acceptodds
Under review as a conference paper at ICLR 2027

Probing Local Conditional Fit for Diffusion Model Membership Inference

Abstract

Membership inference determines whether a candidate example was used to train a text-to-image diffusion model and is an important tool for auditing training-data use. We identify a directional membership signal in conditional prediction shifts, the changes in predicted noise under alternative text conditions at a fixed noisy input: whether movement toward the alternative prediction initially lowers or raises the denoising loss from its value under the original caption. Members exhibit less room for such local improvement. To quantify this response, we propose LoFIT (Local Conditional Fit), which measures the signed initial response of the denoising loss along normalized prediction shifts and aggregates it across multiple text conditions and diffusion states. LoFIT requires no target-model gradients and can be computed using only forward queries to the denoiser. Evaluations on four datasets show strong membership discrimination, particularly for fine-tuned models. LoFIT distinguishes members early in fine-tuning, while the naive denoising-loss baseline remains near chance. Controlled comparisons show that the membership signal is concentrated in the signed first-order response, with little discrimination from prediction-shift magnitude alone.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Membership inference attack, Diffusion models, Prompt landscape, Text-to-Image Diffusion models

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

This paper proposes LoFIT, a membership inference attack for T2I diffusion models that exploits differences in conditional denoising between training members and non-members. The method measures the signed alignment between the denoising residual under the original caption and the prediction shift induced by alternative captions. The authors hypothesize that member samples exhibit less local denoising improvement than non-members. LoFIT aggregates these alignment scores across multiple text conditions and diffusion states without needing model gradients.

Critical strengths
  1. LoFIT is a straightforward interpretation of membership inference through directional changes in denoising loss. It requires only forward denoiser evaluations, without gradients or additional classifier training.

  2. LoFIT achieves good improvements over existing approaches: CLiD, SD-MIA, and PFAMI, particularly in low-FPR membership detection on fine-tuned diffusion models. The results are consistent across multiple datasets and Stable Diffusion backbones. The paper also includes controlled comparisons using identical reference predictions and diffusion states.

Critical weaknesses
  1. The observation that membership information can be extracted from differences in conditional denoising behavior has already been explored by CLiD and, through generation-based sensitivity in SD-MIA. LoFIT is the normalized first-order component of CLiD's conditional loss discrepancy, as dervied in Section 5.3. The main contribution is therefore a modification of an existing membership statistic rather than a new mechanism. The novelty appears relatively incremental.

  2. All evaluated models are in the Stable Diffusion family and use UNet-based denoisers with epsilon-prediction objectives. Modern T2I models use DiTs and flow-matching models, such as SD3 and FLUX. LoFIT's effectiveness across different architectures and prediction objectives is not explored in the paper. This is needed because the proposed membership signal is based on an empirical hypothesis about local conditional denoising fit rather than a theoretically property of diffusion/flow models.

  3. LoFIT achieves amazing membership discrimination on fine-tuned models (Table 1), but performance drops considerably on pretrained SD v1.5 (Table 2), with TPR@5%FPR below SD-MIA. Moreover, the fine-tuning settings are designed such that it exposes training examples multiple times during training. This raises questions about whether the strong membership signal is due to extensive fine-tuning and overfitting rather than LoFIT's discrimination. The relatively weak performance on pre-trained models limits the parctical utility of LoFIT.

Questions for authors
  1. Can the authors clarify whether the LoFIT hypothesis is expected to hold for DiT-based and flow-matching models?

  2. How sensitive is LoFIT to the number of training epochs and the amount of fine-tuning? Section 5.1 shows different training checkpoints, but it would be useful to see whether the strong results survive under fewer repeated exposures of each sample.

  3. The experiments assume access to the original training captions, and Appendix C.3 shows a significant performance drop when using generated captions. Could the authors discuss how this requirement affects LoFIT's real-world MIA utility?

5 comments

We thank the reviewer for your detailed feedback. We clarify the method’s novelty and provide new evidence across model architectures, training exposures, and caption-free settings.

W1. Novelty relative to conditional-loss and sensitivity attacks

The current manuscript does not sufficiently explain the motivation and significance of our method. We will add the following explanation to the revision.

TL;DR: CLiD's gap measures a loss change, the normalized gap an average rate of change, and LoFIT an instantaneous rate. This local-fit formulation improves detection and extends to autoregressive cross-entropy.

Local conditional optimality motivates examining the loss response to conditioning changes. Let L(c)=ℓ(f(c),y)L(c)=\ell(f(c),y), where f(c)f(c) is the prediction under continuous condition representation cc and yy is the fixed target.

