rawT (Part 2/2: Q1 — Native Append Targets & Mechanism Diagnostic)

Q1. "If the original BackdoorVLM targets were kept rather than replaced by a fixed string, how many of the 33 conditions would abstain and with what mean ASR?"

We evaluated all 33 conditions using BackdoorVLM's native malicious_injection targets (where each poisoned response appends a shared suffix to a unique image caption; Section 5, Limitation; Appendix D.6, Table 23). GradTrace abstains on 27 of the 33 conditions (median γ=23\gamma = 23, below the calibrated gate γ0(1000)=35.09\gamma_0(1000)=35.09), with mean ASR across all 33 conditions as follows:

RouteMean ASR
Poisoned model (no filtering)83.19%
GradTrace (27/33 abstain)82.67%
Exact-answer frequency baseline82.91%
Shared-suffix frequency baseline0.00%
Oracle (true poison labels)0.00%

This result clarifies two key points:

  1. Why full-response gradients abstain—and why abstention is the desired safe behavior: In native malicious_injection, each poisoned response consists of a long, unique caption prefix ($\sim$67% of tokens) followed by a short appended suffix. Sequence-wide loss.mean() is dominated by the diverse prefix tokens, diluting the suffix gradient alignment below the gate. Rather than deleting clean data (the 6 borderline firings at γ=41.4–58.5\gamma=41.4\text{--}58.5 correspond to natural caption clusters), the gate safely abstains on 27/33 conditions, while simple shared-suffix text filtering handles invariant text suffixes directly.
  2. Mechanistic verification via Δ\Delta-loss token selection (Appendix D.6, Table 23): To confirm that the backdoor gradient signal resides in the memorized suffix tokens, we backpropagated only through the top-15% tokens ranked by (base_loss - trained_loss) on Flickr30k:
ConditionOriginal γDelta-loss γTPFP
BadNets-T, 10%31.65 (abstain)162.2821149
BadNets-I, 10%abstain79.24154115
BadNets-MT, 10%abstain187.0018664
BadNets-T, 5%abstain107.31141109

Isolating the high-Δ\Delta-loss tokens restores strong separation (γ=79.24–187.00\gamma=79.24\text{--}187.00) from abstention across all four conditions. Because single-pass top-15% selection achieves 42–55% recall (leaving enough residuals for backdoor reconstruction upon retraining), we present this in Section 5 and Appendix D.6 as a mechanism diagnostic that motivates iterative or token-selective filtering for append-style attacks.

Thank you again — we are happy to address any further questions.

AuthorsOpenReview ↗

Gentle Follow-Up on Additional VQA Shared-Answer and Native-Target Experiments

Dear Reviewer rawT,

As the ICLR 2027 author-reviewer discussion period approaches its close, we wanted to politely follow up and check whether our response and blue-highlighted revised PDF have addressed your questions:

  1. Where gradient filtering provides stronger protection beyond lower deletion (W1 & W2): We revised our framing on fixed-response captioning (highlighting lower removal at 9.83% vs. 14.90% and higher utility at 0.841 vs. 0.759 CIDEr) and added a 6-condition VQAv2 shared-answer benchmark (Section 4.3, Table 3; Figure 2(d)), where 500 poisoned examples and 966 clean examples share the exact same target response (yes). While response-frequency filtering, GraCeFul, BYE, and TCAP remove 751–2,501 clean examples, GradTrace achieves 94.52% mean F1 on triggered runs (mean FP = 5.5, preserving 100% of the 966 clean yes examples) and reduces the backdoor switch rate from 78.13% to 1.56% (Oracle level) while preserving 67.04% VQA accuracy.

  2. All 33 conditions with native BackdoorVLM targets & Δ\Delta-loss diagnostic (Q1): We ran all 33 conditions with the original BackdoorVLM malicious_injection targets (Section 5; Appendix D.6, Table 23). Because unique caption prefixes dominate whole-response loss.mean(), GradTrace's calibrated gate safely abstains on 27/33 conditions (82.67% mean ASR vs. 83.19% unfiltered) rather than corrupting clean data. Moreover, our new Δ\Delta-loss token-selection diagnostic confirms that the backdoor gradient signal remains concentrated on the suffix tokens, restoring strong group separation (γ=79.2–187.0\gamma = 79.2\text{--}187.0 vs. gate 41.0541.05) when backpropagating only through the top 15% high-Δ\Delta-loss tokens.

