acceptodds
Under review as a conference paper at ICLR 2027

Source-Only Cross-View Disruption for Transferable Geolocation Privacy

Abstract

Image geolocation capabilities in proprietary vision-language models pose privacy risks by allowing ordinary photographs to reveal their capture locations. Existing protection methods often rely on constructed semantic targets, target images, or detector-defined regions, introducing additional assumptions and preprocessing dependencies. We introduce GL-SD, a source-only objective that disrupts source-image representations across stochastic global and local views using an ensemble of vision encoders, without requiring target images, surrogate text targets, or object detectors. A same-path controlled experiment on GPT-4.1 shows that cross-iteration stochastic view resampling improves mean privacy gain by 2,656.2 km over a fixed-view counterpart on 492 strictly paired images, with a 95% paired-bootstrap confidence interval of [2,179.7, 3,139.3] km. Component ablations further show that local source disruption is weak in isolation but provides a significant additional gain when coupled with the stochastic global objective. Under strictly paired evaluation on IM2GPS3K, GL-SD improves mean privacy gain over GeoShield by 1,111.9 km on GPT-4.1, 554.9 km on Claude-4.5, and 621.5 km on Gemini-2.5, with confidence intervals entirely above zero for all three models. Under matched optimization budgets, Feature-PGD remains substantially weaker than GeoShield. These results demonstrate effective source-only geolocation privacy protection across the evaluated black-box model families and identify stochastic view resampling as a contributor to transfer on GPT-4.1.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
geolocation privacy, vision-language models, multimodal models, transferable adversarial attacks, black-box evaluation

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_GwMWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GL-SD, a source-only cross-view representation suppression method for protecting image geolocation privacy against black-box vision-language models (VLMs). The paper claims that, unlike GeoShield, GL-SD does not require semantic targets or detector-defined regions, instead optimizing global and local feature disruption across multiple surrogate encoders. Experiments on IM2GPS3K show privacy protection on GPT-4.1, Claude-4.5, and Gemini-2.5, with additional ablation and robustness analyses.

Critical strengths
  1. The paper addresses an important privacy concern arising from the ability of VLMs to infer geographic locations from images, even without metadata. The proposed source-only approach may offer a simpler way to protect location privacy without relying on external semantic guidance or object detectors.

  2. The paper provides extensive evaluations across multiple black-box VLMs, supported by paired confidence intervals, component ablations, perturbation-budget analyses, and negative findings. The authors also acknowledge limitations and avoid overstating statistically insignificant improvements.

Critical weaknesses
  1. GeoShield already supports untargeted geolocation privacy protection, while global–local optimization, feature disruption, and stochastic view augmentation have been explored in prior studies. The proposed method mainly combines and simplifies existing techniques, and its independent technical contribution remains unclear.

  2. The experiments do not adequately distinguish the benefits of the proposed source-only objective from those of existing transfer-enhancement techniques. For example, stochastic global-view optimization alone substantially outperforms fixed-view optimization, while GL-SD requires approximately twice as many surrogate forward passes as GeoShield. Stronger input-diversity baselines and compute-matched comparisons are needed to establish whether the improvement comes from the proposed objective rather than existing augmentation strategies or additional computation.

  3. GL-SD directly disrupts general visual representations without explicitly preserving non-geographical information. Although PSNR, SSIM, and LPIPS are reported, these metrics cannot establish whether the protected images retain their original semantic content. Evaluations on image captioning, object recognition, or other downstream visual tasks are needed to assess the practical privacy–utility trade-off.

Questions for authors
  1. Given that GeoShield already supports untargeted geolocation protection and existing studies have explored global–local feature disruption, what is the key technical novelty of GL-SD beyond simplifying and combining existing techniques?

  2. GL-SD requires approximately twice as many surrogate forward passes as GeoShield. Does its performance advantage remain under matched computational budgets or GPU runtime?

  3. Since GL-SD directly disrupts general visual representations, how do the authors ensure that non-geographical semantic information is preserved? Can they provide evaluations on image captioning, object recognition, or other downstream tasks beyond PSNR, SSIM, and LPIPS?

  4. The reported mean PG values of F2 (5,618.4 km) and GL-SD (5,473.5 km) imply a difference of 144.9 km, rather than the reported paired improvement of 206.6 km. Could the authors clarify this discrepancy and confirm whether the comparison uses the same evaluation subset?

Reviewer_kRwuOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates methods of utilizing adversarial perturbations for geolocation privacy against black box vision language models. The privacy gain is measured as the distance in kilometers between the clean and adversarial predicted geolocation. The authors proposed method GL-SD uses three surrogate encoders and optimizes to jointly disrupt both the global and local image representations. Unlike similar prior work, i.e GeoShield, it does not use any targeted location labels or geographic knowledge, while achieving superior transfer to the IM2GPS3K evaluated on GPT-4.1, Claude-4.5, Gemini-2.5.

