acceptodds
Under review as a conference paper at ICLR 2027

M-Drift: Membership Inference via Gradient-Induced Mechanistic Drift

Abstract

Vision-language models (VLMs) and diffusion models are often trained on large web-scale datasets, making it difficult to determine whether a particular example was used during training. We introduce M-Drift, a white-box membership inference framework that uses a model's response to a small, reversible gradient-ascent perturbation as membership evidence. For each candidate example, M-Drift applies a single loss-increasing nudge and measures the resulting changes in loss, internal representations, and cross-modal interactions. For VLMs, we capture representation and cross-attention drift; for diffusion models, we aggregate denoising drift across sampled timesteps and noise realizations. We further use sparse feature projections to localize the internal directions most affected by the intervention. Across multiple VLM and diffusion backbones, M-Drift consistently improves membership detection in low-false-positive regimes over strong output- and gradient-based baselines. These results suggest that intervention-induced representation drift provides a useful and interpretable signal for auditing memorization in modern generative models.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
membership inference, mechanistic drift, model auditing, vision-language models, diffusion models, representation drift, privacy auditing, mechanistic interpretability

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_rcV7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents M-Drift, a white-box membership inference method for vision-language and diffusion models. The method applies a single gradient-ascent parameter update and measures the resulting changes in output loss, internal representations, and cross-modal attention. Sparse autoencoders provide additional feature-level drift statistics, while diffusion models aggregate signals across sampled timesteps and noise realizations. A calibrated classifier combines these signals for instance-level and dataset-level auditing. Experiments on LLaVA, Qwen2-VL, Stable Diffusion, and DiT report improvements over several membership inference baselines, with additional ablations and localization analyses.

Critical strengths

1.The most convincing evidence is Appendix F.2, Table 10: under matched white-box access and scoring, M-Drift improves TPR@1%FPR from 84.6% to 94.9% on LLaVA and from 74.2% to 88.0% on Stable Diffusion relative to combined static features. This directly supports the utility of measuring post-intervention drift.

  1. The authors explicitly define controlled-continuation membership where historical training records are unavailable (Section 5.1 and Appendix E.1). The low-FPR uncertainty analysis (Table 11), component and parameter ablations (Table 12), and localization tests (Table 13) strengthen the empirical case. These experiments provide useful support for the proposed method.
Critical weaknesses
  1. The scope of the membership evidence remains limited. For three of the four backbones, membership ground truth concerns controlled continuation training, not membership in the original pretraining corpus. The authors appropriately acknowledge this distinction. Nevertheless, clearer continuation-training details and an analysis of sensitivity to exposure frequency or calibration shift would help assess the method's applicability beyond the evaluated setting. This primarily limits the generality of the auditing conclusions.

  2. Improvements over the strongest competing attacks are uneven. In Tables 1–2, the TPR@1%FPR advantage is only 0.1 percentage points over MaxR on LLaVA, although the gains are larger on Qwen and remain positive for the diffusion models. Table 11 provides uncertainty estimates for M-Drift, but paired uncertainty for the improvements over the strongest baselines would better establish their practical significance. End-to-end computational costs would also help contextualize these gains.

Questions for authors

1.Could the authors clarify the controlled continuation-training schedule and per-sample exposure frequency, and discuss the expected applicability to lower-exposure or shifted-calibration settings?

2.Could the authors report paired uncertainty for TPR@1%FPR improvements against MaxR/GSA and provide end-to-end inference costs, including the eight diffusion probing trials?

4 comments

We thank the reviewer for recognizing our controlled-membership protocol and matched white-box comparison (Appendix F.2, Table 10). Following, we provide the response regarding generalization, paired uncertainty, and computational cost. We will update these in the manuscript and supplements.

W1: Membership scope and generalization. We explicitly distinguish controlled-continuation membership from unverified inclusion in historical pretraining. LLaVA, Qwen2-VL, and SDv1.5 use exact inclusion/exclusion labels for training continued from released checkpoints; DiT/ImageNet uses the known train/validation split (Sec. 5.1, App. E.1). We do not claim that a continuation non-member was absent from the original pretraining corpus. Source matching, deduplication, and blind temporal/metadata controls (Table 8) mitigate confounding, but cannot establish robustness to arbitrary calibration shift.

Table R4: Current Audit Scope and Future Work.

