acceptodds
Under review as a conference paper at ICLR 2027

Geographic Flow Attacks

Abstract

Modern image-geolocation models can recover location directly from visual content, turning images into latent geographic metadata and creating new privacy risks. We study how to protect images against generative geolocation models using imperceptible adversarial perturbations. We introduce Geographic Flow Attacks (GeoFA), a unified framework that organizes attacks by where they act, what they deviate from or towards, and how deviation is measured. This reveals a mismatch in training-loss attacks: inference follows the model’s predicted velocity field, not the training target. We therefore propose Geographic Flow Attack by Trajectory Deviation (GeoFA-TD), which maximizes angular deviation between clean and perturbed velocities to redirect the sampling trajectory without backpropagating through it. Across YFCC4K and OSV-5M, GeoFA attacks outperform the state of the art. GeoFA-TD outperforms all attacks on OSV-5M and transfers from spherical RFM to diffusion. Code and models will be released.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
geoprivacy, adversarial, flow matching

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GCLwOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper deals with generative geolocation models, which produce a geo location for a given image. Adding adversarial perturbations to the input images can attack the underlying network and produce a different outcome. The paper shows different ways to add these perturbations, demonstrated on Riemannian Flow Models, split between different optimization targets such as the encoder or directly on the flow model itself. The paper shows superior results compared to GeoShield, which demonstrates these perturbations on VLMs.

Critical strengths
  • The decomposition of the L2L^2 difference to magnitude and direction is clear and important, and helps understand the importance of for optimizing on the direction alone, as the paper proposes.
Critical weaknesses
  • The paper collects previous methods that were not proposed specifically for RFM, and adjusts these to operate on such models for a geographic setting. This collection is presented as a “unified framework …”, when this is merely the adaptation of known methods to this architecture. Moreover, throughout the paper (e.g., Tab. 2) these methods are shown as “Ours-GeoFA…”, which is misleading and lets the reader think that these adaptations are different variants of the authors method. As a result, Tab.2 seems as an ablation for different methods, rather than presenting the actual method’s result. This makes the paper hard to follow, as it is hard for the reader to understand the last two rows of the “Ours” block are the authors actual method.
  • Apart from the readability, I think this points on to a problematic comparison to other known methods. GeoShield is a black box model that achieves far inferior results to all the white-box adapted settings. But, these adapted settings even show better or on par results to the paper's method (e.g., Sampling on YFCC4k). Currently, it seems that the paper have implemented known attack methods on the geographic RFM checkpoints of Dufour et. al., and the new suggested method does not significantly improve the results compared to the known methods. Therefore, I think at the current state the paper does not include enough novelty for publication, but I do think that the formulation of adversarial perturbation on RFM model in general can be important and useful, possibly with the paper’s supplied loss term, but demonstrated not only on a narrow benchmark of Geographic attacks on a single model, but on other types of RFM models as well. The current white-box setting, demonstrated on one specific approach for geographic flow model is not enough as it stands, together with the fact that it is not fully shown that it is actually better.

Minor comments: The appearance of the paper should be improved, there are at least a few instances where two words are missing a space in between (e.g., L85), the text itself can feel too repetitive (e.g., L75-59 and L207-210)

Questions for authors

-Can the GeoFA-TD loss be further analyzed to show its significance on other domains? Even toy examples would substantially help to visualize how this approach is tailored for RFM and why it is better than other perturbation methods (if that is even the case).

  • Apart from qualitative examples for the images differences as shown in Fig.2, can the author's supply also quantitative measures such as PSNR, LPIPS on the datasets? It is important to understand that the different optimizations, even when bounded with L∞L_{\infty}, resemble the original image.
3 comments
AuthorsOpenReview ↗

Response to GCLw (1/2)

Dear reviewer,

Thank you for your thorough review of our work. Below, we adress each your questions and pointed weaknesses

1. Significance of GeoFA-TD loss

We add a toy experiment and velocity deviation measurements of the real model attacked by GeoFA-TD and GeoFA-L^2 in Appendix E to allow a better understanding of the GeoFA-TD loss.

Toy model.

The toy model consists in a two mode riemannian flow matching model. We perturb velocities either by scaling ψ\psi by (1+s)(1+s) or by rotating it by θ\theta each step. We show that scaling the velocity essentially accelerates the speed of sampling but only poorly deviates the trajectory. At equal L2L^2 cost per step s=∥ψ~−ψ∥/∥ψ∥s=\lVert\tilde\psi-\psi\rVert/\lVert\psi\rVert, the table below gives the mean shift of the endpoint over 2,000 uniform starts. It considers both a time independent and a time conditioned velocity field so as to better compare to our geographic RFM:

Table 1: results of scale/angular perturbations on our RFM toy model

Direction fieldPerturbations=0.5s=0.5 (29°)s=1s=1 (60°)s=1.5s=1.5 (97°)
fixed in timescaling ×(1+s)100
rotation2262,12910,547
sharpening over timescaling ×(1+s)1,3411,7641,764
rotation1,7594,30510,695

Scaling the velocity has no effect under a time independent velocity field, and a weak effect on the time dependent one compared to rotations.

Real model measures (YFCC4k, 100 images).

