When a location-sharing app blurs your GPS position by a few hundred meters, the usual way to check whether it protects you is to train an attacker model and see how well it guesses where you really went. A useful attacker guesses what kind of place you visited, such as a clinic, a bar, or a place of worship. This note argues that such checks have a hidden variable: how much computation the attacker spends refining its guesses. A real adversary does not classify each blurred visit in isolation. It forms a rough picture of who you are from all your visits, uses that picture to re-guess each visit, updates the picture, and repeats. We propose to build this iterative attacker, measure how much extra semantic information each refinement pass extracts from obfuscated traces, and find where the gains flatten. The broader claim is that privacy should be reported at the attacker's saturation point, not after a single pass.
The idea joins two literatures that have not been connected. Work on semantic location privacy shows that blurring coordinates leaves the type of activity exposed, because shops, restaurants and clinics cluster in space. Work on iterative refinement, across vision, parsing and text generation, shows that repeated self-correction gains most in the first two or three passes and then plateaus, while cost grows linearly with the number of passes. We bring the second literature into the first as a measurement tool. If an attacker's refinement saturates quickly and each pass is cheap, a defense evaluated against a one-pass attacker is being overrated, and auditors can say exactly how many passes they need to run.
Core Claim
The closest existing attack is the study by Wiedemann et al. in Where you go is who you are. It describes each noisy visit with time features and counts of nearby points of interest (POIs, the shops, venues and facilities on a map), trains gradient-boosted trees (XGBoost) to predict one of 12 venue categories, and reaches over 60% accuracy under 100 m of noise. Above-chance accuracy persists at about 1 km. The study also finds that "soft" user profiles, built by averaging predicted probabilities, beat "hard" profiles built from vote counts. That profile is computed once, after classification, and never fed back into the classification of individual visits.
Our claim is that feeding it back helps, and does so iteratively. Let $f(x_i)$ be the single-pass classifier's probability vector over categories for visit $i$ with noisy observation $x_i$. Let $\pi^{u}_{-i,t}$ be user $u$'s soft profile at pass $t$, built from the refined posteriors of all of $u$'s other visits. Excluding visit $i$ keeps the attacker from confirming its own guess. The refined posterior at pass $t$ is
$$ p_t(c_i) \;\propto\; f(x_i)[c_i]\,\big(\pi^{u}_{-i,t-1}(c_i)\big)^{\beta}, $$
where $c_i$ is the candidate category and $\beta$ is a learned weight on the profile. At $t=0$ the attacker uses $f$ alone, which is the Wiedemann et al. attack. This is a form of collective inference. It differs from prior location attacks because it treats the number of passes $T$ as a measured axis of attacker strength.
We predict two things. First, accuracy at saturation will exceed the one-pass accuracy, especially at larger noise radii where $f$ alone is uncertain and the profile carries more of the signal. Second, the gain will saturate within two or three passes, the pattern reported for recurrent feature fusion in DeFusionNET and for two-pass decoding in Deliberation Networks. If both hold, the honest privacy number for any defense is accuracy at the saturation pass, and it can be computed cheaply.
Research Program
The work has three parts.
The first part is the attacker. We reuse the Wiedemann et al. features unchanged for $f$, so the only new ingredient is the feedback loop. We add a second variant for road-constrained data. There, each pass also prunes candidate true locations to map-feasible vertices before scoring categories, in the spirit of the reachability filter in VehiTrack from Protecting Vehicle Location Privacy with Contextually-Driven Synthetic Location Generation. That work showed map constraints alone dismiss 81.69% of candidates produced by standard geometric noise.
The second part is cost. Building the spatial features, meaning the k-nearest-POI counts and distances to each category, is by far the most expensive step, and these features do not change across passes. Following the shortcut idea in iRe2f, which avoids re-encoding the input and applies only a small update to the existing representation, each refinement pass recomputes only the profile term. Its cost is linear in a user's number of visits. We will report wall-clock time and operation counts per pass. This shows whether extra passes are nearly free for an attacker, which would determine whether auditors must assume saturation.
The third part is halting. Following the early-stop classifier in Multi-pass Decoding for Grammatical Error Correction, a small logistic model predicts per user whether another pass will change any argmax label. We expect this to matter only for audit cost at city scale, not for accuracy, and we will test that expectation directly.
The defenses we audit play the role of the "application regime." They are planar Laplace noise (standard geo-indistinguishability) at several radii, and the Graph-Exponential Mechanism (GEM) from Geo-Graph-Indistinguishability. GEM samples pseudolocations on road vertices with probability decaying in shortest-path distance, and its authors explicitly leave semantic privacy open. We also audit a contextual synthetic-location defense in the style of TransProtect from Yadav et al., reimplemented from its description because no code is listed in our sources. Our interest is whether the ranking of these defenses by semantic leakage changes between $T=0$ and $T$ at saturation. A defense that looks strong only against one-pass attackers would be the most consequential finding.
Baselines
The naive reference point has two parts. One is the prior over the 12 categories, guessing the majority category. The other is the attacker applied to un-obfuscated coordinates, which gives an upper bound on leakage.
