acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Model Inversion Attacks Generalize a Privacy-Robustness Tradeoff

Abstract

In this paper, we show that standard evaluations of high-resolution Model Inversion Attacks (MIAs) significantly underestimate training-data privacy leakage. State-of-the-art privacy defenses, standard training techniques such as MixUp and Adversarial Training, and undefended models all leak training images at rates to higher on FaceScrub under simple adaptive changes to the attack, with the largest increases among defenses reporting the strongest privacy. We further show that measured leakage depends on the feature basis of the external classifier used to evaluate reconstructions: for the same reconstructed images, an adversarially trained Inception evaluator identifies the targeted identity at different rates than the standard Inception evaluator. Our results suggest that standard MIA evaluation can mistake optimization and measurement failures for privacy. These underestimated leakage rates also concealed a broader relationship between privacy and adversarial robustness. Once we adapt the attack and vary the evaluator, reconstruction leakage closely tracks adversarial robustness across recent defenses and standard training regimes, suggesting that robustness provides an attack-agnostic proxy for reconstruction vulnerability that applies far more broadly than previously theorized. This raises an open question: can a practical defense reduce training-data reconstruction without paying a corresponding cost in adversarial robustness?

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
model inversion, privacy, reconstruction, adaptive attacks, robustness, adversarial examples, manifold hypothesis, perceptually aligned gradients

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_w2AFOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper revisits the evaluation of high-resolution MIAs. It proposes simple loss-function and learning-rate adaptations, which significantly improve attack performance, showing that prior defense claims may overstate privacy protection under fixed attack setups. The authors also score the reconstructions with both standard and adversarially trained Inception evaluators, demonstrating a strong dependence of measured leakage on the choice of evaluator.

With these adaptations, they show that prior findings on the privacy-robustness trade-off in generic MIA defenses can be generalized to recent anti-gradient defenses as well. Extensive experiments show (1) improved attack performance across multiple datasets, defenses, attacks, and target model architectures, (2) disagreement between standard and robust evaluators, and (3) an empirical association between robustness and reconstruction quality.

Critical strengths
  1. Revealing overfitting issues in existing MIA defenses: This work reveals an issue of MIA defenses overfitting to fixed attack configurations, highlighting the need for adaptive and robust privacy evaluation.
  2. Extending the claim of a privacy–robustness trade-off: The paper extends the findings of prior work on the privacy-robustness trade-off to the most recent MIA defenses, highlighting an important future direction: protection against MIAs without sacrificing adversarial robustness.
Critical weaknesses
  1. Unclear interpretation of evaluator disagreement: While Section 1 states that neither evaluator is treated as ground truth, Section 3 interprets the lower AttAcc of the standard evaluator as evidence that it underestimates privacy leakage. The disagreement demonstrates sensitivity to evaluator choice, but it remains unclear how to interpret the higher scores of the robust evaluator and which better reflects the actual leakage.
  2. Lack of qualitative analysis: This work relies on quantitative metrics across all experiments. Providing qualitative analysis or visualizations of reconstructed samples could help readers better understand the improvement of adaptive attacks and the disagreement between standard and robust evaluators.
Questions for authors
  1. Generalizability of adaptive attacks: Table 5 shows that adaptive attacks only improve IF-GMI against RoLSS and NegLS. Is there any explanation for why the adaptation is much less effective for IF-GMI against other defenses?
  2. Effect of robust evaluators on non-AT models: In Tables 2 and 6, it seems that AttAcc increases for AT models when switching to robust evaluators, but the trend is reversed for many non-AT defenses. Is there any explanation or intuition behind this observation?
7 comments

We sincerely thank you for highlighting how our paper “reveals an issue of MIA defenses overfitting to fixed attack configurations.” We especially appreciate your recognition that this raises “an important future direction: protection against MIAs without sacrificing adversarial robustness.”

