acceptodds
Under review as a conference paper at ICLR 2027

SharpMIA: Critical Sharpness Beyond the Input Hessian for Membership Inference

Abstract

Membership inference seeks to answer if a given example is in a model's training set. It is both an instrument to measure a model's data retention and an attack vector to leak private data. Nearly every attack against vision models reads the model at the example itself and differs only in what it compares that reading against; the one line of work that moves the reading into the loss geometry takes the trace of the input Hessian to extract curvature, still at the example. We find that the loss carries a membership signal not only at the example but also in the local geometry of its neighborhood. To characterize the geometry and derive a signal, we consider the changes in loss over a finite distance across the example to extract critical sharpness, —this is effectively how far one must walk from the example before the loss returns to its starting value. A walk needs a direction. Consequently, computing needs the target model's gradient. But, we show that it is not necessary. So the signal poses a problem by facilitating a black-box attack under Carlini et al. (2022)'s offline, black-box threat model. Effectively, a walking direction can be averaged from the adversary's access to reference models without the example to approximate . Interestingly, we find the critical sharpness and prior work's curvature evidence are strong on different examples. So, we fuse both signals together with a difficulty prior we introduce to aid signal fusion. Reproducing 11 prior attacks and evaluating our method, SharpMIA, with standard models and benchmarks, we show that: i) sharpness leaks membership information; and ii) combining it leads to the strongest attack.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
membership inference, input loss curvature, critical sharpness

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.40 over 5 reviews, above 68% of the others.
Reviewer_QfjfOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper argues in favor of compounding the membership signal (curvature of the loss as defined by Ravikumar et al.[1] and the target's loss) by using critical sharpness (λc\lambda_c) which characterizes the curvature of loss near the neighbourhood of a target sample. With empirical evidence, the author(s) establish superiority of λc\lambda_c - aided SharpMIA w.r.t. SOTA membership inference attacks (MIAs). It differs from the latter in that instead of just using a point estimate, it derives membership information from the geometry of the loss in the locality of the target sample. Additionally, the paper's proposed SharpMIA attack is a black-box attack which does not require access to the model's gradients w.r.t. the target sample for computing a surrogate critical sharpness component (λv\lambda_v). It can be approximated using reference models. Notably, there already exists research that links critical sharpness to memorization [2] but this is the first work to utilize it for improving MIAs.

[1] Deepak Ravikumar et al. Curvature clues: Decoding deep learning privacy with input loss curvature. NeurIPS (2024).

[2] Moosavi-Dezfooli et al. Robustness via curvature regularization, and vice versa. IEEE CVPR (2019).

Critical strengths
  • Computing the critical sharpness does not rely on access to target model gradients for the target samples. The gradients can be approximated from the reference models trained without the target sample. This is cost effective as the reference models can be reused across target samples, amortizing its training compute and avoiding expensive per-sample retraining.
  • Demonstrated marked improvement over other SOTA methods (as listed in Table 1, Page 8). SharpMIA using a compound signal comprising of loss curvature signal (raw and calibrated), target model's confidence score, and the critical sharpness signal (raw and calibrated) is clearly superior to the other baselines across diverse models and datasets.
Critical weaknesses

None

Questions for authors

Suggestions

  • Carlini et al. [3] characterized both member and non-member samples as being hard-to-learn and easy-to-learn and Aerni et al. [4] urge for MIAs to be used to evaluate per-sample vulnerabilities. The argument in lines 277–280 also relies on a high-confidence approximation. Could the author(s) stratify the target samples by the their correct-class confidence and report (i) the agreement between λv\lambda_v and the target-gradient measurement λc\lambda_c, and (ii) the per-sample SharpMIA performance evaluated using only λc\lambda_c and λv\lambda_v? This would help establish whether lower confidence affects the underlying sharpness signal. Otherwise, the proposed approach is robust to sample difficulty.

  • The attack inherits weaknesses of other reference model-based attacks including:

    • You need to train multiple reference models to estimate λv\lambda_v accurately.
    • Adversary is assumed to know the target model architecture, training hyperparameters, and data distribution.

The author(s) should add a Limitations section to acknowledge these inherited weaknesses or add experiments that prove the above arguments to be false.

[3] Carlini et al. Membership Inference Attacks From First Principles. IEEE S&P 2022.

[4] Aerni et al. Evaluations of Machine Learning Privacy Defenses are Misleading. ACM SIGSAC (CCS) 2024.

Reviewer_rTBvOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SharpMIA, a black-box membership inference attack based on the local geometry of the input loss landscape. The main observation is that membership information may be encoded not only in the model’s loss or curvature at a particular example, but also in the geometry of the surrounding neighborhood. Empirically, training examples tend to lie in wider basins and therefore exhibit lower critical sharpness than non-members. Experiments on CIFAR-10, CIFAR-100, and ImageNet show that SharpMIA achieves competitive or improved membership inference performance compared with a broad set of existing attacks, particularly when sharpness and curvature signals are combined.