AspectCurrent study / rebuttal additionFuture work
Ground truthControlled continuation (3 models); exact DiT splitVerified original-pretraining membership
ExposureControlled training; low-exposure sweep [R4]Natural, long-tail exposure histories
CalibrationMatched pools and temporal controls; shift test [R4]Severe domain/language/prompt shifts
AccessWhite-box gradients and activationsAPI-only auditing

According to Table R4, we will narrow the abstract and conclusions to verified training events and distinguish present evidence from future generalization.

Q1: Scope and schedule. Labels are exact for a controlled continuation event (LLaVA, Qwen2-VL, SDv1.5) or the official split (DiT/ImageNet). We do not claim membership in the original pretraining corpus, and we will narrow the abstract and conclusions accordingly. Schedule details (to be added to App. E.1):

Table R5: Controlled continuation-training schedule and per-sample exposure.

BackboneEpochsOptimizer / LRBatchExposures per member
LLaVA-1.5-7B3AdamW / 2e-4163 (non-members: 0)
Qwen2-VL-2B3AdamW / 1e-4163 (non-members: 0)
SDv1.55AdamW / 1e-5165 (non-members: 0)

As mentioned in Table R5, members are shuffled each epoch and seen exactly once per epoch; non-members are never used.

Lower exposure is expected to weaken all MIAs. We will add an exposure sweep (members seen 1/2/4 times) and a calibration-shift test (calibrator fitted on one source pool, audit on another), keeping the protocol fixed. Table 8 controls apply only to the secondary proxy-provenance diagnostic; they do not establish robustness to arbitrary shift.

W2/Q2: Uneven gains and uncertainty. On LLaVA at 1% FPR, M-Drift vs MaxR (94.9 vs 94.8) is a tie, and we will say so. Using Table 11, 95% bootstrap CIs:

  • Qwen: MaxR's 84.1 lies outside M-Drift's CI [87.4, 92.2].
  • LLaVA, SDv1.5, DiT: the best baseline lies inside M-Drift's CI, so we will not claim significance at 1% FPR there. At the stricter 0.1% FPR (Table 1, supporting evidence only per App. G.2), LLaVA separates: 80.2 vs 71.8 (+8.4). We will add paired stratified bootstrap CIs on ΔTPR versus MaxR/GSA (same samples and thresholds as App. G.2).

Our central claim rests on the matched-access comparison (Table 10): drift improves over combined static white-box features, 84.6 → 94.9 (LLaVA) and 74.2 → 88.0 (SDv1.5). This isolates the effect of the intervention rather than implying uniform dominance.

Runtime. M-Drift costs ≈ 4 forward-equivalents for VLMs and K×K\times(2F+1B) for diffusion with K=8 (K=4: 85.8 TPR@1%FPR). We will add measured V100 FP16, batch-1 end-to-end latency, and peak memory for M-Drift and MaxR/GSA, including the update/restore, readouts, and all K trials.

Summary: We will incorporate tables, schedule details, paired inference, and exposure/shift sensitivity into the rebuttal revision. The defensible contribution is that intervention-induced multimodal drift adds membership evidence beyond matched static white-box features, not that every setting yields a large margin or that continuation labels prove historical-pretraining membership.

Thanks again for the detailed review. We have revised the manuscript (see updated file), with new text highlighted in yellow, in line with your points:

  • Scope of the membership claim (W1): The abstract, Sec. 5.1 and the Conclusion now limit our claims to the controlled-continuation labels (plus the exact DiT/ImageNet split) and do not claim membership in the original pretraining corpus. App. E.1 and App. H say the same, and that the blind controls apply only to the secondary diagnostic.
  • Continuation schedule and exposure (Q1): App. E.1 and Table R5 now list the epochs, optimizer, learning rate, batch size and exposure per member for each backbone.
  • Uneven gains and uncertainty (W2/Q2): Sec. 5.2 now states plainly that M-Drift ties MaxR on LLaVA at 1% FPR (94.9 vs. 94.8), where only the 0.1% FPR result separates them (80.2 vs. 71.8). It also says Qwen's gain is clearly significant, since MaxR's 84.1 lies outside our CI [87.4, 92.2], whereas the GSA margins on diffusion lie inside our CIs, so we do not claim significance there. Ties in Tables 1–2 are bolded jointly.
  • Our central claim: The matched comparison against static white-box features (84.6 → 94.9 on LLaVA, 74.2 → 88.0 on SDv1.5) is now presented as the main evidence (Table 3, Sec. 5.3).
  • Cost (Q2): Sec. 6 and Table R6 give the per-candidate costs, including the K = 8 diffusion trials.