We test pertrubating RFM S2\mathbb{S}^2's clean velocity on YFCC4k images via scale and angular perturbations. Here again, rotating velocities has a much stronger effect than scaling it, which is coherent with our toy experiment.

Table 2: results of scale/angular perturbations of the real model's velocities on 100 YFCC4k images

ss0.250.511.5
scale62 km105 km182 km192 km
rotation93 km226 km548 km11,447 km

These perturbation strategies may seem artifical, since we add a fixed invariant perturbation at each step. Especially for rotations, one might argue that rotations could compensate each other if they do not have the same sense over many steps. We therefore measure the effect of GeoFA-TD and GeoFA-L^2 on sampling dynamics on 100 images in Table 3, which compares attacked and clean velocities ψδ=ψ(xk∣c+δ)\psi_\delta=\psi(x_k\mid c+\delta), ψc=ψ(xk∣c)\psi_c=\psi(x_k\mid c) at each point xkx_k (θk\theta_k is the angle between the two guided velocities)::

Table 3: Comparison of clean and attacked velocities for GeoFA-TD and GeoFA-L^2 on 100 YFCC4k images.

Attackbbmean ∣θk∣\lvert\theta_k\rvert (first 100 of 250 steps)share of these steps with ∣θk∣>90∘\lvert\theta_k\rvert>90^\circmedian ∥ψδ∥/∥ψc∥\lVert\psi_\delta\rVert/\lVert\psi_c\rVert (before guidance)shift of the prediction
GeoFA-L2L^21/25590°53%1.239,428 km
GeoFA-TD1/255110°69%1.1513,985 km
GeoFA-L2L^28/25591°53%1.489,849 km
GeoFA-TD8/255114°72%1.2014,609 km

GeoFA-L2L^2 inflates the velocity more (ratio 1.48 vs. 1.20 at 8/255). GeoFA-TD deviates its angle more and its predictions move further. We also check that the GeoFA-TD's rotations keep the same sense over steps. This is the case of 97% of steps, meaning that the deviations add up along sampling. This justifies using same sense angular rotations in our toy example.

GeoFA-TD compared to other methods

We stress that GeoFA-TD:

  • is cheaper to train than the sampling attack and is more scalable to bigger models as it does not require backpropagating through the whole sampling process. The Sample-G is the only attack surpassing GeoFA-TD consistently on YFCC4k.
  • is more transferable to other an architectures it was not trained on, as we show in Figure 5 by evaluating its success rate against a geographic diffusion model sharing the same encoder as the RFM is was trained on.
2. Quantitative measures of image quality

We measure PSNR and LPIPS for GeoFA-TD and GeoShield on YFCC4k as in [1]. GeoFA-TD consistently outperforms GeoShield in PSNR and LPIPs, which is coherent with our qualitative observations. Only at low budget 1/255, the two attacks are close. This is the regime in which GeoShield almost completely fails (see Table 2 of the udpated PDF), while GeoFA-TD remains strong.

Table 4: Quantitative image quality comparison of GeoShield and GeoFA-TD on YFCC4k.

BudgetAttackPSNR ↑LPIPS ↓
1/255GeoShield49.95 ± 0.610.0025 ± 0.0032
1/255GeoFA-TD50.25 ± 2.830.0021 ± 0.0021
4/255GeoShield39.41 ± 0.730.042 ± 0.033
4/255GeoFA-TD41.27 ± 2.140.021 ± 0.022
8/255GeoShield34.03 ± 0.760.117 ± 0.068
8/255GeoFA-TD40.11 ± 2.310.026 ± 0.025
AuthorsOpenReview ↗

Response to GCLw (2/2)

3. Other weaknesses:
Black vs white box attacks' comparability

We have expanded our black box evaluation of GeoFA-TD in order to make the comparison with geoShield and RTA [1], a newly added baseline, fairer. Full results are given in point 2 of our answer to reviewer vEoU.

We highlight that black box evaluations of RTA [1] are limited to retrieval methods that use the same image encoder as the method the attack is trained on, e.g. CLIP (or a variation of it). This is close to our black box evaluation of GeoFA-TD in Figure 5, which evaluates GeoFA-TD against a geographic diffusion variant of RFM S2\mathbb{S}^2 (thus evaluating a generative architecture transfer, not an encoder transfer).

Novelty compared to known attacks adapted from the literature

We aknowledge that naming the attacks we adapted from the literature with a "GeoFA" prefix can cause confusion. We deliberately kept the initial names of the attacks (ex: GeoFA-DPA, now DPA-G, comes from DPA=Diffusion Policy Attack), this was not meant to confuse the reader in thinking these attacks were fully novel. One of our work's contribution is organizing prior work in terms of where and what it attacks, and adapting it to geographic RFM. To the best of our knowledge, prior work on adversarial attacks against a diffusion model on another modality than images does not do that [2, 3].

For clarity purposes, we changed the naming of attacks adapted from the literature to [Attack name]-G to show they were adapted to our setting. We keep the GeoFA prefix for -TD and -L2L^2.

References:

[1] Niccolo Niccoli, Federico Becattini, and Lorenzo Seidenari. Defending from GeoLocalization through adversarial road trips. In Paolo Favaro, Zuzana Kukelova, Atsuto Maki, Anna Rohrbach, Konrad Schindler, and Federico Tombari (eds.), Computer Vision – ECCV 2026, volume 17054 of Lecture Notes in Computer Science, pp. 611–629, Cham, 2026.