At a smooth interior conditional optimum, ∇cL(c0)=0\nabla_cL(c_0)=0. We hypothesize that training pairs exhibit better local fit and less potential for loss reduction through conditional adjustments. However, a small condition derivative can also arise because the prediction barely changes. We separate these effects by normalizing the directional loss response by the corresponding prediction change:

∇cL(c0)⊤Δc∣∇cf(c0)Δc∣2=∇fℓ(f0,y)⊤∇cf(c0)Δc∣∇cf(c0)Δc∣2≈G.\frac{\nabla_c L(c_0)^\top\Delta c} {\\|\nabla_c f(c_0)\Delta c\\|_2} = \frac{\nabla_f\ell(f_0,y)^\top\nabla_c f(c_0)\Delta c} {\\|\nabla_c f(c_0)\Delta c\\|_2} \approx G.

where Δc=cj−c0\Delta c=c_j-c_0, and ∇cf\nabla_c f denotes the prediction Jacobian with respect to conditioning. This approximation follows from the chain rule.

With f0=f(c0)f_0=f(c_0) and nonzero reference shift δ=f(cj)−f0≈∇cf(c0)Δc\delta=f(c_j)-f_0\approx\nabla_c f(c_0)\Delta c, the normalized response is

G=⟨∇fℓ(f0,y),δ∣δ∣2⟩=ddsℓ!(f0+sδ∣δ∣2,y)∣s=0.G=\left\langle\nabla_f\ell(f_0,y),\frac{\delta}{\\|\delta\\|_2}\right\rangle =\left.\frac{d}{ds}\ell\\!\left(f_0+s\frac{\delta}{\\|\delta\\|_2},y\right)\right|_{s=0}.

where ss is output distance. For finite references, GG remains the exact instantaneous loss rate along the observed prediction shift.

LoFIT separates the instantaneous response from the finite loss changes measured by CLiD. CLiD's discrepancy term measures the endpoint loss change ΔL\Delta L [5]. Dividing by ∣δ∣2\\|\delta\\|_2 gives the average rate; LoFIT measures the instantaneous rate at the original prediction. For denoising residual r=f0−yr=f_0-y,

ΔL∣δ∣2=G+∣δ∣2,G=2v,v=⟨r,δ⟩∣δ∣2.\frac{\Delta L}{\\|\delta\\|_2}=G+\\|\delta\\|_2,\qquad G=2v,\quad v=\frac{\langle r,\delta\rangle}{\\|\delta\\|_2}.

The average rate includes the effect of finite displacement, so an endpoint loss increase can coexist with an initial decrease. SD-MIA examines generation stability under textual perturbations [6]; LoFIT directly measures whether the induced direction improves the queried sample's training loss.

On identical Flickr30k predictions, unnormalized first-order scoring reaches AUROC 0.909, compared with 0.906 for normalization alone and 0.813 for the raw gap; combining both reaches 0.935. Over the normalized gap, LoFIT improves TPR@1%FPR by 12.0/9.8/14.5 percentage points on Flickr30k/COCO/Pokemon without extra queries.

The same formulation extends to autoregressive cross-entropy. The cross-entropy gradient p−eyp-e_y replaces the denoising-loss gradient in GG, where pp is the original token distribution and eye_y the observed token's one-hot vector. We keep the image-token prefix and target fixed and change the caption. On LlamaGen GPT-XL trained for 10 epochs on 2,500 COCO images, we test 2,000 members and 2,000 non-members at the same twelve-forward budget:

Method and referencesForwards/imageAUROC / TPR@1%FPR
LoFIT-AR, 11 references120.952 / 0.567
CLiD_th adaptation, 11 references120.861 / 0.239
Kowalczuk et al., our 11-reference extension (Min-K% 50%)120.848 / 0.180

Kowalczuk's statistic [1] is selected on calibration data and fixed across checkpoints; the CLiD_th mixture is also calibrated. The gains show that the scoring contribution extends beyond denoising. We will add the full Jacobian connection, cross-entropy decomposition, AR implementation, and complete comparisons to the revised appendix.

W2 and Q1. DiT architectures, flow matching, and video generation

TL;DR: LoFIT improves detection on transformer-based CogVideoX and flow-matching Wan, extending beyond UNet denoisers.

The same first-order score applies to noise and velocity prediction. We evaluate CogVideoX-2b [2] and Wan2.1 [3], both transformer models. For Wan, we replace the noise target with the flow-matching velocity target u=ϵ−x0u=\epsilon-x_0 and use the same normalized residual interaction.

Each video's training caption and empty reference share noise at 12 fixed levels, costing 24 condition evaluations. CLiD_th uses α=0.5\alpha=0.5.