There are some minor updates we will include them in the camera-ready version.

We hope the narrowed claims and added details resolve your concerns. If they do, we would be grateful if you would reconsider your score.

I appreciate the authors' detailed response and clarifications. However, some of my concerns remain unresolved, so I maintain my original rating.

Thank you for the follow-up, and for acknowledging the matched white-box comparison and the controlled-membership protocol.

These are the changes in the revised manuscript:

  • Membership scope: the abstract, Sec. 5.1 and the Conclusion now limit the claims to verified continuation-training events, and do not claim membership in the original pretraining corpus.
  • Continuation schedule and exposure: App. E.1, Table R5.
  • Uneven gains: we now report the LLaVA 1%-FPR result against MaxR as a tie. We claim a significant gain only for Qwen, where MaxR lies outside M-Drift's 95% CI (Sec. 5.2).
  • Cost: operation counts, including all K diffusion trials (Sec. 6, App. F.1, Table R6).

The central claim does not depend on uniform dominance over every baseline. It is that intervention-induced drift adds membership evidence beyond matched static white-box features: TPR@1%FPR rises from 84.6 to 94.9 on LLaVA and from 74.2 to 88.0 on SDv1.5 (Table 10).

We are committed to include any remaining references in final version. If any specific concern still seems unresolved, we would be grateful if you named it, so we can answer it directly. Thank you again for your time.

Reviewer_c1ZJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper explores membership inference on vision-language models and diffusion models via model response to single-step reversible gradient-ascent perturbations. The authors propose to collect membership evidence from gradient-ascent induced drift in model outputs, representations as well as cross-attention distributions for multi-modal VLMs. To further analyze drifts, the authors project activations into sparse autoencoder (SAE) coordinates to obtain disentangled features and use the SAE-coordinate drift for membership inference. The paper provides tests for datasets with controlled false positives for meaningful auditing.

Critical strengths
  1. The paper is well-written, with good motivations for membership inference in white-box regime.
  2. The paper validates the proposed approach using several metrics on multiple VLMs and diffusion models with detailed ablations.
Critical weaknesses
  1. Limited novelty and contributions: The core idea of using model’s response to gradient-ascent or unlearning is not new and has been explored in existing works [a,b,c,d] to achieve similar objectives. The paper does not discuss existing works in this direction. [a] proposed estimating the changes in internal model representations due to a single gradient-ascent step for membership inference using LLMs. [b] proposed to unlearn using gradient-ascent to trace the influential training data samples. [c,d] uses similar unlearning approaches for diffusion models for training data attribution.

  2. The paper does not discuss the additional complexity incurred and inference time overhead when using the proposed method compared to existing approaches.

[a] G-drift mia: Membership inference via gradient-induced feature drift in llms. ICPR 2026.

[b] Unlearning traces the influential training data of language models. ACL 2024.

[c] The mirrored influence hypothesis: efficient data influence estimation by harnessing forward passes. CVPR 2024.

[d] Data attribution for text-to-image models by unlearning synthesized images. NeurIPS 2024.

Questions for authors
  1. How is the work positioned against existing works [a,b,c,d] mentioned in weaknesses?
3 comments

We thank the reviewer for identifying these relevant studies. We mention very clearly that M-Drift's novelty lies in extending intervention-based membership inference to vision-language models (VLMs) and diffusion models through architecture-specific mechanistic signals, rather than proposing gradient ascent itself.

1. Distinction from [a] and [b]

[a] G-Drift (ICPR 2026) is our closest methodological predecessor, but it exclusively studies text-only LLMs using token-level loss, logits, and hidden-state drift. M-Drift addresses fundamentally different multimodal architectures and learning objectives, introducing (i) image–text binding drift, (ii) sparse-autoencoder (SAE) feature localization, and (iii) diffusion-specific drift aggregation across stochastic denoising timesteps. These are not investigated in G-Drift.

[b] UnTrac (ACL 2024) uses gradient-ascent unlearning to estimate the influence of training datasets on language-model predictions. M-Drift instead targets individual membership inference, using reversible interventions, internal representation changes, and multimodal binding signals.