We understand that your main remaining reservations, and the reason for a score of 3 rather than a clear-accept 4, concern:

  1. Adding qualitative evidence;
  2. How to interpret disagreement between the standard and robust evaluators, including why robust evaluation increases measured leakage for AT models while decreasing it for many non-AT models;
  3. Why some PPA adaptations transfer less consistently to IF-GMI.

To address the qualitative and evaluator concerns, we add three complementary forms of qualitative evaluation (detailed in our comments below):

  1. Example face-reconstruction grids across all 13 defenses × 3 attack settings: vanilla PPA, adaptive PPA C-LR, and adaptive PPA N-LR;
  2. A human validation study comprising 2,225 human qualitative-comparison judgments (with University Human Subjects Exempt approval);
  3. The Ho et al. (CVPR Findings 2026) Gemini-based LLM simulations of human qualitative-comparison judgments across all 39 defense–attack conditions.

We strongly feel that this paper's results advance the MI/privacy field. Its main result shows that leading MIA defenses’ apparent privacy can reflect optimization and measurement failures, and that correcting these failures reveals a fundamental relationship between training-data privacy and adversarial robustness across defenses previously thought to be exceptions. We would appreciate it if you would consider improving your score in light of the new qualitative, human, and Gemini evaluations and the clarifications below.

  • Re: Critical Weaknesses 1–2. Qualitative analysis and which evaluator better reflects actual leakage.

Thank you—we added qualitative face-reconstruction grids for every defense under vanilla PPA, adaptive PPA C-LR, and adaptive PPA N-LR.

Qualitative Results: LINK https://i.imgur.com/1RRuHGS.jpeg

The qualitative examples mirror the quantitative results: adaptive attacks recover more recognizable identity-bearing facial features than vanilla PPA for the defenses on which adaptation produces the largest leakage increases.

We further test this interpretation, and address evaluator concerns, using two independent qualitative evaluations:

  1. 2,225 human responses. We recruited human raters to assess identity, demographics, and family resemblance over 2,225 ratings.
  • (a) Adaptive attacks. Human raters score our adaptive attacks' reconstructions higher than those from vanilla PPA on the NegLS defense. Both adaptive attacks improve human ratings by approximately 7-17 percentage points across the three questions. Across both adaptive attacks, the improvements are statistically significant even after BH correction for the primary Same Individual and Same Demographics questions; Family Resemblance is also higher. Our two adaptive attacks obtain similar performance on each question (BH p ≥\geq 0.421). Human Evaluation of adaptive attack results plot: LINK https://i.imgur.com/FPbTMPr.png

  • (b) Robust evaluator. As described fully in our Response to Reviewer #1 (U2WM) above, we find that our Robust Inception feature-space distance evaluation correlates significantly better with human ratings than the vanilla Standard Inception evaluation model. For Inception distance, the improvement is statistically significant for both the primary Same Individual question (p = 0.008) and Family Resemblance question (p = 0.033). Evaluator confidence shows that our Robust Evaluator offers the same directional improvement over a standard evaluator. Note that this statistical significance was obtained despite the limited statistical power in a small-scale human study with just five human raters. Given the positive point estimates for confidence, we expect its differences to reach greater statistical significance in a larger-scale study. Human Evaluation of robust Inception plot: LINK https://i.imgur.com/rfWHzmx.png

  1. Comprehensive Gemini evaluations. Following Ho et al. (CVPR Findings 2026), we use Gemini to conduct additional qualitative comparisons across all 13 defenses × all 3 (vanilla + adaptive) attacks. As detailed in our Response to Reviewer #1 (U2WM) above, our adaptive attacks generally improve reconstruction success under Gemini as well. Results Plot: LINK https://i.imgur.com/PU70Eh0.png
  • Re: Q2. “In Tables 2 and 6, it seems that AttAcc increases for AT models when switching to robust evaluators, but the trend is reversed for many non-AT defenses. Is there any explanation or intuition behind this observation?”