Critical strengths
  1. The observation that membership information exists in the neighborhood geometry around an example rather than only at the example itself, while not new, is interesting.

  2. The paper constructs the direction using reference-model gradients and estimates the target gradient norm using finite differences. The reported strong agreement between this approximation and the white-box version makes this part of the method compelling.

  3. Extensive experiments are conducted and the results seem sound.

Critical weaknesses
  1. The methodological novelty relative to existing curvature and critical-sharpness work could be better established. Ravikumar et al. already use input-loss curvature for black-box membership inference, while Kalra et al. introduce critical sharpness as a finite-distance measure of loss-landscape curvature in parameter space. The current paper’s main conceptual step is therefore to transfer critical sharpness into input space and use it as a membership signal, followed by fusion with the existing Hessian-based signal. I think this is potentially useful, but the paper could more clearly articulate what fundamentally new insight is obtained beyond this combination.

  2. It is difficult to disentangle the contribution of the new signal from the learned fusion mechanism. Critical sharpness alone is not stronger than the Hessian-trace baseline: for example, Table 1 reports 66.73 vs. 67.83 AUROC on CIFAR-10, 85.94 vs. 86.25 on CIFAR-100, and 71.69 vs. 75.46 on ImageNet. The strongest results come after combining several raw/calibrated signals and a difficulty prior through an MLP.

  3. The query/computational cost deserves a much more explicit analysis. Unlike a point-wise curvature estimate, critical sharpness requires repeatedly querying the target along a trajectory until the loss returns to its starting value. A direct table reporting average/median target queries per example for SharpMIA and baseline attacks, together with attack performance under matched query budgets would be much more helpful.

  4. The threat model still makes fairly strong assumptions despite being described as black-box.The attacker does not need target parameters, but does know the target architecture, has access to data from the same training distribution, and trains a large pool of reference models. Moreover, the black-box walking direction relies specifically on these reference models. It is therefore unclear how robust the attack remains when the reference models differ from the target in architecture, training recipe, or data distribution. This matters for the practical significance of the claimed black-box attack.

  5. The empirical scope is somewhat narrow for a general membership-inference claim. The experiments are limited to image classification using ResNet-18/ResNet-50 on CIFAR-10, CIFAR-100, and ImageNet. Since the proposed signal is explicitly geometric and depends on input-space perturbations, it is not obvious that the same behavior should hold for other architectures or modalities. Even within vision, evaluation on ViTs would help determine whether the observed member/non-member geometry is a general phenomenon or partly specific to convolutional networks.

Questions for authors
  1. How much of SharpMIA’s improvement is attributable specifically to critical sharpness? Could the authors report the same learned combiner with (i) trace only, (ii) trace + difficulty, (iii) critical sharpness only, and (iv) trace + critical sharpness + difficulty, using otherwise identical features and training procedures?
  2. How does SharpMIA perform under a matched target-query budget?
  3. How sensitive is the attack to mismatch between reference and target models? For example, what happens if reference models use a different architecture, initialization/training recipe, augmentation strategy, or slightly shifted training distribution? Since the reference models determine both the walking direction and calibration, this seems important for assessing the practical black-box threat.
  4. Can the authors provide a stronger explanation for why membership produces wider basins in input space? In particular, is the effect primarily associated with memorization, confidence, example difficulty, or generalization?
  5. Does the phenomenon generalize beyond ResNets?
Reviewer_LcSGOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SharpMIA, which characterizes the loss geometry around a given sample by identifying the critical step size—the distance required for the loss to return to its initial level following an input perturbation. This signal is estimated using a reference model and black-box queries; by subsequently integrating information on curvature and sample difficulty, the method determines whether a sample belongs to the training set.

Critical strengths
  1. The paper reproduces 11 baseline attacks, providing a broad empirical comparison that supports the effectiveness of the proposed method.
  2. The paper introduces critical sharpness as an effective membership inference signal, supported by a coherent theoretical motivation.
  3. The black-box attack achieves performance close to its white-box counterpart, which is a strength of the method.
Critical weaknesses
  1. The improvements over state-of-the-art methods are modest in several settings. Do these small gains reflect an inherent limit on achievable attack performance, or limited additional information provided by the proposed signal? If so, some theoretical analysis of potential performance limits would help clarify this.
  2. The ablation study is insufficient. The standalone λv\lambda_v signal performs relatively poorly, whereas the full SharpMIA pipeline achieves substantially better results. It is unclear how much of this improvement comes from the new signal, calibration, the difficulty prior, or the learned combiner. Controlled ablations that isolate these components and clearly document changes in experimental settings are needed.
  3. The method requires a relatively high query budget. Comparisons under matched target-model query budgets would help establish whether its performance gains justify the additional cost.
Questions for authors

See S&W

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

This paper proposes SharpMIA, a black-box membership inference attack based on input-space loss geometry. The authors introduce Critical Sharpness, defined through the distance required for the loss to return to its initial value along a specified direction. They estimate this quantity using reference-model gradients and queries to the target model, then combine it with Hessian trace features and a difficulty prior using an MLP. Experiments are conducted on CIFAR-10, CIFAR-100, and ImageNet.