2. Novelty relative to [c] and [d]

We agree that these diffusion-related data-attribution studies deserve explicit discussion. However, their objectives differ from M-Drift's membership-auditing formulation.

Table R2: Novelty Comparison with Prior Data-Attribution Methods [c,d].

MethodPrimary objectiveM-Drift's distinguishing contribution
[c] Mirrored Influence (CVPR 2024)Efficient training-data influence estimation using reciprocal influence and forward-pass evaluationDirect membership inference through intervention-induced representation, binding, and sparse-feature drift
[d] Unlearning Synthesized Images (NeurIPS 2024)Identify influential training images behind generated outputs via simulated unlearningAudit candidate membership using stochastic denoising drift, mechanistic readouts, and low-FPR calibration

Unlike [c,d], M-Drift integrates architecture-specific mechanistic evidence into a unified VLM/diffusion membership-auditing framework. Its contributions include cross-modal binding analysis, superposition-aware sparse feature localization, diffusion-timestep aggregation, and dataset-level auditing.

Does the multimodal part matter beyond [a]? Table 10 already contains the closest matched proxy for a G-Drift-style transfer (loss and raw hidden-activation drift; same θM\theta_M, layers, calibrator):

Table R3: Matched comparison with a G-Drift-style proxy (TPR@1%FPR, %; from Table 10).

TPR@1%FPRLLaVASDv1.5
Loss drift only68.055.0
Raw activation drift (G-Drift-style proxy)86.777.6
+ binding drift (w/o SAE)92.085.5
Full M-Drift94.988.0

The multimodal components (Table R3) add +8.2 / +10.4 pp over the G-Drift-style proxy. Additionally, we will include a faithful adaptation of [a] on the same splits.

Complexity. Per candidate:

  • Loss-only MIA: 1 forward.
  • M-Drift (VLM): 1 forward + 1 backward + 1 forward ≈ 4 forward-equivalents. Gradients flow only to a rank-4 LoRA probe; the base stays frozen in 4-bit (Table 5).
  • M-Drift (diffusion): K×K\times(2 forward + 1 backward). K trades cost for accuracy (Table 12): K=4 → 85.8, K=8 → 88.0, K=16 → 88.4. Candidates are independent and fully parallelizable. We will report measured V100 end-to-end latency and peak memory against MaxR/GSA, plus a TPR-vs-runtime curve.

Summary: M-Drift's contribution is not gradient ascent alone, but a unified mechanistic membership-auditing framework for multimodal and generative vision architectures. Its novelty lies in cross-modal binding drift, sparse feature localization, and calibrated low-FPR auditing. We will incorporate both comparison tables, acknowledge all four relevant works, and clarify the computational trade-offs in the revision.

References

[a] Ranjan, R., Grover, U., Lin, X., & Polyzou, A. (2026). G-Drift MIA: Membership inference via gradient-induced feature drift in LLMs. ICPR 2026. https://arxiv.org/abs/2604.00419

[b] Isonuma, M., & Titov, I. (2024). Unlearning traces the influential training data of language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (pp. 6312–6325). https://doi.org/10.18653/v1/2024.acl-long.343

[c] Ko, M., Kang, F., Shi, W., Jin, M., Yu, Z., & Jia, R. (2024). The mirrored influence hypothesis: Efficient data influence estimation by harnessing forward passes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 26276–26285). https://doi.org/10.1109/CVPR52733.2024.02483

[d] Wang, S.-Y., Hertzmann, A., Efros, A. A., Zhu, J.-Y., & Zhang, R. (2024). Data attribution for text-to-image models by unlearning synthesized images. In Advances in Neural Information Processing Systems, 37. https://arxiv.org/abs/2406.09408

We would like to thank the reviewer again for the references and the clear questions. We have revised the manuscript (see updated file), with new text highlighted in light yellow, to address both weaknesses:

  • Positioning against [a–d]: A new Related Work paragraph in Sec. 2 covers G-Drift, UnTrac, Mirrored Influence and the synthesized-image unlearning paper. Contribution (i) in Sec. 1 states that we extend G-Drift from text-only LLMs to VLMs and diffusion models, and App. F.3 with Table R2 compares all four works in detail. We now state clearly that gradient ascent itself is not our contribution.
  • What the multimodal readouts add over [a]: The highlighted rows of Table 3 and the "Matched white-box comparison" paragraph in Sec. 5.3 compares against the closest G-Drift-style proxy under identical access. That proxy reaches 86.7 / 77.6 TPR@1%FPR on LLaVA / SDv1.5, versus 94.9 / 88.0 for full M-Drift.
  • Complexity and overhead: The "Computational cost" paragraph in Sec. 6 gives the per-candidate cost: about four forward passes for VLMs, and K × (2F + 1B) for diffusion. The accuracy-vs-K trade-off is also reported there. App. F.1 and Table R6 list the operation counts.