[2] Yipu Chen, Haotian Xue, and Yongxin Chen. Diffusion policy attacker: Crafting adversarial attacks for diffusion-based policies. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang (eds.), Advances in Neural Information Processing Systems, volume 37, pp. 119614–119637. Curran Associates, Inc., 2024b.

[3] Erik Imgrund, Thorsten Eisenhofer, and Konrad Rieck. Adversarial observations in weather forecasting. In Proceedings of the 2025 ACM SIGSAC Conference on Computer and Communications Security (ACM CCS ’25), 2025. doi: 10.1145/3719027.3765076.

AuthorsOpenReview ↗

Summary of the rebuttal

We thank the reviewer again for your detailed and constructive review. We hope the additional results and analysis address your concerns. Specifically,

  • Clarity of contributions: attacks adapted from the literature are now named [Attack]-G (e.g. DPA-G). The GeoFA prefix is kept only for our two new objectives
  • Analysis of the TD loss (App. E): rotation moves predictions 3× more than scaling (548 vs. 182 km); TD rotates more than L² (114° vs. 91°), in the same direction 97% of the time, and shifts predictions further (14,609 vs. 9,849 km).
  • Image quality: GeoFA-TD beats GeoShield at every budget:. PSNR 40.1 vs. 34.0 and LPIPS 0.026 vs. 0.117 at 8/255 (Table 12, App. F.2.5).
  • Black-box: GeoFA-TD with no model access stays >99% (OSV-5M) / >91% (YFCC4k), above GeoShield and RTA at every budget. Transfer across encoders is limited for all methods and now stated explicitly.
  • Advantage over adapted attacks: GeoFA-TD is cheaper than Sampling-G (Table 3), works in the black box setting (Table 2) and transfers across generative paradigm (Figure 5)

We hope these additions address your concerns. If they do, we would be grateful if you would consider updating your assessment. We are happy to clarify anything further.

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

This paper investigates adversarial perturbations for protecting image geolocation privacy against generative geolocation models, particularly Riemannian Flow Matching (RFM) models operating on spherical coordinates. The authors introduce Geographic Flow Attacks (GeoFA), a framework that categorizes attack objectives according to the component being attacked, the reference or target used to define deviation, and the metric used to measure it.

Critical strengths
  1. The paper makes a useful distinction between the velocity targets used during training and the predicted velocity field actually integrated during inference.

  2. The paper evaluates several attack objectives under common perturbation budgets and reports both effectiveness and computational cost.

Critical weaknesses
  1. The primary success metric does not fully distinguish attack-induced privacy protection from pre-existing geolocation errors. The main ASR metric measures whether the attacked prediction is more than 750 or 2,500 km from the ground-truth location. However, a substantial fraction of clean images are already incorrectly localized, particularly on YFCC4K. Table 2 reports a clean ASR above 2,500 km of 25.73% on YFCC4K. These cases already satisfy the attack-success criterion without requiring any perturbation.

  2. The headline ASR still combines images that were initially vulnerable to geolocation with images that were already geographically ambiguous or incorrectly predicted

  3. The experiments primarily consider white-box access to the target geolocation model, including its encoder and denoiser. More importantly, the attack. construction uses the image's true geographic coordinates to sample intermediate states along the forward process. This is reasonable for an image owner who knows the original location, but it is a significant assumption.

Questions for authors
  1. The paper includes clean-versus-attacked FSD analyses, but could the authors report ASR and geographic displacement on subsets with clean localization errors below 25 km, 200 km, and 750 km?
3 comments
AuthorsOpenReview ↗

Response to vEoU (1/2)

Dear reviewer,

Thank you for your thorough review of our work. Below, we adress each your questions and pointed out weaknesses

1. Evaluation on correctly localized images

We evaluated our methods and baselines on subsets of well localized images as you suggested. Results are available in the updated PDF (Table 2 of the updated pdf; ASR on all images and single-prediction ASR in Tables 9-11, Appendix F) and in our response 2A to reviewer WdCu. Results show that on these subsets, the general ordering of our attacks remains very similar to our initial results.

2. Other weaknesses: favorable assumptions of GeoFA

GeoFA indeed suppose access to the true location and white box access to the model.

True location access.

By default, the -G attacks and GeoFA-TD use the image's true location: Sampling-G maximizes the distance to it, DPA-G targets its antipode, and GeoFA-TD uses it to draw the noised positions xtx_t. GeoFA-TD is the only one of the three whose objective does not involve the true location (it only needs it for xtx_t). We therefore ran all three attacks without the true location. GeoFA-TD draws xtx_t from the clean model's own samples. Sampling-G and DPA-G use the clean model's prediction in place of the true location.

Table 1: ASR$_{>2500}$ (%, ↑\uparrow, all images, one attempt per image) with (✓) and without (✗) the true location.