Our human study provides supporting evidence: Robust Inception distance correlates more strongly with human ratings on all three questions. But the broader contribution is not to replace one evaluator with another. This directional pattern is exactly what the evaluator feature-basis account predicts. Standard and robust Inception measure leakage through different feature bases: the standard evaluator places more weight on non-robust predictive features, while the robust evaluator places more weight on robust, visually meaningful features. This explains why evaluator substitution can lower measured leakage for models that rely more heavily on non-robust features while sharply increasing it for AT models.

Our manuscript states this point explicitly: “Neither evaluator is treated as ground truth; disagreement between them measures how strongly reported leakage depends on the evaluator’s feature basis” (lines 127–128). A low leakage score under one arbitrarily selected evaluator—such as the standard Inception evaluator used to assess virtually all defenses—therefore does not establish privacy when another evaluator detects significant leakage in the same reconstructions. As the manuscript concludes, reconstructed images should be “evaluated by more than one arbitrarily chosen classifier feature basis” (lines 158–160). Each evaluator detects leakage through one particular feature basis, and using multiple evaluators reveals leakage through different types of features (we will clarify the description of this point in Section 3--thank you for pointing out this opportunity to improve clarity).

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

The paper shows that the standard evaluation of high-resolution Model Inversion Attack (MIA) underestimates training data leakage. The adaptive attack, with changing in loss and learning rate, paper proposed can increase training image leakage 1.16x to 6.59x higher on Facescrub in undefended models as well as models with state-of-the-art defense. The paper also studies the dependence to evaluator’s feature bias of measurement and show that neither standard and adversarially trained Inception Evaluator are considered as ground truth. In addition, Figure 1 also shows the failure comes from standard evaluation. The paper also provides analysis on trade off between privacy and robustness and gradient and show that this is more robust and makes more leakage, and a question on is it possible to design a defense against training data reconstruction that obtains privacy without a corresponding sacrifice of adversarial robustness? Is raised. Finally, the paper recognizes the problem of the field is to evaluate defense with fixed parameters attack, which is not good, and the paper shows that with adaptive attack, the attack accuracy can be recovered with only changing setup of attacks.

Critical strengths

The research question on problems in standard evaluation for Model Inversion Attack is clear and important.

The adaptive attack is simple but effective and shows it can increase attack utility.

Analysis provides meaningful comparison on attack setups, the reason for adaptive version is simple and reasonable.

The experiments are comprehensive with 13 target-models ×3 PPA variants ×2 datasets ×2 architectures = 156 conditions.

The adaptive attack’s gain is consistent across defenses.

Critical weaknesses

The paper should include paper [A] to compare the argument on MIA evaluation.

The paper does not provide metrics by humans to see if it is also misleading. Since some defenses use user study, it is useful to see if adaptive attack actually can improve the quality of attack image under evaluation with humans.

The paper can not provide a better evaluation, which should be stated as an open question.

The stronger attack can create false positive samples as the paper [A] observe, the paper should do analysis on that to see if adaptive attacks create them.

The "adaptive" word in adaptive attack methods seems too strong since the paper only adapts with learning rate and negative label smoothing instead of building adaptive attack for specific defense method one by one.

The paper claims that adaptive attack can improve attack accuracy. However, the defense also can be finetuned again if they have adaptive attack to try, and defense has data to try to defend against adaptive attack is reasonable since they can have both private and a best public dataset while attackers in reality do not have.

[A] Ho, Sy-Tuyen et al. “Revisiting Model Inversion Evaluation: From Misleading Standards to Reliable Privacy Assessment.” 2026 IEEE/CVF Conference on Computer Vision and Pattern Recognition- FINDINGS Track (CVPRF) (2025): 8073-8082.

Questions for authors

Can the authors discuss and compare with the paper [A]?

Can the authors provide user study for the images generated by adaptive attack ?

Can the authors provide adaptive defense with hyperparameters finetuning privacy reports when an adaptive attack results are included to finetune the adaptive defense? Since defense normally also need finetune to achieve best privacy-utility trade-off, only adapt attacks is not fair with defense and do not consider adaptability of current defenses.