ModelDatasetTraining updatesMembersNon-membersCLiD_th AUROCCLiD_th TPR@5%FPRLoFIT AUROCLoFIT TPR@5%FPR
CogVideoX-2bAnime-Shooter2005004,0000.8700.5300.9480.692
CogVideoX-2bOpenVid4,0001,0004,0000.5940.1140.6560.116
Wan2.1Anime-Shooter2005003,9980.5640.0740.6100.112
Wan2.1OpenVid7,0001,0004,0000.5470.0720.6040.105

LoFIT improves both metrics over CLiD_th across these settings. We will add these results to the main experimental summary and document the noise/velocity formulations and full training settings in the revised appendix.

W3. Extensive fine-tuning and weaker pretrained discrimination

TL;DR: LoFIT's AUROC advantage appears before extensive fine-tuning, and additional sampling improves pretrained-model detection without further training.

The advantage is present within the first few training epochs. The three-seed COCO/Pokemon experiments in Q2 show earlier AUROC separation for LoFIT than CLiD_th under identical predictions. Strong low-FPR detection develops with greater exposure.

The pretrained result improves by increasing attack sampling alone. On the unchanged pretrained model, AUROC/TPR@5%FPR rises from 0.602/0.101 to 0.635/0.114, versus SD-MIA's 0.571/0.112. The new run uses 340×4×2=2,720340\times4\times2=2{,}720 condition evaluations, compared with 4,800 for SD-MIA. We will report these results separately from the fine-tuning study in the revision.

Q2. Sensitivity to training exposure

TL;DR: Figure 3 shows LoFIT improving earlier than the baselines, and new 1–10-epoch experiments establish an early AUROC advantage.

LoFIT reaches high discrimination earliest among the baselines in Figure 3. On Flickr30k, it reaches AUROC 0.994 at 40k updates, while official CLiD_th continues improving through 80k.

We additionally evaluate much lower exposure, starting at one epoch. Both attacks share eleven references (five combined edits, five random captions, and empty) and 48 timestep/noise states, costing 576 condition evaluations. CLiD_th's alpha is calibrated separately at each checkpoint. COCO/Pokemon use training seeds 1/2/3 and test sets of 2,000/2,000 and 333/334; Flickr30k uses one run and 8,000/8,000. AUROC is mean ± SD where multiple seeds are available; TPR is the mean.

DatasetEpochs (updates)LoFIT AUROCCLiD_th AUROCLoFIT TPR@1%FPRCLiD_th TPR@1%FPR
COCO1 (10)0.511 ± 0.0010.497 ± 0.0000.0100.011
COCO2 (20)0.522 ± 0.0010.497 ± 0.0000.0110.011
COCO5 (50)0.550 ± 0.0010.510 ± 0.0210.0150.010
COCO10 (100)0.590 ± 0.0010.552 ± 0.0020.0180.009
Pokemon1 (2)0.496 ± 0.0030.452 ± 0.0020.0120.010
Pokemon2 (4)0.517 ± 0.0020.459 ± 0.0020.0110.015
Pokemon5 (10)0.549 ± 0.0040.490 ± 0.0010.0040.014
Pokemon10 (20)0.575 ± 0.0030.515 ± 0.0040.0090.012
Flickr30k1 (625)0.5240.5160.0160.012
Flickr30k2 (1,250)0.5500.5310.0170.012
Flickr30k5 (3,125)0.6140.5660.0280.015

The early AUROC advantage precedes strong low-FPR detection. At 100 epochs, LoFIT's TPR@1%FPR reaches 0.604/0.307 on COCO/Pokemon, versus 0.087/0.074 for CLiD_th. We will add the full exposure study, including empty-reference controls and training settings, to the revision.

Q3. Practical effect of missing original captions

TL;DR: Generated captions enable image-only LoFIT audits, and LoFIT further improves MoFit using the same optimized conditions.

LoFIT retains an advantage without the original training caption. In the manuscript's BLIP-2 comparison, AUROC is 0.641/0.669/0.621 on Flickr30k/COCO/Pokemon, versus 0.603/0.650/0.576 for official CLiD_th with oracle alpha. Gaussian embedding probes raise it to 0.842/0.904/0.802 on these fine-tuned models; on pretrained LAION-MI, caption references give 0.546 versus 0.532 for Gaussian probes.

LoFIT also improves MoFit, an existing caption-free attack. We keep MoFit's optimized conditions fixed and replace only its loss-gap terms [4]. Both methods fit their scaler and mixing weight on the same calibration split. Scoring costs 15 forwards after MoFit's 1,300/1,200 gradient-based optimization passes per COCO/Pokemon image.