The strongest existing attack is the one-pass XGBoost classifier of Wiedemann et al., with its soft profiling applied after the fact. The reported figure is over 60% categorization accuracy among 12 categories at 100 m of noise on their data. We will retrain it with the same features, data splits and hyperparameter search budget as our refined attacker. All refinement variants share the identical trained $f$, so any difference comes from the feedback loop alone. Beating this baseline means a statistically reliable accuracy gain at matched noise radius, measured by paired bootstrap over users, at one or more radii between 100 m and 1 km.
The strongest competitor in spirit is a single-pass sequence model over a user's whole trace, a bidirectional LSTM like the reconstruction phase of VehiTrack in Yadav et al.. It can in principle learn the same cross-visit dependencies in one forward pass. We will give it the same per-visit features and a training compute budget at least equal to the refined attacker's. For it, Yadav et al. report only relative inference-error reductions on taxi location reconstruction: 61.20% against Laplace and 47.35% against linear-programming obfuscation. No reported result was found on Lacuna for semantic category accuracy for VehiTrack or TransProtect.
Two ablations isolate the key ingredient. The first is one-round stacking: XGBoost with the leave-one-out user profile added as an input feature, trained once. This is exactly $T=1$ of our scheme with the weight learned inside the trees. The second sets $\beta=0$, which recovers $T=0$ while keeping the code path identical. A third ablation drops the leave-one-out exclusion, to measure how much apparent gain comes from self-confirmation.
Why not just…?
Why not just add the user profile as a feature once (stacking) and stop? This is the real threat. If one round captures nearly all the gain, the "iterative" part of the attacker is unnecessary. The experiments settle this directly, because stacking is our $T=1$ point. Even in that case the audit conclusion survives in weaker form: privacy evaluations that use one-pass attackers, as in Wiedemann et al., still underestimate leakage. The methodological recommendation would then reduce to "always include one profile round." We will state the result as it falls.
Why not just train a bigger single-pass sequence model over the whole trace? It might match the refined attacker's accuracy. Our claim is not that refinement is the only way to exploit cross-visit structure. It is that leakage should be reported at the point where extra attacker effort stops paying. We will compare both attackers at matched training and inference compute. If the sequence model wins, it becomes the reference attacker for the audit, and the pass-count curve is replaced by a curve over model capacity. The audit methodology is unchanged.
Why not just raise the noise radius until the attack fails? Wiedemann et al. report that re-identification risk becomes manageable between 200 m and 400 m, but category predictability persists out to 1 km because of spatial clustering of venues. The service-quality costs of such large radii are what GEM and TransProtect exist to avoid. Measuring where the refined attacker's curve meets chance level tells practitioners how much more noise is needed than a one-pass audit suggests.
Why not just run a fixed three passes and skip the halting classifier? For accuracy this is probably fine, and we expect fixed $T=3$ to match adaptive halting. The halting classifier is tested only as a cost saving for city-scale audits, and it is dropped if it saves little.
Proposed Experiments
The main experiments use the Foursquare check-in data and the Yumuv GNSS tracking data, the two datasets of Wiedemann et al., so our $T=0$ numbers can be checked against theirs. We add GeoLife stay points matched to OpenStreetMap POIs as a held-out third city with irregular mobility. The road-constrained variant runs on the Rome and San Francisco taxi traces over OpenStreetMap road networks, as in Yadav et al. Category labels there come from POIs near drop-offs, so those results are treated as secondary.
For each dataset, defense and noise level, we report three quantities. The first is 12-way category accuracy and macro-F1 at each pass from 0 to 5. The second is the saturation pass, defined as the first pass after which the gain falls below the bootstrap confidence width. The third is attacker cost per pass. Service quality for each defense is reported as mean displacement and road-distance distortion, so defense comparisons stay at matched utility.
Three decisive checks determine the outcome. The first is whether saturation accuracy exceeds both the one-pass baseline and the stacking ablation at any noise level. If it exceeds neither, the claim fails beyond one profile round. The second is whether the gain grows with noise radius, as the posterior equation predicts when $f$ becomes less informative. The third is whether any pair of defenses swaps order between $T=0$ and saturation. A swap, for instance GEM's graph-shaped noise turning out less robust to profile feedback than Laplace noise of equal utility, would show that current semantic privacy evaluations can mis-rank defenses. That would be the strongest reason for the field to report leakage against a saturated attacker.
References
- Where you go is who you are: a study on machine learning based semantic privacy attacks
- DeFusionNET: Defocus Blur Detection via Recurrently Fusing and Refining Multi-Scale Deep Features
- Deliberation Networks: Sequence Generation Beyond One-Pass Decoding
- Protecting Vehicle Location Privacy with Contextually-Driven Synthetic Location Generation
- iRe2f: Rethinking Effective Refinement in Language Structure Prediction via Efficient Iterative Retrospecting and Reasoning
- Multi-pass Decoding for Grammatical Error Correction
- Geo-Graph-Indistinguishability: Location Privacy on Road Networks with Differential Privacy