10 comments

We thank you for the careful and generally positive review, and for highlighting that the research addresses a “clear and important” problem, that the “analysis provides meaningful comparison on attack setups,” and that the motivation for our adaptive attacks is "simple and reasonable."

Your remaining reservations, and the apparent reason for a score of 3 rather than a clear-accept 4, concern:

  1. Ho et al. (CVPR Findings 2026) discussion and Gemini-based evaluations;
  2. Human evaluations;
  3. Clarification regarding whether defenses can be retuned against our adaptive attacks if our attack adaptations are known by defenders in advance.

To address these points, we add new discussion and results from the Ho et al. Gemini evaluation and a human validation study (detailed in our comments below), and clarify how our paper directly addresses the defense-tuning problem you identify.

We strongly feel that this is a high-impact paper: it shows that a fundamental relationship between two core Trustworthy ML concepts—adversarial robustness and training data privacy—extends across all mainstream defenses, including all supposed exceptions, rather than the small handful identified in previous work (e.g., Torp et al., ICLR 2026). We would appreciate it if you would consider improving your score in light of the new MLLM and human evaluations and the clarifications below.

  • Re: Q1. “Can the authors discuss and compare with the paper [A]?”

We thank you for suggesting this complementary paper (Ho et al. CVPR Findings 2026) and evaluation. We discuss the relationship to our paper below, and we add the Gemini evaluation (see new plot below) to affirm our adaptive attack effectiveness. Our paper shows how measured leakage depends on the feature basis of the external classifier used to evaluate reconstructions: for the same reconstructed images, an Inception evaluator that is adversarially trained to have robust features finds different leakage rates than the standard Inception evaluator. This complements the Ho et al. finding that the standard evaluation classifier can incorrectly count visually unrelated reconstructions as successful attacks, including because of transferable Type-I adversarial features. Incidentally, those adversarial features are, by definition, non-robust ones that our robust evaluator is trained to deemphasize.

Following Ho et al.’s Gemini experiment, we compare each reconstructed face with four real photographs of the targeted individual using their Gemini image-comparison prompt. We use the same model family (Gemini-2-Flash) with the thinking budget set to 0 as in Ho et al.'s setup (note that Google changed Gemini 2's available sub-version since Ho et al.'s publication). We evaluate all 3 attacks × 13 defenses in our paper, including vanilla PPA and our adaptive PPA N-LR and adaptive PPA C-LR attacks.

GEMINI RESULTS PLOT: https://i.imgur.com/PU70Eh0.png

PPA N-LR raises the Gemini score on all 13 targets; e.g., NegLS rises from 0.026 to 0.100 and TL-DMI from 0.030 to 0.114.

The adaptive attacks generally improve reconstruction success under Gemini evaluation as well. Because Gemini does not rely on the brittle standard Inception feature basis exploited by the false positives identified in Ho et al. (2026), this provides independent evidence that our adaptive gains are robust across different types of evaluators.

We also test Gemini on a positive control: one real FaceScrub photograph as Image A and four other real FaceScrub photographs of the same individual as Set B. Gemini incorrectly rejects 16.8% of these genuine matches, compared with approximately 3.2% reported by Ho et al. This shows that the current Gemini API with its settings for inference and preprocessing (changed by Google since Ho et al.’s publication) now yields more false negatives on real face images. Together, the Inception, Gemini, and human evaluations support our paper’s recommendation that (1) leakage should not be assessed through a single evaluator, and (2) defenses should not be designed to overfit to just one evaluator.

  • Re: Q2. “Can the authors provide user study for the images generated by adaptive attack?”