MethodCOCO AUROC [95% CI]COCO TPR@1%FPR [95% CI]Pokemon AUROC [95% CI]Pokemon TPR@1%FPR [95% CI]
CLiD_th, BLIP-2, published protocol0.648 [0.631, 0.665]0.041 [0.028, 0.060]0.571 [0.527, 0.612]0.045 [0.015, 0.075]
LoFIT, BLIP-2, no MoFit optimization0.757 [0.741, 0.772]0.098 [0.066, 0.125]0.594 [0.549, 0.635]0.033 [0.003, 0.093]
MoFit0.677 [0.660, 0.692]0.062 [0.047, 0.089]0.599 [0.556, 0.641]0.036 [0.015, 0.075]
MoFit + LoFIT0.815 [0.802, 0.827]0.140 [0.104, 0.177]0.692 [0.653, 0.731]0.057 [0.009, 0.102]

The improvement on unchanged conditions shows that LoFIT complements caption-free condition construction. We will distinguish supplied-caption and image-only audits in the revised abstract and introduction and add the full MoFit comparison and settings.

References

[1] Antoni Kowalczuk et al. Privacy Attacks on Image AutoRegressive Models. ICML, 2025.

[2] Zhuoyi Yang et al. CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer. ICLR, 2025.

[3] Team Wan et al. Wan: Open and Advanced Large-Scale Video Generative Models. arXiv:2503.20314, 2025.

[4] Joonsung Jeon et al. No Caption, No Problem: Caption-Free Membership Inference via Model-Fitted Embeddings. ICLR, 2026.

[5] Shengfang Zhai et al. Membership Inference on Text-to-Image Diffusion Models via Conditional Likelihood Discrepancy. NeurIPS, 2024.

[6] Tao Qi et al. Black-box Membership Inference Attacks on the Pre-training Data of Image-generation Models. CVPR, 2026.

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

The paper studies gray-box membership inference for text-to-image diffusion models, where the auditor can query the noise predictor at chosen noisy inputs, timesteps and text conditions and holds the original caption, but has no weights or gradients. It proposes LoFIT: at a fixed noisy state, compare the prediction under the original caption with predictions under reference captions (LLM rewrites, random captions, the empty prompt), and score the signed, shift-normalized inner product between the original residual and the prediction shift, i.e., half the initial directional derivative of the denoising loss toward the alternative prediction (Eqs. 7–8). Members are hypothesized to have less room for such local improvement. The score averages this response over timesteps, noise draws and references (Eq. 10), with no classifier or calibration. On fine-tuned SD v1.4, v2.1 and SD-XL (Flickr30k, COCO, Pokemon), LoFIT reaches AUROC 0.81–0.98 and TPR@1%FPR up to 0.72, well above CLiD (even with an oracle-selected α), PFAMI, SD-MIA, SecMI and PIA; on pretrained SD v1.5 (LAION-MI) it reaches AUROC 0.602 vs 0.571 for SD-MIA, with slightly lower TPR at 5% FPR. Matched-budget comparisons (Fig. 2), identical-prediction readout comparisons (Tables 3, 5), early-checkpoint detection (Fig. 3), Gaussian conditioning probes (Table 4) and a sign ablation (Table 13) analyze where the signal comes from.

Critical strengths
  1. The empirical case is carefully controlled and favours the baselines where it matters: CLiD_th is given an oracle α chosen on the evaluation labels (App. A.3), CLiD is re-run with LoFIT's own references and states (Table 3) and at matched query budgets (Fig. 2), and the readout comparison on identical predictions (Table 5) isolates the scoring rule from the prediction pool. The margins on fine-tuned models are large and consistent across three backbones and three datasets, which makes the main claim convincing for that setting.
  2. The method is simple and practical within its threat model: forward queries only, no gradients, no shadow models or calibration, and the paper is candid about its limits (second place at 5% FPR on LAION-MI; Gaussian probes requiring embedding access; rewrites hurting on LAION-MI).
Critical weaknesses
  1. The headline results are for image–caption pair membership under heavy fine-tuning, not for the image audit the paper motivates. All main results require the exact training caption. With BLIP-2 generated captions (Sec. 4.2, Table 17), AUROC drops to 0.641 / 0.669 / 0.621 on the fine-tuned SD v1.4 models (from 0.935 / 0.954 / 0.896), only marginally above CLiD_th (0.603 / 0.650 / 0.576), and to 0.546 on LAION-MI. On the pretrained model, which is the realistic audit target, the gain is modest (AUROC 0.602 vs 0.571) and TPR at 5% FPR is below SD-MIA (0.101 vs 0.112). The fine-tuned targets are trained for roughly 30–120 epochs on 0.4–10k images with a single fixed caption (Table 6), and Fig. 3 shows AUROC saturating to 0.994 with further training. The abstract and introduction should state the pair-membership requirement and the fine-tuning regime explicitly; as written, "auditing training-data use" promises more than the image-only numbers deliver.
  2. The conceptual contribution is thinner than the framing suggests; most of the gain is a better readout of CLiD's conditional loss gap. On identical predictions (Table 5, Flickr30k), normalizing CLiD's gap by the shift magnitude raises AUROC from 0.813 to 0.906, while isolating the first-order term adds a further 0.03 (0.935). The empty prompt alone yields 0.900 (Table 3), so the expensive LLM-rewrite pool contributes little (single-reference AUROC 0.64–0.74 for token, style, paraphrase and semantic edits, Table 14), and Gaussian directions do nearly as well as caption edits on COCO/Pokemon (Table 4). The optimality criterion of Sec. 3.1 is a sufficient, worst-case condition over references, but the score in Eq. 10 is an unweighted mean, and Fig. 4 shows both classes violate the criterion (median responses −0.18 and −0.55); LoFIT measures a degree of violation. None of this makes the method unsound, but the paper presents "local conditional fit" as the discovery when the operative change is normalization plus a larger, mostly free, reference set.
Questions for authors

