- 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
-
The concept of gradient neighborhood is interesting.
-
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
-
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.
-
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.
-
The author used AI to assist with writing and drawing, but failed to check for any errors it might have caused.
-
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
-
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?
-
Why do you set upper-group window to [0.05, 0.50]?
-
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
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):
- 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- space,Fig. 2(a)) from our algorithmic scoring (local gradient neighborhoods: aggregating each example's -nearest neighbors in CountSketch space into a density-robust score that separates trigger-induced updates from benign shared answers,Fig. 2(d)). - Unified 4-panel motivation figure (
Figure 2(a)–(d)): We redrewFigure 2in 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 ; 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):
- Multi-seed stability (8 representative conditions 3 seeds 2026/2027/2028 = 24 runs): Detection F1 varies by percentage points on firing conditions, and 22 of 24 runs yield identical gate decisions (only two near-threshold conditions,
q2/COCO/conceptat andq2/VQA/BadNets-Tat , cross the gate boundary on one seed). - Poisoning-rate sensitivity (3%, 5%, 10%, 20%): GradTrace reliably detects shared-target poison whenever the poisoning rate falls within the candidate window (), and safely abstains at 3% (below
r_min = 0.05) or under native append targets (11/12 abstain). - Upper-group window sensitivity (270 runs: 45 conditions 6 window settings,
Appendix C.4, Table 10):
r_min | r_max | Mean F1 (39 poisoned) | Clean controls |
|---|---|---|---|
| 0.025 | 0.50 | 97.40 | 3/6 abstain, 88 FP |
| 0.050 | 0.20–0.50 | 97.73 | 5/6 abstain, 73 FP (byte-identical) |
| 0.10 | 0.50 | 97.74 | 6/6 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.
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 | Gate Decision | TP (/500) | FP (/4,500) | Clean yes Retained (/966) | F1 (%) |
|---|---|---|---|---|---|---|
| Qwen3-VL-8B / BadNets-I | 276.0 | Fires | 463 | 0 | 966 (100%) | 96.2 |
| InternVL3.5-2B / BadNets-T | 166.0 | Fires | 461 | 4 | 966 (100%) | 95.5 |
| Qwen3-VL-2B / BadNets-T | 83.8 | Fires | 447 | 0 | 966 (100%) | 94.4 |
| Qwen3-VL-8B / BadNets-T | 71.7 | Fires | 441 | 18 | 966 (100%) | 92.0 |
| InternVL3.5-2B / BadNets-I | 36.5 | Safely abstains | 0 | 0 | 966 (100%) | — |
| Qwen3-VL-2B / BadNets-I | 28.2 | Safely abstains | 0 | 0 | 966 (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 (), 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- 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):
- Lower bound
r_min = 0.05: Balances coverage of realistic poisoning budgets () against small-sample order-statistic noise. Loweringr_minto0.025admits 3 clean-only false firings (88 FP across 6,000 clean control samples vs. 5/6 abstaining atr_min = 0.05), whereas raisingr_minto0.10achieves 6/6 clean abstentions (0 FP) at the cost of excluding 5%–9.9% poisoning budgets. - Upper bound
r_max = 0.50: Serves as a loose minority-fraction cap. Because the dominant gap ratio occurs near the true poison fraction, anyr_maxin[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:
| Attack | Trigger Vocabulary | Residual Poisoned Examples (/100) | Residual Examples per Trigger Token | Post-Retraining Mean ASR |
|---|---|---|---|---|
| BadNets-T | 1 shared token (BadMagic) | 3–5 examples | 3–5 on BadMagic | 54.17% |
| BadNets-MT | 4 distinct tokens (BadMagic, cf, mn, bb) | 3–4 examples | 0–2 per token | 4.17% |
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.
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:
-
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 ) 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 poisonedyessamples from the 966 cleanyessamples (94.52% mean F1 on triggered runs, preserving 100% of cleanyesexamples) and outperforms GMM-2/GMM-BIC/MAD-3 gate replacements across 54 conditions (Appendix C.4). -
Multi-seed, poisoning-rate, and upper-group window sensitivity (W4 & Q2): We added 24 multi-seed runs ( 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 and mean F1 variation across the grid. -
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), butBadNets-Tconcentrates all 3–5 residuals on a single trigger token (BadMagic), whereasBadNets-MTsplits them across 4 distinct triggers (0–2 residuals per trigger, below the memorization threshold). -
Introduction, Figure 2, and proofreading (W2 & W3): We rewrote
Section 1to clarify the link between local directional alignment and local gradient neighborhoods, redesignedFigure 2(a)–(d)with explicit model/dataset/attack annotations, and corrected theAppendix A.3dimension notation.
Please let us know if there is anything else we can clarify before the discussion phase ends!
Best regards, Authors of Submission 32279
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.
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:To evaluate these differences empirically, we conducted 18-condition head-to-head comparisons (3 MLLMs × 2 triggers × 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:While the unmasked control trains normally, RepGuard's activation masking collapses CIDEr and VQA accuracy to ≈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).yes)This comparison highlights the structural difference between forced partitioning and calibrated abstention:
yesvs.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.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.