Thank you—we added a human validation study specifically comparing images produced by the standard and adaptive attacks, with University Human Subjects Exempt approval. We recruited five undergraduate raters and built an online survey that presented an MI-reconstructed face from a ResNet-152 target alongside four original FaceScrub photographs of the targeted individual. Raters provided a total of 2,225 responses by answering three questions on a scale from 0% to 100%:

  1. Same individual: Confidence that the reconstruction depicts the same person as the reference images.
  2. Same demographics: Similarity in apparent race, gender, and age.
  3. Family resemblance: Resemblance of facial features.

The first question follows Ho et al., replacing binary yes/no with graded confidence. The other two test whether reconstructions reveal demographic attributes or recognizable facial features even when they do not perfectly recover identity.

To make a human study feasible during the brief rebuttal period, we focus directly on the conditions relevant to your concern: standard PPA, PPA C-LR, and PPA N-LR against NegLS, where adaptation produced the largest improvement per Fig. 2 on p.7 in our paper. We also include standard PPA against AT*, where changing the evaluator produced the largest difference, and NoDef as the baseline.

Human Evaluation of adaptive attack results plot: LINK https://i.imgur.com/FPbTMPr.png

Human raters score reconstructions from our adaptive attacks higher than those from vanilla PPA on the NegLS defense. Both of our adaptive attacks improve human ratings by approximately 7-17 percentage points across the three questions. The improvements are statistically significant for both adaptive attacks even after Benjamini-Hochberg correction on the primary Same Individual question and also the Same Demographics question. Family Resemblance is also higher for both adaptive attacks. Our two adaptive attacks obtain similar performance on each question (all BH p ≥\geq 0.421). This directly shows that the increased attack success reflects improved reconstructions rather than only additional evaluator false positives.

The study also addresses your broader concern regarding whether the paper provides a better evaluation. Robust Inception feature-space distance correlates more closely than standard Inception distance with human ratings on all three questions:

Human evaluation correlation with standard and our robust evaluator: LINK https://i.imgur.com/rfWHzmx.png;

We emphasize that no automated evaluator should be treated as ground truth, particularly because humans often find identity resemblance to be difficult to measure. Our manuscript therefore states: “Neither evaluator is treated as ground truth; disagreement between them measures how strongly reported leakage depends on the evaluator’s feature basis.” The new results support evaluating reconstructions using both feature bases, rather than replacing one arbitrarily chosen evaluator with another. Put differently, any one evaluator conveys a single distribution over the properties of leakage (e.g., likeness, robust features, non-robust features, demographic features, etc.); a better evaluation would test under multiple evaluators to avoid overfitting to any one.

Reviewer_U2WMOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies whether standard evaluations of high-resolution model inversion attacks (MIAs) underestimate privacy leakage. The authors test simple adaptations to PPA by changing its loss function and learning rate. They also evaluate the same reconstructed images using standard and adversarially trained Inception-v3 classifiers.

Experiments cover 13 target-model conditions across FaceScrub and CelebA, using ResNet-152 and DenseNet-169. The adaptive attacks often achieve higher reconstruction accuracy than standard PPA. The authors further show that these evaluation changes improve the prediction of reconstruction leakage from adversarial robustness on held-out defenses and training regimes.

Critical strengths
  1. I appreciate that the paper questions a common practice in MIA evaluation. The results are quite striking. For example, on FaceScrub with ResNet-152, NegLS AttAcc@1 increases from 0.105 to 0.692 under adaptive PPA. This is a useful finding for researchers evaluating MIA defenses.

  2. The authors test a wide range of defenses and training methods across two datasets and two architectures. I also appreciate that they report L2-Face alongside attack accuracy, which provides additional evidence about reconstruction quality.

  3. Rather than only improving an attack, the paper revisits the privacy–robustness relationship. The improved prediction on held-out defenses is interesting and provides a useful direction for further research.

Critical weaknesses
  1. The reliability of the robust evaluator is not fully established. I agree that different evaluators can give very different results. However, I am not sure which evaluator better measures recognizable identity information. For example, in Table 2, the AttAcc@1 of NoDef decreases from 0.823 to 0.463, while that of AT* increases from 0.430 to 0.901. These results show evaluator dependence, but do not necessarily establish that the robust evaluator gives a more accurate estimate of privacy leakage.