Clarifications that could change my assessment:

  1. Please present the ablation chain explicitly in the main text — official CLiD → CLiD with LoFIT's references and states → normalized gap → first-order response — and report the min (or soft-min) aggregation over references implied by Eq. 5, as well as a weighted or selected reference pool, since the empty prompt alone recovers most of the performance.
  2. Please reconcile the CLiD baseline values (AUROC 0.70–0.75 on fine-tuned SD v1.4) with those reported by Zhai et al. (2024) for comparable settings, and specify what differs (fine-tuning budget, caption handling, VAE sampling).
  3. Will code, the generated reference pools (produced with a closed LLM) and the fine-tuned checkpoints be released, and can the main AUROCs be reported with variability over target-model training seeds rather than only over reference slots?

Suggestions (not affecting the score): the fallback S = 0 for an empty pair set in Eq. 10 is not neutral, since responses are negative for almost all images (Fig. 4); and because v = ‖r‖·cos θ, the unweighted average over timesteps is dominated by states with large residuals — a per-state standardization would make the aggregation cleaner.

6 comments

We thank the reviewer for your detailed review. We clarify LoFIT’s contribution and evaluation settings, supported by new controlled comparisons and training-seed results.

W1. Supplied captions, fine-tuning exposure, and pretrained-model performance

TL;DR: LoFIT supports image-only audits through generated or optimized conditions, improves detection at low training exposure, and strengthens pretrained-model results with additional sampling.

LoFIT retains an AUROC advantage with generated captions. In the manuscript's BLIP-2 comparison, AUROC is 0.641/0.669/0.621 on Flickr30k/COCO/Pokemon, versus 0.603/0.650/0.576 for official CLiD_th with oracle alpha. Gaussian embedding probes further raise it to 0.842/0.904/0.802 when embedding access is available. We will distinguish supplied-caption and image-only audits explicitly in the revised abstract and introduction.

LoFIT also improves MoFit, an existing caption-free attack. MoFit [1] optimizes conditions; we keep them fixed and replace only its loss-gap terms with first-order responses. Both methods fit their scaler and mixing weight on the same calibration split and use 15 scoring queries. These follow MoFit's 1,300/1,200 gradient-based optimization passes per COCO/Pokemon image.

MethodCOCO AUROC [95% CI]COCO TPR@1%FPR [95% CI]Pokemon AUROC [95% CI]Pokemon TPR@1%FPR [95% CI]
CLiD_th, BLIP-2, published protocol0.648 [0.631, 0.665]0.041 [0.028, 0.060]0.571 [0.527, 0.612]0.045 [0.015, 0.075]
LoFIT, BLIP-2, no MoFit optimization0.757 [0.741, 0.772]0.098 [0.066, 0.125]0.594 [0.549, 0.635]0.033 [0.003, 0.093]
MoFit0.677 [0.660, 0.692]0.062 [0.047, 0.089]0.599 [0.556, 0.641]0.036 [0.015, 0.075]
MoFit + LoFIT0.815 [0.802, 0.827]0.140 [0.104, 0.177]0.692 [0.653, 0.731]0.057 [0.009, 0.102]

The improvement with unchanged optimized conditions shows that LoFIT complements caption-free condition construction.

Additional sampling improves pretrained-model detection without further target training. The new run uses 340 timesteps, four noise draws, and two conditions: 340×4×2=2,720340\times4\times2=2{,}720 condition evaluations. SD-MIA's 2,400 conditional/unconditional pairs count as 4,800 evaluations under the same convention.

MethodCondition evaluations per imageAUROCTPR@5%FPR
SD-MIA4,8000.5710.112
LoFIT, submitted setting3840.6020.101
LoFIT, additional run2,7200.6350.114

LoFIT's advantage also appears at lower training exposure. Figure 3 shows LoFIT approaching saturation by 40k updates while official CLiD_th continues improving through 80k. After just five epochs, three-seed AUROC is 0.550 ± 0.001 on COCO and 0.549 ± 0.004 on Pokemon, versus 0.510 ± 0.021 and 0.490 ± 0.001 for CLiD_th on the same predictions. We will add the caption-free, pretrained-budget, and low-exposure results and their full settings to the revision.