We will carefully review all the contents and if missing anything will add before the camera-ready deadline.

We believe the paper's contribution is now clearly separated from [a–d]. If you agree, we would be grateful if you would reconsider your score.

I thank the authors for the rebuttal responses. The experiments added during the rebuttal improves the paper.

After reading the authors response, I believe the comparison to the suggested papers is better clarified. The novelty of the work is still limited to application of drift-induced membership inference to VLMs and Diffusion Models.

The authors added analysis of the complexity of the method which shows that the proposed method incurs several times higher computation (about four forward passes for VLMs over 1 forward pass for loss-only MIA). The authors also do not present a comparison of computational cost compared to existing approaches as asked in my initial review.

Finally, the presentation of the paper is still not great with several figures in poor resolutions and readability (also pointed out by other reviewers). Based on all of this, I maintain the same rating.

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

This paper introduces M-Drift, a white-box framework for identifying whether an image or image–text example was used to train a vision-language or diffusion model. For each candidate, it applies a small, reversible gradient-ascent update that increases the sample’s loss, then measures changes in output loss, internal representations, and cross-modal attention. Optional sparse autoencoders localize these changes to interpretable feature coordinates, while diffusion-specific aggregation combines measurements across sampled timesteps and noise realizations. A calibrated classifier integrates the drift signals into membership scores, which also support dataset-level contamination tests. The central contribution is using a model’s response to a targeted intervention as membership evidence. Experiments report improved detection at low false-positive rates across several VLM and diffusion backbones.

Critical strengths
  1. The theoretical aim is ambitious. The authors tried to build a unified theoretical analysis paradigm over training data auditing in both visual understanding (VLMs) and genration (diffusion models). The proposed M-Drift, a theoretical framework under perturbed model parameters over gradient ascent tries unify the paradigms at higher up levels, while leaving the implementation remaining distinct.
  2. The theoretical analysis is sufficient. The authors provide sufficient analysis and induction in both the main paper and the supplementary material, making the induction theoretically sound.
  3. Sufficient quantitative experiments are provided to back the proposed framework in a condensed manner.
Critical weaknesses
  1. The underlying method beneath the proposed general framework at the model level remain a little bit scattered without a sufficiently elegant unification. At implementation, the authors used the auditing of the loss under perturbed model parameters, LoRA for VLMs and denoiser params respectively, while seeking the change of the loss in either answer token logits or noise prediction to guide the data auditing. However, this still remains one step away from the entire unification, as both of the two kinds of the models currently uses transformer as the backbone. The reviewer thinks this can be further unified at the decoder token logit level.
  2. The illustration of the paper is quite difficult for the general ML audience to quickly understand. The authors focus on the theoretical induction in both the main paper and the supplementary material, while leaving the a general illustraion of the method behind in the main text.
  3. The resolution and the layout of the images are problematic. The resolution of the images are ill-zoomed, as well as the tables. The placement of the images are not traditional either. It seems that Figure 1 is rarely put above the abstract in ICLR papers.
Questions for authors
  1. Can the authors try to unify the auditing formulation of both the VLMs and the diffusion models at the level of decoder logits? As the modern architecture in both models are now transformers, placing the methodology at this level will enahnce the unity of the proposed method.
  2. It is still possible that the attackers can mimic their attack data by imitating the loss activation patterns while remaining other traits intact. Is the reviewer understanding it correct? Or, the authors could theoretically prove that such attacking method will eventually lead to the data being identical to the training data?
3 comments

We thank the reviewer for recognizing the theoretical analysis and experimental evidence. We agree that the unified formulation and presentation can be clearer. We will incorporate these changes into the manuscript during rebuttal where permitted and in the final revision.

