acceptodds
Under review as a conference paper at ICLR 2027

NOWHERE: Evidence-Space Protection for Geolocation Privacy

17 views

Abstract

Multimodal large reasoning models can infer the locations of ordinary images from fine-grained visual clues, posing serious geoprivacy risks. Existing methods disrupt overall image features or use coarse text descriptions to define attack targets. However, they may overlook key location clues in individual scenes. Moreover, static attacks may keep targeting already weakened clues while leaving evidence that still supports localization. Therefore, effective protection requires targeting specific visual evidence and adjusting the optimization focus as this evidence changes. Although models use different internal representations, their location predictions rely on visual evidence from the same physical scene. This evidence provides a shared target for cross-model attacks. Building on this insight, we propose NOWHERE, a black-box geoprivacy protection framework based on Evidence-Space Transfer. Evidence Space Construction (ESC) locates image-specific geographic clues and builds multidimensional evidence subspaces in each surrogate model’s feature space. It also uses a complementary scene reference to guide content preservation. Dynamic Evidence Tracking (DET) compares clue features before and after perturbation, shifting optimization toward clues that remain insufficiently suppressed. Extensive experiments show that NOWHERE provides stronger geoprivacy protection while preserving visual quality and non-geographic semantic consistency. On Im2GPS3k, NOWHERE offers stronger protection against GPT-5.6-Terra’s location inference. The model locates only 4.2% of protected images within 1 km, compared with 6.8% under the strongest baseline. This helps users reduce location leakage while keeping their photos suitable for sharing.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Geolocation, Geoprivacy, Black-box attacks, Multimodal large reasoning models

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_P4XVOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes NOWHERE, a framework that protects image location privacy through transferable adversarial perturbations. Its Evidence Space Construction module builds subspaces for each image using local geographic clues. Its Dynamic Evidence Tracking module adjusts the allocation of local views based on the remaining clue responses. The framework also uses a complementary scene reference to preserve content unrelated to location. In the experiments, NOWHERE provides stronger location privacy protection than the compared baselines. The paper also includes ablation studies, semantic consistency tests, and image transformation tests. These tests further assess each module’s contribution, the preservation of image meaning, and the robustness of the protection.

Critical strengths
  1. The method is well designed and introduces a useful technical innovation. ESC preserves multiple feature directions from geographic clues, while DET updates clue priorities during optimization. Together, they allow the attack to target distinct clues and shift attention toward those that remain strong. The complementary scene reference also helps preserve content unrelated to location. This design gives a clear purpose to each component.

  2. The experiments are extensive and support the main claims. The evaluation covers two datasets, four black-box models, and five distance thresholds, with comparisons against GeoShield and ReasonBreak. On Im2GPS3k, GPT-5.6-Terra’s accuracy within 1 km falls from 6.8% under the strongest baseline to 4.2% with NOWHERE. The rank, fixed-view-budget, and shuffled-priority ablations support the main design choices. Semantic consistency tests and JPEG compression and blur tests also assess content preservation and robustness.

Critical weaknesses
  1. The limits of the protection could be discussed more clearly. The paper reports strong overall results, but gives little detail on cases where the model still finds the correct location. A few examples from the existing evaluation would help explain which visual clues remain difficult to suppress and when the protection may be less effective.

  2. The role of the evidence subspace in the final objective could be clearer. ESC keeps multiple feature directions, but the geographic suppression term uses a single reference vector eme_m. Table 4 supports the use of multiple directions. Still, the paper could better explain how they contribute to this reference vector and why this is more effective than simply averaging the clue features.

Questions for authors
  1. Could you provide a failure case analysis? A few examples from the existing evaluation would help show which visual clues are difficult to suppress and when the protection becomes less effective.
  2. How do the multiple directions in the evidence subspace contribute to the reference vector eme_m? Why is this better than simply averaging the clue features?

If my issue is resolved, I would be willing to increase my rating.

3 comments
AuthorsOpenReview ↗

Response to Reviewer P4XV

Dear Reviewer,

Thank you for your helpful suggestions.

WK1 & Q1. Provide failure case analysis.