W2. Scoring contribution, reference value, and local fit

The current manuscript does not sufficiently explain the motivation and significance of our method. We will add the following explanation to the revision.

TL;DR: CLiD's gap measures a loss change, normalization gives an average rate of change, and LoFIT measures an instantaneous rate. Additional references contribute information beyond the empty prompt.

Local conditional fit motivates measuring the loss response to conditioning changes. For prediction f(c)f(c) under a continuous condition representation cc and target yy, write L(c)=ℓ(f(c),y)L(c)=\ell(f(c),y).

At a smooth interior conditional optimum, ∇cL(c0)=0\nabla_cL(c_0)=0. We hypothesize that training pairs exhibit better local fit and less potential for loss reduction through conditional adjustments. However, a small condition derivative can also arise because the prediction barely changes. We separate these effects by normalizing the directional loss response by the corresponding prediction change:

∇cL(c0)⊤Δc∣∇cf(c0)Δc∣2=∇fℓ(f0,y)⊤∇cf(c0)Δc∣∇cf(c0)Δc∣2≈G.\frac{\nabla_c L(c_0)^\top\Delta c} {\\|\nabla_c f(c_0)\Delta c\\|_2} = \frac{\nabla_f\ell(f_0,y)^\top\nabla_c f(c_0)\Delta c} {\\|\nabla_c f(c_0)\Delta c\\|_2} \approx G.

where Δc=cj−c0\Delta c=c_j-c_0, and ∇cf\nabla_c f denotes the prediction Jacobian with respect to conditioning.

With f0=f(c0)f_0=f(c_0) and nonzero reference shift δ=f(cj)−f0≈∇cf(c0)Δc\delta=f(c_j)-f_0\approx\nabla_c f(c_0)\Delta c, the normalized response is

G=⟨∇fℓ(f0,y),δ∣δ∣2⟩=ddsℓ!(f0+sδ∣δ∣2,y)∣s=0.G=\left\langle\nabla_f\ell(f_0,y),\frac{\delta}{\\|\delta\\|_2}\right\rangle =\left.\frac{d}{ds}\ell\\!\left(f_0+s\frac{\delta}{\\|\delta\\|_2},y\right)\right|_{s=0}.

where ss is output distance. For finite references, GG remains the exact instantaneous loss rate along the observed prediction shift.

LoFIT isolates the instantaneous rate from the normalized gap's average rate. For denoising residual r=f0−yr=f_0-y and loss change ΔL=∣r+δ∣22−∣r∣22\Delta L=\\|r+\delta\\|_2^2-\\|r\\|_2^2,

ΔL∣δ∣2=G+∣δ∣2,G=2v.\frac{\Delta L}{\\|\delta\\|_2}=G+\\|\delta\\|_2,\qquad G=2v.

Here vv is the LoFIT response. The average rate includes the effect of finite displacement; the instantaneous rate identifies an initial improvement even when the endpoint loss is higher. On identical predictions, LoFIT improves TPR@1%FPR over the normalized gap by 12.0/9.8/14.5 percentage points on Flickr30k/COCO/Pokemon. Q1 reports both ablation paths and their overlapping gains.

Negative responses closer to zero imply smaller attainable improvements along the tested directions. Along a unit prediction direction, a negative response vv places the line minimum at s∗=−vs^*=-v, with attainable improvement v2v^2. Thus the reported member/non-member responses of -0.18/-0.55 support less local improvement for members. Equation 5 requires every direction to satisfy a sufficient condition; our score measures the average response. Q1 tests this choice: mean and softmin outperform the hard minimum. We will clarify this distinction in Section 3.

Additional references improve detection when the query budget is increased. Empty-only sampling already captures substantial signal. Starting from 240 states (480 queries), we spend 240 more queries on either 120 additional empty-only states or four captions at 48 states, using a fixed 1:1 combination weight. Both branches cost 720 queries. Further empty-only sampling adds only 0.0005 AUROC on COCO and 0.0019 in the Pokemon random/combined comparisons; the table shows the gains from other references over these equal-cost controls.