DatasetAttackTrue loc.?b=1/255b=4/255b=8/255largest drop (both thresholds)
OSV-5MSampling-G✓ / ✗97.67 / 97.3599.85 / 99.5899.88 / 99.520.36
OSV-5MDPA-G✓ / ✗96.47 / 95.7899.98 / 99.6599.98 / 99.650.69
OSV-5MGeoFA-TD✓ / ✗99.67 / 99.5099.92 / 99.7299.95 / 99.720.23
YFCC4kSampling-G✓ / ✗95.93 / 94.3099.50 / 98.0299.45 / 98.051.63
YFCC4kDPA-G✓ / ✗93.92 / 91.8598.85 / 97.8298.85 / 98.282.07
YFCC4kGeoFA-TD✓ / ✗93.70 / 93.5897.35 / 96.9097.15 / 97.280.45

All three attacks remain strong without the true location, so this assumption is not what drives their success. GeoFA-TD is the least affected: it loses at most 0.45 points, against 1.6 for Sampling-G and 2.1 for DPA-G. This matches the fact that its objective never references the true location. Without the true location, GeoFA-TD is the best attack on OSV-5M and is on par with Sampling-G and DPA-G on YFCC4k (see also Table 2 of the updated PDF).

White box access.

We aknowledge that GeoShield's black box access makes its comparison with GeoFA methods unfavorable. We mainly include GeoShield because it is the only existing baseline (with RTA against retrieval models [1], which we add to our baselines).

For this rebuttal, we evaluate GeoFA-TD with neither the true location nor access to the evaluated model, in two transfer settings (Table 2 in the next comment; Table 2 of the updated PDF):

  • GeoFA-TD (diff. transfer): the perturbation is computed against RFM S2\mathbb{S}^2 and evaluated against its diffusion variants by Dufour et al. [2] that shares its encoder. It transfers almost fully: ASR$_{>2500}$ stays above 99% on OSV-5M and above 91% on YFCC4k at every budget, above GeoShield and RTA at every budget and close to the white box setting. We highlight that on OSV-5M, GeoShield and RTA have access to the encoder since they train on CLIP variants, making them comparable with GeoFA-TD (diff.transfer). (The recent work RTA [1] similarly evaluates black box performance against geolocation methods that share the same encoder it attacks, e.g. CLIP variants.)
  • GeoFA-TD (enc. transfer): the perturbation is computed against the RFM S2\mathbb{S}^2 model of the other dataset (different encoder and training data: DinoV2 vs. StreetCLIP) and evaluated against the dataset's own model.

GeoFA-TD therefore protects against geographic generative models sharing the encoder it is computed on, whatever their generative framework (RFM or diffusion), but its transfer across image encoders is limited.

AuthorsOpenReview ↗

Response to vEoU (2/2)

Table 2: ASR (%, ↑\uparrow, all images) of GeoFA-TD with and without access to the true location and to the evaluated model, compared with the black box baselines.

DatasetAttackTrue loc.?White box?b=1/255 >750b=1/255 >2500b=4/255 >750b=4/255 >2500b=8/255 >750b=8/255 >2500
OSVGeoFA-TD✓✓99.9899.67100.0099.92100.0099.95
OSVGeoFA-TD✗✓99.9299.5099.9099.7299.9099.72
OSVGeoFA-TD (enc. transfer)✗✗40.3322.0581.7564.3385.3569.23
OSVGeoFA-TD (diff. transfer)✗✗99.8399.2099.8599.5299.9099.55
OSVGeoShield✗✗55.4334.8898.4792.4299.7596.78
OSVGeoShield + VLM✗✗58.4837.8298.6293.5899.7297.28
OSVRTA✓✗41.7721.5775.5252.0585.2566.45
OSVClean24.1810.20
YFCC4KGeoFA-TD✓✓96.5593.7098.8597.3598.6797.15
YFCC4KGeoFA-TD✗✓96.2893.5898.6296.9098.8897.28
YFCC4KGeoFA-TD (enc. transfer)✗✗43.5526.6747.7730.7348.9832.42
YFCC4KGeoFA-TD (diff. transfer)✗✗95.7391.4098.9595.6799.1295.93
YFCC4KGeoShield✗✗44.1027.8265.1548.0881.5366.07
YFCC4KGeoShield + VLM✗✗43.6026.9263.6047.0279.6264.65
YFCC4KRTA✓✗42.9826.3047.0530.3853.0535.55
YFCC4KClean41.9825.72

References:

[1] Niccolo Niccoli, Federico Becattini, and Lorenzo Seidenari. Defending from GeoLocalization through adversarial road trips. In Paolo Favaro, Zuzana Kukelova, Atsuto Maki, Anna Rohrbach, Konrad Schindler, and Federico Tombari (eds.), Computer Vision – ECCV 2026, volume 17054 of Lecture Notes in Computer Science, pp. 611–629, Cham, 2026.

[2] Nicolas Dufour, Vicky Kalogeiton, David Picard, and Loic Landrieu. Around the world in 80 timesteps: A generative approach to global visual geolocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 23016–23026, June 2025.

AuthorsOpenReview ↗

Summary of the rebuttal

We thank the reviewer again for your detailed and constructive review. We hope the additional results and analysis address your concerns. Specifically,

  • Results in well-localized subsets (<25/200/750 km): The attack ordering is unchanged (Table 2 and Tables 9–11, App. F). GeoFA-TD reaches 99.06% on OSV-5M at 1/255 (<25 km).
  • Use of the true location: without the true location, GeoFA-TD loses ≤0.45 pts (vs. 1.6 for Sampling-G, 2.1 for DPA-G, Table 2)
  • Black-box: GeoFA-TD with no model access stays >99% (OSV-5M) / >91% (YFCC4k), above GeoShield and RTA at every budget. Transfer across encoders is limited for all methods.