A1. We have added a qualitative failure analysis to Section 5, “Limitations and Discussion,” on page 10, lines 511–518. Figure 6 on the same page shows the two protected images and the model’s explanations.

The examples show prominent signs that remain readable after protection. GPT-5.6-Terra localized the protected Sheraton Saigon hotel image and the Franklin D. Roosevelt Presidential Library and Museum sign image with errors of 2.82 km and 2.83 km, respectively. Its explanations cited the visible place names as location clues.

These examples suggest that readable signs can leave useful location clues even after geographic features are suppressed. This is particularly challenging when the goal is to preserve the photo's appearance. Removing the visible information more directly may require changing or obscuring parts of the image.

We include both protected images in Figure 6 to make this limitation concrete. The explanations suggest that signs contributed to localization, but do not establish that text was the only cause.

WK2 & Q2. How do the multiple directions in the evidence subspace contribute to the reference vector em ? Why is this better than simply averaging the clue features?

A2 (1). How the directions contribute to the reference. Let um,1,…,um,rmu_{m,1},\ldots,u_{m,r_m} be the retained orthonormal directions, and let gmg_m be the clean-image feature. The reference in Appendix A.2, Eq. (18), can be expanded as:

\qquad e_m = \frac{P_m g_m}{\|P_m g_m\|_2}.$$ Thus, $e_m$ combines the retained directions. Each direction contributes according to its alignment with the clean-image feature. Although the final reference is one vector, it depends on the full retained basis, not just the first direction. **A2 (2).** **Why this differs from averaging.** Mean pooling gives equal weight to the clue features and combines them into one vector directly. ESC first retains multiple directions from those features, then uses the clean-image feature to determine their contributions to the reference. This allows the reference to reflect how the image is represented within the evidence space. Table 4(a) tests this difference using the same clue features, a fixed scene-preservation reference, and static view allocation. Its Acc@25 km results are reproduced below. | Reference | GPT-5.6-Terra ↓ | Claude-Opus-4.8 ↓ | | :--- | ---: | ---: | | Mean | 14.8 | 14.6 | | Rank limit = 1 | 12.9 | 13.1 | | Rank limit = 2 | 12.0 | 11.7 | | Rank limit = 4 | 11.0 | 10.3 | | Rank limit = 8 | 10.8 | 10.6 | Retaining up to four directions improves protection over both averaging and the rank-one reference on both models. We have clarified how the retained directions define the reference in Section 3.2 on page 5, lines 232–233. Appendix A.2 now gives the expanded formula in Eq. (19) and explains the difference from mean pooling on page 17, lines 884–898.

Thank you for the clear responses and revisions. The failure cases and the explanation of how the evidence directions contribute to the reference vector address my concerns. I also read the responses to the other reviewers and appreciate the additional experiments and careful discussion. I am raising my score from 3 to 4.

AuthorsOpenReview ↗

Response to Reviewer P4XV

Thank you for your thoughtful review and for raising your score. We are glad that the revisions and additional experiments addressed your concerns.

Reviewer_zp42OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes NOWHERE, a transfer-based framework for protecting image geolocation privacy. Evidence Space Construction represents localized geographic clues using multidimensional visual subspaces, while Dynamic Evidence Tracking reallocates local views according to the remaining evidence during perturbation optimization. A complementary scene reference guides content preservation. Experiments on Google Street View and Im2GPS3k show substantial reductions in geolocation accuracy across four black-box VLMs, with additional analyses of evidence construction, view allocation, semantic consistency, and image fidelity.

Critical strengths

The evidence representation and adaptive optimization form a coherent contribution. Representing individual visual clues in a multidimensional subspace provides a concrete alternative to averaging clue features. Table 4 separately examines subspace construction and clue-dependent allocation, making the roles of these design choices reasonably clear.

The privacy improvements are substantial across several evaluation settings. For example, on Im2GPS3k, GPT-5.6-Terra’s accuracy within 25 km decreases from 15.5% under the strongest baseline to 9.0% with NOWHERE. The evaluation across two datasets and four target models makes the contribution practically relevant.