  2. The connection to recent MIA evaluation research could be stronger. Ho et al. (Revisiting Model Inversion Evaluation, CVPR 2026 Findings) showed that conventional MIA evaluators can incorrectly classify visually unrelated reconstructions as successful attacks. I think this work is particularly relevant because the current paper also questions evaluation reliability. The authors study a different issue, but I would like to understand how their evaluation accounts for such false positives.

  3. I am not fully convinced by the generality of the privacy–robustness relationship. The regression uses 16 observations for fitting and 32 held-out observations, all from the same two facial datasets and two architectures. The results support a stronger relationship across the tested defenses, but I find the broader claim of an attack-agnostic privacy–robustness tradeoff less clear. The more limited improvements under IF-GMI also suggest that the conclusions may depend on the attack setting.

  4. The contributions of the individual attack changes are not well separated. Both C-LR and N-LR change the loss function, learning rate, and learning-rate schedule together. Although the results show that these combinations work, I cannot tell which change is mainly responsible for the improvement. A small ablation would help explain the observed optimization failure.

Questions for authors

Q1. Could the authors provide an independent evaluation of reconstruction quality, such as human identity judgments or additional face recognition models? I would like to understand whether the higher Robust-AttAcc@1 scores reflect more recognizable identity information.

Q2. Could the authors discuss the differences between their evaluation method and Ho et al. (CVPR 2026 Findings)? In particular, how can we tell whether the additional successful reconstructions include false positives?

Q3. Could the authors provide a more detailed analysis of the regression results across different defense families or attack settings? This would help me understand how far the claimed privacy–robustness relationship generalizes.

Q4. Could the authors separately evaluate the effects of changing the loss function and learning rate? This would help identify which modification leads to the improvement.

6 comments

We sincerely thank you for your thoughtful and constructive review, your generally positive assessment, and your comments that the paper’s results are “quite striking” and a “useful finding,” and that its contribution to the generalized privacy–robustness relationship is “interesting and provides a useful direction for further research.” We also appreciate your many positive comments about the breadth and “wide range of defenses and training methods” in the experiments.

We understand that your remaining reservations and reason for a score of 3 instead of “full accept” arise from four questions. Below, we added new results addressing all four:

NEW RESULTS detailed in our comments below:

  1. Human evaluations, including University Human Subjects Exempt approval;
  2. The new Gemini evaluation proposed by Ho et al. (CVPR Findings 2026);
  3. Extensive new regression robustness tests across 127 allowed fitting/held-out defense splits;
  4. New separate loss-function and learning-rate ablations.

We would be grateful if you would consider improving your score accordingly.

  • Re: Q1. “Could the authors provide an independent evaluation of reconstruction quality, such as human identity judgments or additional face recognition models?”

Thank you - we added the human validation study you suggested, with University Human Subjects approval. We recruited five undergraduate raters and built an online interface that presents an MI-reconstructed face from a ResNet-152 target model alongside four original FaceScrub photographs of the targeted individual. Raters provided 2,225 total responses by answering three questions on a scale from 0% to 100%:

  1. Same individual: Does Image A depict the same individual as the images in Set B? How confident are you that the answer is yes?
  2. Same demographics: To what degree does the person in Image A share the same demographics (e.g., apparent race, gender, age) as the person in the Set B images?
  3. Family resemblance: To what degree do the facial features of the person in Image A bear family resemblance to the person in the images in Set B?

Question Example: LINK https://i.imgur.com/6PJR7XE.png

Survey question 1 above is the question from Ho et al. (CVPR Findings 2026), with graded confidence in place of a binary yes/no answer. Questions 2 and 3 assess whether reconstructions reveal (2) demographic attributes or (3) robust facial features even when they do not perfectly recover identity. To make a rebuttal-period study feasible, we focus the scope on the settings most relevant to your question: standard and adaptive PPA on the defense where (per Fig. 2 and Tables 1 and 2 in the manuscript) the attack adaptations produced the largest improvement (NegLS); standard PPA against AT*, where Robust evaluator substitution produced the largest change; and standard PPA against NoDef as the baseline.