We hope these results address your concerns. If they do, we would be grateful if you would consider updating your assessment. We are happy to answer any further questions.

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

The paper studies how to protect images with imperceptible L∞ perturbations so that a generative geolocation model, Riemannian Flow Matching on the sphere (RFM-S²), cannot recover where they were taken. The GeoFA framework organizes previous attacks on diffusion and flow matching models by three questions: where the attack acts, what it deviates from, and how deviation is measured. The key insight is that training-loss attacks push the model away from the training target v, while at inference the sampler follows the model's own prediction ψ. This motivates two new objectives: GeoFA-L2, which moves away from the clean prediction, and GeoFA-TD, which pushes the clean and perturbed velocity directions apart without backpropagating through the sampler. Experiments on YFCC4k and OSV-5M show that the GeoFA attacks outperform the strongest existing defense and reach ~100% success even at the smallest perturbation budget, and that GeoFA-TD also transfers to diffusion models sharing the same encoder.

Critical strengths
  1. Well-chosen problem, a real gap. Existing geo-privacy work (GeoShield, BPPM, RTA) targets VLMs or retrieval models, so the threat model here is new. The attack is also simple to set up: L∞ with PGD, one denoiser call, and no backpropagation through the 250-step sampler.
  2. The central insight is sound. The sampler follows the model's own prediction, not the training target, so attacking the prediction directly is both cheap and effective at small budgets. At b=1/255, GeoFA-TD reaches 99.97/99.60 on OSV-5M and 96.45/94.33 on YFCC4k, at 193 ms and 1.93 GiB per step. Fig. 7 is the most convincing evidence in the paper: TD pushes away exactly those images the model localizes well (r = -0.42), which is also where privacy protection matters most.
Critical weaknesses
  1. The GeoShield comparison is unfair. Appendix D says GeoShield was tested without region detection and with a constant text term. GeoFA can access the target model, but GeoShield cannot. This cannot support the claim of beating the state of the art.
  2. TD is not consistently better. In Table 4, with five restarts on YFCC4k at 8/255, DPA scores 100.00/100.00, while TD scores 99.98/99.92. DPA also wins at 4/255. TD's advantage clearly depends on the setting. A random-noise baseline is also missing.
  3. A high ASR does not mean the location is hidden. The paper averages the errors of 128 predictions before judging success. Some predictions may still be correct. Also, 42% of clean YFCC4k images already have errors above 750 km. Report how often at least one prediction finds the location, especially for images the clean model gets right.
  4. The paper does not explain why TD works. Equation 13 uses ψ, but sampling uses ψ_w with CFG = 10. Does the attack target the velocity used at test time? Proposition 1 and Equation 12 are algebra, not evidence that changing velocity direction gives larger location errors.
  5. The new method adds little. Six of the eight attacks come from earlier work. The two new losses mainly change what is compared and how. Relevant work, including AdvFM and RTA, is missing.
  6. Practical use is unclear. The attacks need access to the target model. Transfer is tested only between models with the same encoder. There is no evidence that the method protects photos against an unknown model.
Questions for authors
  1. Is the ψ in Eq. 13 the guided field ψ_w? If not, why is this not the same inconsistency the paper criticizes? If the objective uses ψ_w at w=10, does TD's advantage over L2 and DPA survive?
  2. Please report: (a) on the subset the clean model localizes correctly, the protection rate of a single prediction and the fraction of images with at least one of 128 draws within d km (thresholds, subset size, noise pairing); (b) the ASR of random noise at the same budget and of single-step encoder PGD; (c) the extra cost of best-of-five selection over a single initialization.
  3. Table 4 shows that DPA matches or beats TD with five restarts. On OSV-5M, several attacks reach 100%. What evidence supports TD being better than these existing methods?
  4. Appendix D states that GeoShield was tested without detector regions and with placeholder descriptions. How can this comparison support the SOTA claim? Can the authors evaluate the full method or revise their claims?
  5. Eq. 12 separates changes in velocity magnitude and direction, but does not show which matters more for the final prediction. Can the authors measure both effects and show that direction changes lead to larger location errors?
4 comments
AuthorsOpenReview ↗

Response to WdCu (1/3)

Dear reviewer,

Thank you for your thorough reviewof our work. Below, we adress each of your questions:

1. Attacked field

You ask whether ψ\psi in Eq. 13 is the guided field. It is not, we use the plain field ψ\psi (as in all other attacks). This indeed induces a small inconsistency between attack training and sampling. It however allows to train an attack that is not specifically tied to sampling parameters (in that case w=10w=10, we ablate the sampling guidance scale in Figure 6). For the sampling attack, we also test training it with w=10. This significantly degrades attack success, as is shown in Figures 6 and 9 (Appendix F.1.2).

2. Additional evaluations
A. Protection rate of a single prediction

We report the protection rate for single prediction by the attacked model (Tables 1 and 2). Overall numbers are nearly identical to that of the 128 prediction batch, the ordering of the attacks stays very similar to our initial table 2: Sample-G, DPA-G and GeoFA-TD lead. GeoFA-TD still scores best on OSV-5M.