DatasetAdded referencesFinal AUROCAUROC gain over extra empty-only [CI]TPR@1%FPR gain [CI]
COCORandom captions0.9664+0.0199 [0.0169, 0.0230]+0.193 [0.087, 0.265]
COCOCombined edits0.9637+0.0172 [0.0142, 0.0205]+0.142 [0.027, 0.205]
COCOSemantic edits0.9532+0.0067 [0.0036, 0.0100]+0.107 [0.027, 0.175]
COCOParaphrases0.9521+0.0056 [0.0015, 0.0097]+0.081 [−0.051, 0.186]
COCOToken-level edits0.9415−0.0050 [−0.0089, −0.0011]−0.001 [−0.104, 0.071]
PokemonRandom captions0.8709+0.0477 [0.0356, 0.0602]+0.126 [0.009, 0.267]
PokemonCombined edits0.8701+0.0469 [0.0341, 0.0599]+0.183 [0.063, 0.282]
PokemonSemantic edits0.8612+0.0376 [0.0234, 0.0510]+0.115 [−0.003, 0.234]
PokemonParaphrases0.8570+0.0334 [0.0148, 0.0516]+0.100 [−0.012, 0.234]
PokemonToken-level edits0.8407+0.0171 [0.0019, 0.0319]+0.112 [−0.003, 0.219]

Other references therefore carry complementary information after empty-only sampling largely saturates. Empty is a sensible single-reference choice under a small budget; multiple references help when scaling the budget for stronger detection. The benefit depends on reference type, as the COCO token-edit result shows.

Caption references avoid model-specific Gaussian scale tuning. Gaussian probes can be competitive on fine-tuned models, but their useful scale changes across targets:

Gaussian scale hhSD v1.4 Pokemon AUROCCogVideoX-2b AUROC
200.6470.522
400.7670.579
800.8600.692
1600.8560.783
3200.7360.855
Empty caption reference0.8100.940

Caption references avoid selecting this embedding scale. They also perform better in our pretrained comparison: empty reaches AUROC 0.606 versus 0.534 for Gaussian probes at h=160h=160.

The same instantaneous-rate formulation applies to autoregressive cross-entropy. We replace the denoising-loss gradient in GG with p−eyp-e_y, the cross-entropy gradient for token probabilities pp and observed token yy. On LlamaGen GPT-XL trained for 10 epochs on 2,500 COCO images, we test 2,000 members and 2,000 non-members using identical references and twelve forwards:

Method and referencesForwards/imageAUROC / TPR@1%FPR
LoFIT-AR, 11 references120.952 / 0.567
CLiD_th adaptation, 11 references120.861 / 0.239
Kowalczuk et al., our 11-reference extension (Min-K% 50%)120.848 / 0.180

Kowalczuk's statistic [2] is selected on calibration data and fixed across checkpoints; the CLiD_th mixture is also calibrated. These comparisons establish the scoring contribution beyond denoising loss. We will add the full Jacobian connection, AR derivation, implementation, and complete comparisons to the revised appendix.

Reviewer_ufjbOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes LoFIT (Local Conditional Fit), a membership inference method for text-to-image diffusion models. It measures the local denoising-loss response along prediction shifts induced by alternative captions to infer whether an image–caption pair was used for training.

Critical strengths
  1. The paper is clearly written and easy to follow.
  2. LoFIT achieves better membership inference performance than evaluated baselines such as CLiD.
Critical weaknesses

The practical application scenarios appear limited. Model developers with complete training records can directly check whether a sample was used for training. External users of closed-source services typically lack access to the denoising predictions required by LoFIT. For open-weight models, running the attack may be costly for individual users, and the actionable benefits of identifying training membership remain unclear. The paper would benefit from explaining who would use the method and what practical decisions its results could support

Questions for authors

Can the underlying idea be extended to autoregressive image generation models or diffusion-based text generation models?

4 comments

We thank the reviewer for your sincere feedback. We clarify LoFIT’s practical use and present new results on autoregressive image generation.

W1. Practical users, access requirements, and actionable benefits

TL;DR: LoFIT uses forward queries to efficiently audit membership leakage in open-weight models, supporting release decisions and audits by data owners and model recipients.

LoFIT helps developers measure whether releasing a model exposes sensitive training membership. For example, revealing that a person's image belongs to a clinic's private dataset can itself disclose sensitive information, even without reproducing that image. Training records establish which examples were used; an attack measures whether a model recipient can infer that fact. The strength of the attack therefore matters: on the same Flickr30k target, LoFIT detects 45.4% of members at 1% FPR, versus 6.3% for official CLiD_th. These measurements help developers assess a release and compare whether fewer training epochs, regularization, or data-removal procedures reduce observable leakage.

The same audit also helps data owners and model recipients who lack training records. An artist can screen their image collection for possible training use, while an organization receiving a third-party checkpoint can check for exposure of known sensitive examples before deployment. Access to weights enables these tests but does not reveal the training dataset. When training captions are unavailable, our image-only experiments use generated or optimized conditions.