Critical weaknesses

Some consequential optimization settings remain unspecified. Appendix A.3 introduces a search over decreasing preservation weights, gradient-conflict adjustment, and candidate acceptance with a numerical tolerance. However, the candidate-weight sequence and acceptance tolerance are not specified. These choices determine the actual updates and the balance between suppression and preservation, so they matter for reproducing the reported results.

The visual-fidelity cost deserves more prominence in the main assessment. Appendix C already reports that NOWHERE has lower SSIM and higher LPIPS than GeoShield: 0.8302 versus 0.8832, and 0.2625 versus 0.1909, respectively. The stronger privacy protection is valuable, but readers should encounter this tradeoff alongside the main localization results, rather than primarily through qualitative examples and caption consistency.

Questions for authors
  1. Please specify the preservation-weight candidates, their search order, and the numerical acceptance tolerance used in Appendix A.3 and Algorithm 2. Which exact settings produced Table 1?
  2. Could the main results discussion explicitly connect Table 1 with the existing fidelity results in Table 5, describing the additional localization protection together with its measured visual cost?
3 comments
AuthorsOpenReview ↗

Response to Reviewer zp42

Dear Reviewer,

Thank you for pointing out these missing details.

WK1 & Q1. Please specify the preservation-weight candidates, their search order, and the numerical acceptance tolerance used in Appendix A.3 and Algorithm 2. Which exact settings produced Table 1?

A1. The weight search uses only surrogate-model scores, with no target-model queries.

  1. Preservation weights and search order. We try 1, 0.5, 0.25, and 0.125 on the adjusted preservation gradients, accepting the first passing candidate. Rejected trials leave the image and Adam states unchanged.

  2. Acceptance checks and tolerance. Before projection, the update uu must satisfy g⊤u>0g^\top u>0 and h⊤u≥0h^\top u\geq0, where gg and hh are the summed suppression and original preservation gradients. After projection, using the same sampled views, suppression must increase by at least 10−610^{-6} and preservation may decrease by at most 10−610^{-6}. These are fixed absolute thresholds.

  3. Fallback. If all candidates fail, we use weight 0 and the current suppression gradients, retaining Adam history without repeating the checks.

  4. Settings for Table 1. All NOWHERE results use these settings with 300 optimization steps.

We have specified these settings in Appendix A.3 on pages 19–20, lines 1013–1070, and updated Algorithm 2 on page 20 to match.

WK2 & Q2. Could the main results discussion explicitly connect Table 1 with the existing fidelity results in Table 5, describing the additional localization protection together with its measured visual cost?

A2. In the revised manuscript, we connect Tables 1 and 5 in Section 5, “Limitations and Discussion,” on page 9, lines 481–485. Page 10 presents the fidelity and human evaluation results in Tables 5–6, with the study described in lines 504–510. We summarize these additions below.

(1) Protection and image fidelity. On GPT-5.6-Terra, NOWHERE achieves an Acc@25 km of 9.0%, compared with 15.5% for GeoShield and 27.9% for ReasonBreak (Table 1). The fidelity results are:

MethodPSNR ↑ (dB)SSIM ↑LPIPS ↓
AnyAttack30.95900.89710.2712
M-Attack34.24720.88990.1817
SSA-CWA34.67740.90790.2457
ReasonBreak32.70380.86000.3360
GeoShield34.17560.88320.1909
NOWHERE32.31360.83020.2625

NOWHERE provides stronger location protection but has lower image fidelity than GeoShield. Compared with ReasonBreak, it has slightly lower PSNR and SSIM but better LPIPS. The added discussion explicitly presents these quality differences alongside the protection gains.

(2) Human evaluation. To assess how metric differences affect perceived quality, five raters evaluated 100 images per method, giving 500 responses per method. Method names were hidden and image order randomized. Raters scored visual quality from 1 to 5 and willingness to share based on appearance. We report results for three geolocation-privacy methods below.

MethodMean visual-quality scoreSharing acceptance
ReasonBreak3.6677.2%
GeoShield4.4290.2%
NOWHERE4.4092.2%

NOWHERE and GeoShield received close mean quality scores (4.40 vs. 4.42), with sharing acceptance rates of 92.2% and 90.2%. These results support the visual acceptability of NOWHERE within this sample and complement the objective metrics.