Critical strengths
  • The GL-SD is a simpler algorithm than GeoShield and geography agnostic, yet outperforms the geolocation specific method by a decent margin. Discovering which added complexity is unnecessary, is valuable for the field.
  • The evaluation on 3 separate API based black-box models is convincing gives me confidence in the robustness of the results.
Critical weaknesses

Weaknesses:

  • The main methodology and general contributions have modest novelty. It directly builds on top of GeoShield from Liu et al. (2026). Although to the authors' credit, they do not claim novelty for global-local optimization. Because of this I believe the overall contribution of this paper is incremental.
  • I do not find the paper to have sufficient explanations and reasoning of why GL-SD outperforms GeoShield, except the experiments that show that it does so on aggregate on IM2GPS3K.

Minor weaknesses:

  • I think figure 1 is a little messy, too dense, and could be improved. There are many overlapping lines with text and in general too much unnecessary information and text in it. It looks more like a whole research poster than a figure.
  • The abstract could be improved. It has too many technical details better suited to other places and reads more like a results section than an abstract
Questions for authors
  • Although GeoShield and GL-SD seem to use the same ensemble of surrogate models, they are not matched in terms of computation and queries to said models. Could the authors run an ablation to see if GeoShield gets improved performance if given the same budget as GL-SD?
  • Could the authors supply the attack effectiveness on the surrogate models (white-box) by e.g. reporting clean to adversarial feature similarity? This would help distinguish between stronger perturbations in general, or specifically stronger black-box transfer.
Reviewer_2MdaOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GL-SD, a source-only cross-view representation suppression method for transferable geolocation privacy. Unlike GeoShield, GL-SD does not require a target/decoy location, semantic geographic target, or detector-defined regions. It optimizes stochastic global and local views against cached clean representations using a three-CLIP surrogate ensemble.  On the strict 2,972-image GPT-4.1 evaluation, GL-SD achieves 5,473.5 km mean privacy gain vs. 4,361.7 km for GeoShield, a paired improvement of +1,111.9 km [921.3, 1,303.4]. The improvement also replicates Claude-4.5 and Gemini-2.5 holdouts.

Critical strengths
  1. GL-SD substantially outperforms the paper-exact GeoShield baseline on a large paired evaluation. 
  2. The advantage replicates on Claude-4.5 (+554.9 km) and Gemini-2.5 (+621.5 km), strengthening the transfer claim. 
  3. The use of strict common-image intersections, frozen clean predictions, paired statistics, confidence intervals, and clearly separated evaluation populations is convincing. 
  4. GL-SD improves over G-only by +1,181.6 km mean PG, and the Fixed-G/Stochastic-G comparison (+2,656.2 km) provides particularly interesting evidence for the importance of stochastic cross-view resampling. 
  5. The authors clearly distinguish empirical transfer from mechanistic explanations and report negative mechanism results rather than over-interpreting them.
Critical weaknesses
  1. Global/local optimization, feature-space attacks, transformation diversity, surrogate ensembles, and CLIP feature disruption have substantial precedent. The main novelty appears to be the source-only cross-view formulation and its application to geolocation privacy. The paper should more precisely articulate what conceptual contribution goes beyond combining established ingredients.
  2. The GL-SD vs. GeoShield comparison does not fully isolate the source of improvement. Multiple factors change simultaneously, including source-only representation suppression, semantic/target conditioning, detector-based regions, and stochastic view construction. A controlled hybrid/factorial ablation swapping these components would considerably strengthen the central claim.
  3. The improvement is heterogeneous. Although mean GL-SD−GeoShield improvement is +1,111.9 km, the median is only +146.6 km, with W/D/L = 1,649/211/1,112. A characterization of which images benefit most would help explain the result. 
  4. Perceptual quality remains incompletely evaluated. GL-SD has slightly worse PSNR/LPIPS than GeoShield, and the paper appropriately notes that these metrics do not establish human imperceptibility. A small human perceptual evaluation would strengthen the practical privacy-defense claim.
Questions for authors
  1. What is the precise methodological novelty of GL-SD? Is it primarily the unpaired source-only objective, stochastic cross-view optimization, their combination, or the empirical finding that this formulation transfers strongly for geolocation?
  2. Can the authors provide a controlled experiment that independently swaps source-only vs. target/semantic guidance and random local views vs. detector-selected regions?
  3. Why does the local term produce a large mean improvement over G-only (+1,181.6 km) but essentially zero median improvement (+0.5 km)? Can the authors characterize the subset producing the large positive tail? 
  4. How stable are the reported PG differences across repeated black-box queries? A more controlled isolation of why GL-SD outperforms GeoShield would make the contribution substantially stronger.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.