We also report in Tables 3,4 the share of images for which all 128 predictions are pushed more than 2500 km away. This variant does not significantly change results.

Table 1: Single-prediction ASR$_{>2500}$ (%, ↑\uparrow) on well-localized predictions, OSV-5M. Each of the 128 (clean, attacked) prediction pairs per image is a trial, kept if its clean prediction is within 25 / 200 / 750 km of the truth. Kept trials (in image-equivalents, /128): <25 km: 425, <200 km: 2163, <750 km: 3036.

Attackb=1/255 <25 kmb=1/255 <200 kmb=1/255 <750 kmb=4/255 <25 kmb=4/255 <200 kmb=4/255 <750 kmb=8/255 <25 kmb=8/255 <200 kmb=8/255 <750 km
GeoShield (2026)23.2326.4727.8491.1091.0691.7696.4996.2396.30
Encoder-G (2023)69.5077.4978.7796.8797.4197.7797.3198.0198.33
Sampling-G (2023)95.9796.66<u>97.0299.5399.7799.7799.4799.7899.84
AdvDM-G (2023)72.9476.7076.9878.3582.0682.2878.8282.2082.71
ACE-G (2025)75.0880.8681.4878.3884.9685.6779.7285.1485.56
DPA-G (2024b)93.6094.4594.9599.7699.8999.90100.0099.9399.86
UniDef-G (2026)68.0767.9166.4781.3579.0578.1579.7478.9177.59
GeoFA-L²90.1093.0693.2192.4794.3694.4093.6594.9195.03
GeoFA-TD99.0699.5899.6499.7699.9199.93100.00100.00100.00

Table 2: Single-prediction ASR$_{>2500}$ (%, ↑\uparrow) on well-localized predictions, YFCC4k. Each of the 128 (clean, attacked) prediction pairs per image is a trial, kept if its clean prediction is within 25 / 200 / 750 km of the truth. Kept trials (in image-equivalents, /128): <25 km: 1260, <200 km: 1709, <750 km: 2325.

Attackb=1/255 <25 kmb=1/255 <200 kmb=1/255 <750 kmb=4/255 <25 kmb=4/255 <200 kmb=4/255 <750 kmb=8/255 <25 kmb=8/255 <200 kmb=8/255 <750 km
GeoShield (2026)2.963.974.4330.9232.0433.8058.0757.8458.96
Encoder-G (2023)69.4869.0669.4879.4979.9081.3681.5581.4881.96
Sampling-G (2023)93.2293.7694.5199.1399.3099.4099.3799.3699.32
AdvDM-G (2023)67.6166.7966.8975.9274.9074.9776.5375.3575.48
DPA-G (2024b)89.9090.5190.9198.4298.3498.3498.1998.1898.43
UniDef-G (2026)51.0050.2748.7163.1062.1661.3865.4163.7762.79
GeoFA-L²70.7872.7874.4475.9977.7779.2176.7278.4680.04
GeoFA-TD92.1392.4493.1696.8997.3697.7196.4396.9797.28
AuthorsOpenReview ↗

Response to WdCu (2/3)

Table 3: Share of well-localized images (%, ↑\uparrow) for which all 128 predictions are pushed more than 2500 km away, OSV-5M. Images whose clean error (mean over the 128 predictions) is below 25 / 200 / 750 km; same images as Table 11 of the paper at 200 km. Images: <25 km: 424, <200 km: 2160, <750 km: 3033.

Attackb=1/255 <25 kmb=1/255 <200 kmb=1/255 <750 kmb=4/255 <25 kmb=4/255 <200 kmb=4/255 <750 kmb=8/255 <25 kmb=8/255 <200 kmb=8/255 <750 km
GeoShield (2026)21.9325.4226.7890.5790.4691.1395.7595.4695.35
Encoder-G (2023)67.9276.2077.3293.1693.6694.3092.6994.1794.96
Sampling-G (2023)95.5296.1196.5799.5399.7799.7499.2999.7299.80
AdvDM-G (2023)72.8876.6276.9278.3082.0482.2678.7782.1882.69
ACE-G (2025)73.1179.7280.4578.0784.7785.4979.2584.8685.20
DPA-G (2024b)91.0492.3192.9899.7699.6899.70100.0099.8699.77
UniDef-G (2026)67.2266.9965.7479.2578.0177.2278.5477.9676.66
GeoFA-L²89.8693.0193.1892.4594.3594.3993.6394.9195.02
GeoFA-TD99.0699.5899.6499.7699.9199.93100.00100.00100.00

Table 4: Share of well-localized images (%, ↑\uparrow) for which all 128 predictions are pushed more than 2500 km away, YFCC4k. Images whose clean error (mean over the 128 predictions) is below 25 / 200 / 750 km; same images as Table 11 of the paper at 200 km. Images: <25 km: 1253, <200 km: 1698, <750 km: 2321.