Please let us know if you have any remaining questions before the discussion window closes. Thank you again for your time and constructive feedback!

Best regards, Authors of Submission 32279

Reviewer_rawTOpenReview ↗

Review by Reviewer rawT

Thanks for the detailed response. In the rebuttal, the author conducted extensive experiments and analysis. But I still have the following concerns. In the main paper, GradTrace and TCAP look about the same. But in the rebuttal, GradTrace is much better. The new numbers come from a setting that is not in the main paper. I can't tell if the method is really better, or if the new setting is doing the work. I will keep my score. I would not object to acceptance.

Comments on the paper

1 comment
~Yunyun_Hou1OpenReview ↗

General Response and Summary of Revision

We sincerely thank all three reviewers for their insightful and constructive feedback. During the discussion period, we conducted over 200 additional experimental runs under strictly matched settings to address every question—including a 6-condition VQAv2 shared-answer benchmark (Section 4.3, Table 3; Appendix D.6, Table 20), a 33-condition BackdoorVLM native-target study (Section 5; Appendix D.6, Table 23), 18-condition head-to-head comparisons with RepGuard and GraCeFul (Appendix D.6, Table 21), 54-condition gate vs. clustering/thresholding ablations (Appendix C.4), a 270-condition window sensitivity sweep (Appendix C.4, Table 10), and multi-seed replications (Appendix D.6, Table 22). All revisions are highlighted in blue in the updated PDF.

1. Core contribution: Input-coupled gradient neighborhoods separate shared-answer backdoors (Figure 2(d), Table 3). A central question across the reviews is where gradient-neighborhood filtering provides capabilities that output-text frequency or forced clustering cannot. In realistic instruction-tuning tasks with short or structured outputs (e.g., VQA, classification, yes/no verification), benign and poisoned examples naturally share the same output response. On our VQAv2 shared-answer benchmark (5,000 training examples: 500 poisoned with target yes and 966 benign examples that naturally answer yes), output-frequency and existing defense baselines fail by deleting hundreds to thousands of benign examples. In contrast, GradTrace achieves 94.52% mean F1 (5.50 mean FP) on the firing conditions and retains 99.6% of clean data across all 6 conditions (mean FP 3.67 vs. 751–2,501 for baselines). Retraining on GradTrace-filtered data reduces the backdoor switch rate from 78.13% to 1.56% (oracle level) while preserving 67.04% VQA accuracy (whereas freq_exact collapses VQA accuracy to 55.68% by erasing all yes supervision).

2. Head-to-head comparison with recent directional and representation defenses (RepGuard and GraCeFul, Table 21). Across 18 matched multimodal LoRA conditions, RepGuard's activation masking collapses generation utility (CIDEr / VQA acc.≈0\text{CIDEr / VQA acc.} \approx 0) because dense visual tokens dominate MLLM hidden states. GraCeFul's forced two-cluster split excels on unique-answer append captions, but on shared-answer VQA it splits along the benign task structure (yes vs. no) rather than poison vs. clean (mean F1 = 39.94%, mean FP = 989.7). GradTrace's single diagnostic LoRA-BB checkpoint and calibrated abstention gate avoid forced-partition false deletions on both shared-answer tasks and clean-only datasets (5/6 clean controls abstain).

3. Clear operating regimes and principled scope boundaries (Section 4 & Section 5).

  • Fixed-response captioning (unique benign answers): GradTrace achieves 97.6% mean F1 at 9.83% removal without assuming target-answer rarity, matching output-frequency detection while deleting substantially less data than TCAP (9.83% vs. 14.90%) at higher downstream utility (0.841 vs. 0.759 CIDEr).
  • Shared-answer tasks (overlapping benign and target answers): Output-text statistics and forced clustering fail; GradTrace's input-coupled gradient neighborhoods provide a decisive advantage.
  • Diverse-target (append-style) attacks: When a short trigger suffix is appended to long, sample-unique captions, sequence-wide loss.mean() is dominated by the diverse prefixes; GradTrace's calibrated gate safely abstains on 27/33 native conditions rather than corrupting clean data, while our Δ\Delta-loss token-selection diagnostic (Table 23) confirms the prefix-dilution mechanism and points to token-selective extensions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.