Human Validation Results: The human results favor the robust inception evaluation metric over the standard inception metric on all three questions.

Results: LINK https://i.imgur.com/rfWHzmx.png

Our Robust Inception evaluation model correlates significantly better with human ratings than the vanilla Standard Inception evaluation model. For Inception distance, the improvement is statistically significant for both the primary identity question (p = 0.008) and family resemblance question (p = 0.033), which is designed to capture the robust features humans care about. It is also borderline statistically significant for the demographics question (p = 0.057). The primary identity question result remains significant even after BH correction (adjusted p = 0.047). Evaluator confidence (top row) shows that our Robust Evaluator offers the same directional improvement over a standard evaluator. We emphasize that this statistical significance was obtained despite the limited statistical power in a Rebuttal-scoped human study with just five human raters. Given the positive point estimates, we expect the evaluator-confidence differences to reach greater statistical significance in an expanded study with more human raters.

The new human results corroborate our paper's point that the field has overfit the design of its defenses against a single evaluation classifier feature basis (i.e., vanilla Inception), allowing apparent privacy to reflect optimization and measurement failures that poorly correlate with human judgment. The manuscript states: “Neither evaluator is treated as ground truth; disagreement between them measures how strongly reported leakage depends on the evaluator’s feature basis.” Its evaluation recommendation is that reconstructed images should be “evaluated by more than one arbitrarily chosen classifier feature basis.”

  • Re: Q2. “Could the authors discuss the differences between their evaluation method and Ho et al. (CVPR 2026 Findings)? In particular, how can we tell whether the additional successful reconstructions include false positives?”

Thank you - we added the Ho et al. Gemini evaluation and a positive-control experiment to assess its reliability in our setting.

Ho et al. show the interesting result that conventional evaluation classifiers can label visually unrelated reconstructions as successful attacks, including through transferable Type-I adversarial features. Our paper provides complementary contributions by showing that measured leakage depends on (and mainstream defenses are overfit to) the feature basis of the fixed external Inception classifier used to evaluate reconstructions.

We replicate Ho et al.’s (CVPR Findings 2026) protocol by submitting their image-comparison prompt with one reconstructed face and four real reference images to Gemini-2-Flash with the thinking budget set to zero. We apply this Ho et al. (2026) evaluation to all {13 defenses ×\times 3 attacks} (i.e., vanilla PPA + our two adaptive attacks) in our paper.

Our results differ somewhat from Ho et al.'s, as Google changed the available Gemini 2 model sub-version (2.0 → 2.5) as well as its inference configuration and image preprocessing, since the publication of Ho et al.'s results.

GEMINI RESULTS PLOT: https://i.imgur.com/PU70Eh0.png

Our adaptive attacks generally improve success under this evaluator as well. PPA N-LR raises the Gemini score on all 13 targets, and relative to NoDef under the same attack, NegLS rises from ~27% (0.026 vs. 0.096) to ~78% (0.100 vs. 0.128), so this result does not depend on Gemini’s absolute calibration. However, the relatively low absolute Gemini scores across the board require an additional check. We therefore repeat the comparison with a real FaceScrub image as Image A and four other real FaceScrub images of the same individual as Set B. Gemini rejects 16.8% of these genuine positive comparisons, versus approximately 3.2% reported by Ho et al. This shows that the current Gemini API and its current inference and preprocessing settings (changed by Google since Ho et al.'s publication) now yield appreciable false negatives on natural images.