Thank you for your response. It solved my concern. I choose to maintain my positive score.

AuthorsOpenReview ↗

Response to Reviewer zp42

Thank you for following up and for your positive assessment. We appreciate your constructive feedback, which helped us improve the clarity and completeness of the paper.

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

The paper studies geolocation privacy: multimodal reasoning models can infer where a photo was taken from small visual clues. It proposes NOWHERE, a transfer-based adversarial perturbation method. It first locates clue regions in the image (using a VLM plus segmentation) and builds a low-rank "evidence subspace" in several CLIP encoders. During optimization, it tracks which clues are still insufficiently suppressed and shifts local crops toward them. Experiments on Google Street View and Im2GPS3k against four closed models show substantially lower geolocation accuracy than prior methods, with comparable caption-based semantic similarity.

Critical strengths
  1. The empirical gains are large and consistent. Across four black-box models and two datasets, NOWHERE beats GeoShield and ReasonBreak by a clear margin. Given how strong current VLMs are at geolocation, this is a meaningful practical result.

  2. The ablations are reasonably designed. The comparisons of mean vs. subspace targets, random vs. clue-guided crops, and shuffled priorities make it plausible that each component contributes.

Critical weaknesses
  1. The threat model is too weak for a privacy defense. Once a photo is posted, it cannot change, while the attacker can wait for newer models or simply denoise or purify the image first. This is the known weakness of the cloaking line of work (Fawkes and its successors). The paper only tests JPEG compression and blur. Without diffusion-based purification, resizing, or other adaptive attacks, the protection claim is hard to trust.

  2. Image quality is not really measured. There is no PSNR, SSIM, or LPIPS, and caption similarity is not better than the baselines. Since the whole point is "protect location but keep the photo shareable," this part needs to be stronger. The method is also mostly a combination of existing pieces (SVD subspace, DINO-style multi-crop, M-Attack-V2 aggregation), so the conceptual contribution is limited.

Questions for authors
  1. How does protection hold up under diffusion purification (e.g., DiffPure) or other denoising? This would most affect my score.

  2. Can you report standard image-quality metrics (PSNR, SSIM, LPIPS) against the baselines?

  3. GPT-5.5 is used to find clues, and two of the four targets are GPT models. Does performance change if the clue finder is an open model from a different family?

  4. What is the per-image cost (API calls, GPU time)?

4 comments
AuthorsOpenReview ↗

Response to Reviewer WRNw — Part 1 (WK1 & Q1)

Dear Reviewer,

Thank you for your careful review and helpful suggestions.

WK1 & Q1. How does protection hold up under diffusion purification or other denoising?

A1. We tested resizing, NLM (h=10), and DiffPure (t=100) on Im2GPS3k, applying the same processing to fixed protected images and unprotected images before geolocation. Table R1 reports Acc@25 km (%) on GPT-5.6-Terra; lower values indicate stronger protection.

Table R1. Protection under post-processing on Im2GPS3k.

MethodNo post-processingResize 256Resize 1280NLM 10DiffPure (t=100)
Clean55.348.056.442.245.9
M-Attack37.531.442.736.235.4
ReasonBreak27.926.831.129.127.9
GeoShield15.56.218.622.518.5
NOWHERE9.04.511.517.314.8

Denoising and purification weaken NOWHERE’s protection. NLM and DiffPure raise geolocation accuracy on its protected images from 9.0% to 17.3% and 14.8%, respectively, compared with 42.2% and 45.9% for unprotected images. Still, NOWHERE provides the strongest location protection among the compared methods in all five settings.

We have added these results to Appendix E, Table 7, on page 23, lines 1223–1237.

AuthorsOpenReview ↗

Response to Reviewer WRNw — Part 2 (WK2 & Q2)

WK2 & Q2. Can you report standard image-quality metrics (PSNR, SSIM, LPIPS) against the baselines? The method is also mostly a combination of existing pieces, so the conceptual contribution is limited.

A2. We address image quality and the method contribution below.