Attackb=1/255 <25 kmb=1/255 <200 kmb=1/255 <750 kmb=4/255 <25 kmb=4/255 <200 kmb=4/255 <750 kmb=8/255 <25 kmb=8/255 <200 kmb=8/255 <750 km
GeoShield (2026)2.553.363.7529.9030.9432.7957.2656.9858.12
Encoder-G (2023)68.4568.1868.6378.6478.7980.2580.2880.2480.69
Sampling-G (2023)93.0693.6494.3698.9699.1299.2799.1299.1899.18
AdvDM-G (2023)67.4966.5566.8175.9274.7774.9176.4975.2075.44
DPA-G (2024b)88.0388.6389.1997.4597.5397.5497.5397.5997.80
UniDef-G (2026)49.8849.1247.7462.1761.1360.3664.8063.2562.04
GeoFA-L²70.7172.6774.4575.9877.7479.2376.7078.4580.01
GeoFA-TD92.0292.3493.1196.8197.2997.6796.4196.9497.24
B. Performance of random noise

We added a random noise baseline in our main Table 2. Random noise moves attack sucess rate by about 10 points on OSV, and is close to RTA but well below our methods. On YFCC4k, it has barely an effect - this is probably due to DinoV2's increased robustness to noise compared to StreetClip (OSV's encoder).

The encoder attack's results for single prediction are in our updated Table 2 in the pdf.

C. Extra cost of best of five selection over single initialization

The extra cost of best out of five selection is 5 times the compute time if runs are run sequentially, which we do. This brings the attack training cost of GeoFA-TD close to GeoShield, which allows an iso-time attack comparison.

3. Evidence of TD being better than other attacks

Using the metric you suggested in question 2, that is single prediction on well localized images, GeoFA-TD beats the second best attack by 2-3 points at budget 1/255 on OSV-5M, and is the only attack consistently reaching 100% success at higher budget. On YFCC4k, GeoFA-TD is within the three best attacks (DPA, Sample, GeoFA-TD), although it does not lead.

Moreover, we stress that GeoFA-TD

  • is cheaper to train than the sampling attack and is more scalable to bigger models as it does not require backpropagating through the whole sampling process. The Sample-G is the only attack surpassing GeoFA-TD consistently on YFCC4k.
  • is more transferable to other architectures it was not trained on, as we show in Figure 5 by evaluating its success rate against a geographic diffusion model sharing the same encoder as the RFM is was trained on.

We underline that using best-of 5 initialization results is useful for measuring attack reliability, but it is very expensive in practice if one aims at protecting many images. This is why we put forward single initialization results.

AuthorsOpenReview ↗

Response to WdCu (3/3)

4. GeoShield evaluations

GeoShield was evaluated using the default instantiation of the paper's authors (https://github.com/thinwayliu/Geoshield). This does not include detector regions nor VLM descriptions. Detection regions are based on a call to a third party repo, GroundingDINO, which is broken in the current code and reported as optional by the authors. VLM descriptions are also out of the default version, and require (to reproduce GeoShield's method exactly) access to a proprietary VLM (GPT-4o, GPT-4.1, Claude-3.5, and Gemini-2.5), making large scale evaluations difficult.

Additionally, we have added an open-source VLM, Gemma4-12b, to generate image descriptions used in GeoShield, and have re-evaluated this pipeline on both our datasets. Results are in the updated PDF under the the name GeoShield+VLM. The VLM component does not enhance GeoShield's performance across our two datasets.

5. Analysis of velocity/magnitude perturbations

We add an analysis of our GeoFA-TD attack's dynamics in Appendix E of the updated PDF.

We measure the effect of velocity and magnitude perturbations on sampler trajectories of RFM S2\mathbb{S}^2 on 100 YFCC4k images. Angular/magnitude pertubations amount to the same relative L2L^2 amount s=∥ψ~−ψ∥/∥ψ∥s=\lVert\tilde\psi-\psi\rVert/\lVert\psi\rVert. We either inflate it (ψ~=(1+s)ψ\tilde\psi=(1+s)\psi) or rotate it in the tangent plane by θ\theta, with 2sin⁡(θ/2)=s2\sin(\theta/2)=s. We then measure how far the prediction moves, and observe that magnitude only perturbations have a very weak effect compared to angular ones. This is due to the fact that increasing the magnitude of velocity essentially accelerates the sampling process rather than deviating it.

Table 5: Rotations vs inflation of velocity along RFM S2\mathbb{S}^2 sampling trajectory on 100 YFCC4K images.

ssinflationrotation
0.2562 km (×1.25)93 km (14°)
0.5105 km (×1.5)226 km (29°)
1182 km (×2)548 km (60°)
1.5192 km (×2.5)11,447 km (97°)

We also measure the effect of GeoFA-TD and GeoFA-L^2 on sampling dynamics on 100 images in Table 6, which compares attacked and clean velocities ψδ=ψ(xk∣c+δ)\psi_\delta=\psi(x_k\mid c+\delta), ψc=ψ(xk∣c)\psi_c=\psi(x_k\mid c) at each point xkx_k (θk\theta_k is the angle between the two guided velocities)::

Table 6: Comparison of clean and attacked velocities for GeoFA-TD and GeoFA-L^2 on 100 YFCC4k images.