Gray-box attacks such as LoFIT, CLiD [1], SecMI [2], PIA [3], and SimA [4] are practical and computationally efficient with open-weight access. A released checkpoint lets an auditor evaluate the denoiser directly at selected inputs, timesteps, and captions. These methods extract membership signals from forward evaluations, avoiding the backward passes and gradient storage required by gradient-based white-box attacks. They also avoid the full sampling trajectories used by generation-based black-box attacks. Gray-box attacks thus make efficient use of the available model access: users can audit training-data leakage without the computational cost of gradient analysis or repeated image generation.

Within this setting, LoFIT improves detection at the same query budget without training an attack classifier. At 15 denoiser evaluations, it reaches AUROC 0.834 on Flickr30k versus 0.767 for the strongest evaluated CLiD variant. Auditors can also avoid LLM reference generation: the empty-only variant gives AUROC 0.900 on Flickr30k and 0.914 on COCO. We will make these use cases and computational advantages explicit in the introduction.

Q1. Extension to autoregressive image generation and diffusion-based text generation

TL;DR: LoFIT extends to autoregressive image generation and improves detection at both two and twelve forwards per image.

LoFIT replaces denoising loss with next-token cross-entropy. We supply the candidate image's observed tokens as context (teacher forcing), keep each prefix and target token fixed, and change only the caption. For the resulting shift δij\delta_{ij} in logits, original token probabilities pi0p_{i0}, and observed token's one-hot vector eyie_{y_i},

vijAR=⟨pi0−eyi,δij⟩∣δij∣2.v^{\mathrm{AR}}_{ij} =\frac{\langle p_{i0}-e_{y_i},\delta_{ij}\rangle}{\\|\delta_{ij}\\|_2}.

The cross-entropy gradient pi0−eyip_{i0}-e_{y_i} replaces the denoising residual. This measures the instantaneous loss response per unit logit displacement; members exhibit less potential for local improvement. We average over tokens and references using forward predictions alone.

We fine-tune LlamaGen GPT-XL on 2,500 COCO images for 10 epochs (100 updates) and test on 2,000 members and 2,000 non-members. Eleven references require twelve forwards per image; the empty reference requires two.

Kowalczuk et al.'s attack [5] scores conditional-minus-reference logits; we extend it to eleven references by averaging those differences. Each statistic is selected by mean calibration AUROC across 1/2/5/10 epochs and fixed across checkpoints: Loss for the empty reference and Min-K% 50% for eleven references. The CLiD_th adaptation calibrates its mixture of original loss and the negative mean CE gap.

StatisticForwards/imageAUROC / TPR@1%FPR
LoFIT-AR, 11 references120.952 / 0.567
CLiD_th adaptation, 11 references120.861 / 0.239
Kowalczuk et al., 11 references (Min-K% 50%)120.848 / 0.180
Raw CE gap, 11 references120.852 / 0.234
LoFIT-AR, empty reference only20.891 / 0.229
Kowalczuk et al., empty reference (Loss)20.794 / 0.157
Token NLL10.664 / 0.036

The gains at both forward budgets show that the first-order response extends to next-token prediction. We will add the full AR derivation, implementation details, and 1–10-epoch results to the revised appendix. Diffusion-based text generation remains future work.

References

[1] Shengfang Zhai et al. Membership Inference on Text-to-Image Diffusion Models via Conditional Likelihood Discrepancy. NeurIPS, 2024.

[2] Jinhao Duan et al. Are Diffusion Models Vulnerable to Membership Inference Attacks?. ICML, 2023.

[3] Fei Kong et al. An Efficient Membership Inference Attack for the Diffusion Model by Proximal Initialization. ICLR, 2024.

[4] Mingxing Rao, Bowen Qu, and Daniel Moyer. Score-based Membership Inference on Diffusion Models. arXiv:2509.25003, 2025.

[5] Antoni Kowalczuk et al. Privacy Attacks on Image AutoRegressive Models. ICML, 2025.

Thank you for clarifying the practical use cases and providing the additional autoregressive experiments. These address my main concerns, so I will raise my rating from 2 to 3 (weak acceptance).

Dear Reviewer ufjb,

Thank you for your constructive feedback and for considering our responses and additional experiments. We are glad that these addressed your main concerns. Your comments helped us clarify the practical relevance of our work and strengthen its evaluation.

Sincerely, The Authors

Comments on the paper

2 comments

Dear Reviewers,

Thank you again for your thoughtful reviews and constructive feedback. We have submitted our revised manuscript, incorporating the clarifications and additional results presented in our rebuttal.

We would greatly appreciate your consideration of the revision and our responses, and would be grateful to hear whether they fully address your concerns. We are happy to provide any further clarification.

Sincerely, The Authors

Thank you very much for your careful and helpful feedback. As the discussion period is coming to an end, we would greatly appreciate hearing whether our responses and the revised manuscript have satisfactorily resolved your concerns. If there is anything else you would like us to address, please let us know. We would be more than happy to clarify further.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.