(1) Image quality and shareability. We have moved and updated the visual-quality discussion from Appendix C to Section 5, “Limitations and Discussion,” in the revised manuscript. Table 5 now appears on page 10, alongside the new human evaluation results.

Table R2. Image quality and protection on Im2GPS3k. Results from Tables 1 and 5 of the revised manuscript.

MethodPSNR (dB) ↑SSIM ↑LPIPS ↓25 km (%)
ReasonBreak32.70380.86000.336027.9
GeoShield34.17560.88320.190915.5
NOWHERE32.31360.83020.26259.0

The image-quality metrics show only small differences between NOWHERE, GeoShield, and ReasonBreak, with broadly comparable results. Meanwhile, NOWHERE provides stronger location protection, as shown in Table 1 on page 7 of the paper.

To assess whether these metric differences affect perceived quality, five raters evaluated 100 images per method in a blind study with randomized order. They rated visual quality from 1 to 5 and willingness to share based on appearance. We report results for ReasonBreak, GeoShield, and NOWHERE.

Table R3. Human evaluation of protected images.

MethodMean visual qualitySharing acceptance
ReasonBreak3.6677.2%
GeoShield4.4290.2%
NOWHERE4.4092.2%

NOWHERE and GeoShield received close mean quality scores (4.40 vs. 4.42). Their sharing acceptance rates were 92.2% and 90.2%, respectively. These results support the visual usability of NOWHERE within this sample, despite the differences in objective metrics.

We added this discussion to Section 5, page 9, lines 481–485, and the human study to page 10, lines 504–510. Tables 5 and 6 report the results.

(2) Method contribution. We drew inspiration from DINO and M-Attack-V2 solely for the multi-crop construction. Related sampling strategies also appear in iBOT, GeoShield, and SU. These examples show that existing sampling tools can support new method designs.

In NOWHERE, the contribution lies in ESC’s representation of geographic evidence and DET’s evidence-guided allocation of local views during optimization. The ablations in Table 3 of the paper support the value of these designs. Table 4(b) further isolates the allocation strategy with a fixed view count. Dynamic clue-guided allocation reduces accuracy from 11.0% to 9.0% on GPT-5.6-Terra.

Table R4. ESC and DET ablations. Acc@25 km (%) from Table 3 on page 8; lower values indicate stronger protection.

VariantGPT-5.6-TerraClaude-Opus-4.8
w/o ESC14.316.0
w/o DET (Global-only)13.113.5
NOWHERE9.09.8
AuthorsOpenReview ↗

Response to Reviewer WRNw — Part 3 (Q3 & Q4)

Q3. Does performance change if the clue finder is an open model from a different family?

A3. We replaced GPT-5.5 with Qwen3.8-27B to extract clues. Qwen is an open model from a different family and runs locally. We tested the protected images on Im2GPS3k using GPT-5.6-Terra.

Table R5. Effect of the clue finder on Im2GPS3k.

Clue finder1 km25 km200 km750 km2500 km
Clean19.755.367.880.290.9
GPT-5.54.29.010.817.130.5
Qwen3.8-27B4.510.111.217.134.6

Both versions greatly reduce geolocation accuracy compared with clean images. Their results are close at 1, 25, 200, and 750 km. Protection is weaker with Qwen at 2500 km. However, NOWHERE with Qwen still provides better protection than all compared baselines at all five distances.

Q4. What is the per-image cost (API calls, GPU time)?

A4. NOWHERE uses one VLM API call per image to identify geographic clues. GeoShield requires two such calls. The two methods have roughly comparable GPU runtimes, but NOWHERE provides better protection with fewer API calls.

Table R6. Per-image cost.

MethodAPI calls / imageGPU time (s / image)
GeoShield270.32
NOWHERE179.83
AuthorsOpenReview ↗

Follow-up on Our Responses

Dear Reviewer WRNw,

Your comments have been very valuable in helping us improve the quality of our paper. As the author-reviewer discussion period is set to end in less than one day, we would greatly appreciate it if you could take a moment to review our rebuttal and let us know whether our responses have addressed your concerns or if any issues remain.

Thank you again for your time and consideration.

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.