W1/Q1: Unification at the decoder-logit level. We will make the shared interface explicit in Sec. 4. M-Drift (Eqs. 2–8) is defined on the model's task-native prediction head: for output oθ(x,c)o_\theta(x,c), target yy and Lθ=ℓ(oθ,y)\mathcal L_\theta=\ell(o_\theta,y),

θM+=U(θM,∇θMLθ;η),Δo=oθ+−oθ,ΔL=Lθ+−Lθ,\theta_M^+=\mathcal U(\theta_M,\nabla_{\theta_M}\mathcal L_\theta;\eta),\quad \Delta o=o_{\theta^+}-o_\theta,\quad \Delta\mathcal L=\mathcal L_{\theta^+}-\mathcal L_\theta,

with representation, binding and SAE drift read from the same intervention (for diffusion, averaged over K=8K{=}8 paired (t,ϵ)(t,\epsilon) trials).

Table R1: Shared prediction-head interface across backbones.

BackboneHead outputLoss
LLaVA / Qwen2-VLDecoder token logitsTeacher-forced NLL
SDv1.5 (U-Net) / DiT (transformer)Predicted noise ϵθ\epsilon_\thetaDenoising MSE

As mentioned in Table R1, for VLMs this is the decoder-logit level the reviewer suggests. For diffusion, a transformer backbone (DiT) still has no vocabulary head: it regresses continuous noise, and SDv1.5 is a U-Net. Imposing token logits would require changing the generative objective, so we would no longer audit the model as trained. Predicted-noise drift Δϵ\Delta\epsilon (Table 6) is the continuous analogue of logit drift. We will state this mapping in Sec. 4 and note discrete-token image generators as a direct extension.

W2: Accessibility. We will open Sec. 4 with a plain-language overview and simplify Fig. 2 to four stages (candidate → reversible one-step nudge → output/internal drift → calibrated score), moving derivations to the appendix.

W3: Figures/layout. We have created a new Fig. 1 and will move from above the abstract into Sec. 1. No results change required.

Q2: Mimicry. We agree with the reviewer's understanding. Distinct inputs can share loss, activations, or even drift, so matching these statistics cannot prove that an input is identical to a training sample, and M-Drift does not claim this. It provides calibrated statistical evidence of membership at a controlled FPR, not an identity certificate or adversarial-robustness guarantee. Two points limit, but do not remove, this risk:

  1. Matching loss alone is insufficient: under matched access (Table 10), static loss gives only 8.0% / 12.0% TPR@1%FPR on LLaVA / SDv1.5, versus 94.9% / 88.0% for full drift. A forger must match the response of loss, representations and cross-attention to the update, which depends on gradients rather than loss values alone.
  2. Forging all drift channels requires white-box gradient access to the audited model.

We will add this as an explicit limitation and, time permitting, report an adaptive attack that optimizes non-members to match the full drift vector.

Reviewer_DuSeOpenReview ↗

Thanks for your reply. Looking forward to see the updated manuscript.

Thanks a lot for your reply. This helps better clarify my understandings. However, as my concern mostly lies in the illustration of the paper. I would like to see the updated paper. The author could upload an update of the paper with the revised text marked blue.

Thank you for the follow-up and for asking to see the updated paper.

We have uploaded a revised manuscript in which all new or changed text is highlighted (in light yellow, not blue, so it stays readable over the blue colored text and links). Since your main concern was the illustration, here is where it changed:

  • Clearer illustration (W2): The new four-stage Figure 1 (candidate → reversible nudge → drift → decision) is in Sec. 1, below the abstract, with a rewritten caption. A plain-language "Overview" now opens Sec. 4 and walks through the same four stages.
  • Figure/layout (W3): Figure 1 is moved out of its old position above the abstract and Sec. 6 now uses its notation (Δθ_m, Δθ_nm). Tables 1–3 are still scaled to fit the page, so we have not fully fixed their resolution.
  • Logit-level unification (W1/Q1): Sec. 4.6 and the new Table R1 define a shared prediction-head interface. For VLMs it is exactly the decoder-logit level you suggested. Diffusion models regress continuous noise and have no token-logit head, so we use predicted-noise drift as the analogue.
  • Mimicry (Q2): Sec. 6, "What drift evidence does not establish", and App. H state that matching loss or drift cannot prove an input equals a training sample. We do not claim adversarial robustness, and the adaptive attack is listed as future work.

We hope the revisions addresses your concern. If it does, we would be grateful if you would reconsider your score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.