Attackbbmean ∣θk∣\lvert\theta_k\rvert (first 100 of 250 steps)share of these steps with ∣θk∣>90∘\lvert\theta_k\rvert>90^\circmedian ∥ψδ∥/∥ψc∥\lVert\psi_\delta\rVert/\lVert\psi_c\rVert (before guidance)shift of the prediction
GeoFA-L2L^21/25590°53%1.239,428 km
GeoFA-TD1/255110°69%1.1513,985 km
GeoFA-L2L^28/25591°53%1.489,849 km
GeoFA-TD8/255114°72%1.2014,609 km

GeoFA-L2L^2 inflates the velocity more (ratio 1.48 vs. 1.20 at 8/255). GeoFA-TD deviates its angle more and its predictions move further. We also check that the GeoFA-TD's rotations keep the same sense over steps. This is the case of 97% of steps, meaning that the deviations add up along sampling. This justifies using same sense angular rotations in our toy example.

Other weaknesses

We had missed the very recent RTA work [1]. We now included RTA in our baselines (see table 2), and show it does not perform well against geographic RFM models compared to our methods and to GeoShield. It is best on OSV-5M, which can be attributed to the fact that RFM S2\mathbb{S}^2 uses a CLIP based encoder on this dataset (RTA attacks CLIP based retrieval methods). On YFCC4k, it is very close to the effect of random noise.

References:

[1] Niccolo Niccoli, Federico Becattini, and Lorenzo Seidenari. Defending from GeoLocalization through adversarial road trips. In Paolo Favaro, Zuzana Kukelova, Atsuto Maki, Anna Rohrbach, Konrad Schindler, and Federico Tombari (eds.), Computer Vision – ECCV 2026.

AuthorsOpenReview ↗

Summary of the rebuttal

We thank the reviewer again for your detailed and constructive review. We hope the additional results and analysis address your concerns. Specifically,

  • Single prediction, well-localized images: GeoFA-TD leads by 2–3 pts on OSV-5M at 1/255 and is the only attack reaching 100% at 8/255. It is top-3 on YFCC4k.
  • Random noise: We report it in Table 2. It is far below our attacks.
  • GeoShield: We use the official implementation. Adding a VLM (GeoShield+VLM) does not help.
  • Black-box: GeoFA-TD with no model access stays >99% (OSV-5M) / >91% (YFCC4k), above GeoShield and RTA at every budget (Table 2). Transfer across encoders is limited for all methods.
  • Why TD works (App. E): rotation moves predictions 3× more than scaling (548 vs. 182 km); TD rotates more than L² (114° vs. 91°), in the same direction 97% of the time, and shifts predictions further (14,609 vs. 9,849 km).
  • RTA is added as a baseline, and the adapted attacks are renamed [Attack]-G to separate them from our two new objectives.

We hope these additions address your concerns. If they do, we would be grateful if you would consider updating your assessment. We are happy to clarify anything further.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General comment

We thank the reviewers for their time spent on our work. Their questions led to several new experiments, all included in the updated PDF (modifications in blue):

  • Evaluation of a new Baseline on top of GeoShield, RTA, a very recent visual geoprivacy work (R-WdCu)
  • As proposed by R-WdCu, we now report single prediction results rather than averaged result over a 128 predictions batch in Table 2 (R-WdCu).
  • Results reported on the subsets of images that the attacked model localizes well in Table 2 as suggested by R-WdCu and R-vEoU.
  • Both white and black box experiments: we add an evaluation of GeoFA-TD in a black box setting, without the true location, by computing it on a different version of the geographic model it is evaluated against: GeoFA-TD (enc. transfer) changes the encoder and training data, GeoFA-TD (diff. transfer) changes the generative framework (RFM to diffusion). Results are reported in Table 2. We also remove its access to the true location of the image for fair comparison with GeoShield. Results are reported in Table 2 (R-WdCu, R-vEoU, R-GCLw)
  • We report compute time and GPU memory utilization per attack to compare training costs in a dedicated Table 3.
  • As requested by R-GCLw and following the recent RTA paper, we include PSNR and LPIPS measures for GeoSHield and GeoFA-TD. Results are given in Table 12, Appendix F.2.5.
  • We add a toy model in Appendix E which shows how scale and angular deviations of a velocity path impact trajectory deviation, and measure angular/norm deviations of our attacks on YFCC4k images to test alignement with the toy model (R-WdCu, R-GCLw).
  • To clarify our contributions, e.g. adapting existing attacks from the image literature to geographic RFM and proposing new attacks in that setting, we renamethe adapted attacks as [Attack name]-G (instead of GeoFA-[attack name]) (R-GCLw).
  • We provide results with two geoShield [2] instantiations, including one with image descriptions from an open source VLM (see Table 2). Results are very close to our previous GeoShield implementation. (R-WdCu)

References:

[1] Niccolo Niccoli, Federico Becattini, and Lorenzo Seidenari. Defending from GeoLocalization through adversarial road trips. In Paolo Favaro, Zuzana Kukelova, Atsuto Maki, Anna Rohrbach, Konrad Schindler, and Federico Tombari (eds.), Computer Vision – ECCV 2026, volume 17054 of Lecture Notes in Computer Science, pp. 611–629, Cham, 2026.

[2] Xinwei Liu, Xiaojun Jia, Yuan Xun, Simeng Qin, and Xiaochun Cao. Geoshield: safeguarding geolocation privacy from vision-language models via adversarial perturbations. AAAI, 2026.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.