Under vanilla PPA, the evaluators disagree on whether AT* leaks more or less than NoDef. Gemini assigns AT* slightly lower leakage than NoDef (0.076 vs. 0.096), whereas standard Inception assigns it far lower leakage (AttAcc@1: 0.430 vs. 0.823). Robust Inception reverses this ranking, assigning AT* nearly twice the leakage of NoDef (Robust-AttAcc@1: 0.901 vs. 0.463). Thus, evaluator choice changes both the magnitude and the direction of the apparent privacy benefit, supporting our recommendation against certifying privacy with any single evaluator.

Comments on the paper

2 comments
AuthorsOpenReview ↗

**Overall Response to the Area Chair and All Reviewers**

We thank all reviewers for their thoughtful, detailed, and positive reviews, and for highlighting that the paper presents “quite striking” results [U2WM]; addresses a “clear and important” problem using adaptive attacks that are “simple but effective” [HUtF]; and “reveals an issue of MIA defenses overfitting to fixed attack configurations” [w2AF], while our extension of the privacy–robustness relationship to recent defenses "is interesting and provides a useful direction for further research" [U2WM]. We also appreciate reviewers’ recognition of the breadth of our experiments, described as covering a “wide range of defenses and training methods” [U2WM] and as “comprehensive” with 156 experimental conditions [HUtF].

Collectively, the reviews affirm the paper’s central contribution: standard evaluations of high-resolution Model Inversion Attacks substantially underestimate training-data privacy leakage. Simple adaptive changes increase measured leakage by 1.16–6.59× on FaceScrub, while the same reconstructed images receive sharply different leakage scores depending on the evaluator’s feature basis. Once we adapt the attacks and vary the evaluator model, reconstruction leakage closely tracks adversarial robustness across all mainstream defenses and standard training regimes. This suggests that standard MIA evaluation has mistaken optimization and measurement failures for privacy. It also reveals a key open research frontier for the field: whether a defense can meaningfully improve privacy without resorting to brittle modifications, and without becoming vulnerable to the separate threat of adversarial examples.

We add six new sets of results to address every remaining concern:

  1. Human evaluations (including University Human Subjects Exempt approval) showing that our Robust Inception evaluator obtains significantly better correlation as a proxy for 2,225 collected human judgments in terms of Inception distance: Example question: https://i.imgur.com/6PJR7XE.png Results: https://i.imgur.com/rfWHzmx.png; (Now added to revised manuscript in Appendix G: Figs. 9 and 10, pages 26 and 27)
  2. Human evaluations on our adaptive attacks: https://i.imgur.com/FPbTMPr.png; (Now App. G: Fig. 11, p. 27)
  3. The new Gemini-based privacy evaluation proposed by Ho et al. (CVPR Findings 2026): https://i.imgur.com/PU70Eh0.png; (Now App. G: Fig. 8, p. 26)
  4. Extensive new regression cross-validation tests showing that the privacy–robustness result holds across 127 allowed fitting/held-out defense splits; https://i.imgur.com/llsmYmR.png; (Now App. G: Tab. 11, p. 25)
  5. New separate loss-function and learning-rate ablations: https://i.imgur.com/hTIUoB9.png; (Now App. G: Tab. 12, p. 28)
  6. Qualitative analysis and example face reconstructions across all 13 defenses × 3 attack settings: vanilla PPA, adaptive PPA C-LR, and adaptive PPA N-LR. https://i.imgur.com/1RRuHGS.jpeg. (Now App. G: Fig. 7, p. 25)

All six new sets of results corroborate the submitted findings and significantly strengthen the empirical evidence, as we describe in detail below.

We strongly believe that this is a high-impact paper. We address each reviewer’s remaining questions in detail below. We would be grateful if reviewers would consider improving their scores in light of the new human, Gemini, cross-validation, ablations, qualitative reconstructions, and clarifications below.

Area_Chair_uu53OpenReview ↗

Reviewer's responses to authors

Dear reviewers,

If not done already, could you please take a look and reply briefly to the "Official Comment(s) by Authors", possibly by an acknowledgement that you have seen it.

Feel free to respond in more detail, in particular, if you have further questions or opinions to the replies from the authors.

With best regards, an AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.