Critical strengths
  1. The use of finite-distance loss behavior as a membership inference signal is an interesting direction for exploring input-space geometry.
  2. The paper combines theoretical analysis with empirical validation to extend the proposed metric to black-box settings.
  3. The evaluation covers multiple datasets and considers attack performance at low false-positive rates.
Critical weaknesses
  1. Critical Sharpness is closely related to Hessian-based curvature and underperforms Hessian trace as a standalone signal. It remains unclear whether it provides additional membership information beyond existing curvature features. A controlled ablation removing Critical Sharpness from the full framework is needed to establish its independent contribution.
  2. The paper does not explain why reference-model gradients can reliably approximate the target model's Critical Sharpness. In particular, local rank-one structure neither guarantees cross-model gradient alignment nor extends directly to the finite-distance loss trajectory. The theoretical justification remains incomplete.
  3. A convincing evaluation would be to incorporate the proposed Critical Sharpness features into several existing MIA methods, such as LiRA, RMIA, RAPID, and LDC-MIA, and compare their performance before and after feature integration under consistent evaluation settings.
  4. Performance gains over existing methods are limited in some settings. Compared with RAPID, SharpMIA improves AUROC by approximately 1.30, 0.29, and 0.05 percentage points on CIFAR-10, CIFAR-100, and ImageNet, respectively. On ImageNet, the overall performance is nearly identical, and SharpMIA does not achieve the best TPR at 0.1% FPR.
  5. The derivation relies on a truncated second-order Taylor expansion but subsequently uses exact equalities without accounting for higher-order terms. In particular, η=2/R(g)\eta = 2/R(g) is exact only for the quadratic approximation, not necessarily for the original loss function.
Questions for authors
  1. Can the authors provide an ablation study that removes Critical Sharpness from the full framework while keeping all other components unchanged?
  2. Why can reference-model gradients reliably approximate the target model's Critical Sharpness?
  3. How do the authors justify the additional computational and query costs?
Reviewer_sgW7OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

This article proposes a membership-inference attack based on the behavior of the loss around an input. Conceptually, the attack measures how far the input can be moved in a selected direction before the model’s loss returns to its original value. The inverse of this return distance is used as a sharpness score. To obtain the sharpness score without accessing the target model’s gradients, the authors average normalized gradients from reference models that were not trained on the candidate example. This provides the direction of the walk, while black-box queries estimate the target model’s gradient magnitude and return distance. In the experiments, the resulting score is highly rank-correlated with the score computed using the target model’s true gradient. The final attack relies on a learned classifier that combines the raw and reference-calibrated sharpness with corresponding curvature scores and a reference-loss measure of example difficulty.

Critical strengths
  • Substituing the direction of the target model gradients with the reference gradients makes sense informally and also works surprisingly well in practice. The resulting sharpness scores have Spearman correlations between 0.9780.978 and 0.9900.990.

  • The evaluation compares against eleven baselines on standard vision benchmark datasets. SharpMIA is consistently competitive. On CIFAR-100, and the false-positive rates appear subtantially lower.

Critical weaknesses
  • I think that the proposed definition of critical-sharpness has potentially degenerate edge cases that are not fully addressed. Suppose the target gradient at a particular input is zero such that the loss remains at its initial value locally, the sharpness score is now undefined (since the true return distance is 0). Its unclear how the definitions need to be modified to account for this edge case. Further, the size of local region under around an input where the loss of the target model is constant has implications for membership inference but at present the definitions appear to not distinguish between the case when such a local region is infinitesimally small vs large.

  • The black-box construction for finding the sharpness score depends on the reference-derived direction being aligned with the target model’s gradient. Unlike the true negative-gradient direction, it is not guaranteed to reduce the target loss initially. Although the reported agreement is strong when the reference and target models share the same task, architecture, and data distribution, but it is unclear whether the observed correlation ican extrapolate to other choices of task/architecture.

  • The method may also confuse exact membership with proximity to a training example. If a nonmember is very close to a member, training on the member may create a broad low-loss neighborhood containing both examples. They may then have similar return distances even though only one was observed during training.

  • The proposed attack requires far more target-model evaluations than the baseline comparisons for example i think RAPID attack only needs 2 evaluations per candidate but SharpMIA might require on the order of 100 evaluations.

Questions for authors
  • The final attack combines multiple ideas - sharpness, calibration with reference models, a learned classifier, etc. How should we understand the contribution of the return-distance statistic alone?

  • how often does the reference-derived direction fail to decrease the target loss?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.