Cities estimate how many trips go from each origin to each destination (an "origin–destination matrix") by tuning a traffic simulator until its simulated road counts match what roadside sensors recorded. Increasingly, the data behind that tuning come from phones and connected vehicles, and aggregated counts of that kind are not safe to publish as they are: individual trajectories can be rebuilt from them. The standard fix is differential privacy, which adds calibrated random noise so that no single person's trips change the release much. Nobody knows what that noise does to simulator calibration, or which privatized release a data holder should publish if the goal is a well-calibrated city model.
This note proposes to answer that with a controlled study and one specific method. The method is to publish two cheap private releases, noisy sensor counts and a noisy origin–destination matrix, and reconcile them through the simulator using an objective that weights each by its known noise level. The hypothesis is that this hybrid beats each single release, and beats publishing differentially private synthetic trajectories, once origin–destination cells become sparse. It is in exactly that sparse regime that synthetic-trajectory fidelity scores, the "synthetic benchmark" of the privacy literature, should stop predicting real downstream usefulness.
Core Claim
Two research communities each test on stand-ins for the real problem. Optimization researchers have shown that clean synthetic test functions mislead algorithm design. One such study found real hyperparameter-tuning landscapes so different from standard test functions that a decision tree separated them with 96.46% accuracy (HPO ˟ ELA). Trajectory-privacy researchers judge their generators by distributional fidelity scores, such as divergence between trip-length histograms or relative error of cell densities. They rarely ask whether the generated data support the decisions it is released for.
BO4Mob gives us a realistic downstream task. It couples the Simulation of Urban MObility (SUMO) traffic simulator with real Caltrans PeMS freeway counts in San Jose, and poses origin–destination estimation as black-box optimization in 3 to 10,100 dimensions. Its targets, however, are raw and unprotected. The opportunity is to bring privacy into that task.
We claim three things, each testable:
- Different releases win in different regimes. Which privatized release is best depends mainly on how many trips fall into a typical released cell. A noisy origin–destination matrix wins when cells are dense. Noisy sensor counts calibrated through the simulator win when cells are sparse, because each sensor sums many trips and its relative noise is far smaller.
- A variance-weighted hybrid beats both. Combining the two releases with weights set by their known noise levels dominates either one alone.
- Fidelity scores mispredict utility. Fidelity scores of private synthetic trajectories rank generator configurations poorly with respect to calibration error in the sparse regime.
Prior Work and Its Role Here
The motivation for privatizing counts at all comes from Trajectory Recovery From Ash. It recovered individual trajectories from un-noised aggregate counts with 73%–91% accuracy, and showed two frequent locations identify over 95% of users. Privacy-Friendly Mobility Analytics using Aggregate Location Data found aggregate location data vulnerable to membership and re-identification attacks. Both rule out releasing unprotected aggregates. Neither measures what noise costs a downstream user, which is what we measure.
The trajectory-release competitors are PrivTrace and DPTraj-PM. PrivTrace builds a private adaptive first/second-order Markov model on a two-level grid. DPTraj-PM combines a noisy prefix tree with an order-$m$ Markov process. Both report fidelity metrics only, both note that fine grids drown sparse cells in noise, and DPTraj-PM explicitly leaves downstream use untested. We use them as baselines and as the test bed for the fidelity-versus-utility question.
Two other papers constrain the design. PateGAIL++ privatizes generative imitation learning and checks only individual-level statistics, not city-scale flows. We include it only if its code runs on simulator traces. Arcolezi et al. find that noise added during training (DP-SGD) beats noise added to inputs for forecasting. That is a reminder that where noise enters matters. Our setting differs, though: the downstream user here is a simulator calibrator, not a trained network, so the release itself must be private.
From the optimization side, BO4Mob supplies the scenarios, optimizers, and its "visibility gap" observation: origin–destination pairs that no sensor sees give the optimizer no feedback. A private origin–destination prior is a natural way to pin those pairs down. MOEA/D normalization shows how optimizers break when objective terms differ wildly in scale. Our two data terms do differ by orders of magnitude. Unlike that work, we do not tune the balance heuristically: the known noise of each privacy mechanism fixes it.
Research Program
We build a semi-synthetic private version of BO4Mob so that ground truth exists.
Ground truth and data. For the 21-, 44-, and 151-dimensional scenarios, we fix a reference matrix $\theta^\*$, taken as the best non-private TuRBO solution on real PeMS counts. SUMO then generates individual vehicle trips from $\theta^\*$. A trusted data holder holds these trips. We vary trip density by scaling total demand and by splitting the time window into finer slots. This moves the typical trips-per-cell across two orders of magnitude without changing the optimization dimension.
Privacy unit and releases. The privacy unit is one trip. We clip each trip to at most $k$ sensor-slot crossings, so the count release has L2 sensitivity $\sqrt{k}$ and the origin–destination release has sensitivity 1. Both use the Gaussian mechanism. Budget is split between the two releases by sequential composition. Post-processing is free under differential privacy, so we then project the noisy counts onto freeway flow-conservation constraints (upstream equals downstream plus on-ramps minus off-ramps) and onto non-negativity.
Calibration objective. The calibrator minimizes
$$ J(\theta) = \sum_{s} \frac{\big(c_s(\theta) - \hat c_s\big)^2}{\sigma_c^2 + \tau_s^2} \;+\; \sum_{ij} \frac{\big(\theta_{ij} - \hat o_{ij}\big)^2}{\sigma_o^2}. $$
Here $c_s(\theta)$ is the simulated count at sensor-slot $s$ for candidate matrix $\theta$, and $\hat c_s$ is the released, projected noisy count. $\sigma_c^2$ is the post-projection noise variance, computable from the mechanism and the projection. $\tau_s^2$ is SUMO's own run-to-run variance at that sensor, estimated from replicate runs. $\hat o_{ij}$ and $\sigma_o^2$ are the noisy origin–destination entry and its variance. This is a Gaussian negative log-likelihood, so the relative weight of the two terms is fixed by the mechanisms, not tuned.
Expected crossover. For a cell with $n$ trips, the relative error of a direct release scales like $\sigma_o/n$. A sensor aggregating $N$ trips has relative error near $\sigma_c/N$ with $N \gg n$. Simulator calibration converts the second error into origin–destination error through the network's routing structure, which predicts the regime crossover in claim 1.
Optimization-side question. We also ask whether conclusions drawn on clean BO4Mob survive privatization. We run BO4Mob's optimizers (vanilla Bayesian optimization, TuRBO, SAASBO, the stochastic-approximation method SPSA, random search) on the private objectives. We compute inexpensive landscape features on the 21- and 44-dimensional scenarios, where sampling is affordable. Privatization noise is a fixed perturbation of the target, not fresh noise at each evaluation. We therefore predict it mostly moves the minimizer and raises the error floor rather than roughening the landscape, which would leave optimizer rankings stable. Finding the opposite would itself be a useful negative result for the field.
Baselines
Reference points. The naive floor is BO4Mob's initial origin–destination guess with no data, together with the same guess calibrated on un-noised counts. The latter is the non-private ceiling. Privacy cost is reported as the fraction of the ceiling's improvement retained.
Plain noisy-count calibration. TuRBO minimizes BO4Mob's error metric, normalized root mean squared error (NRMSE), directly against noisy sensor counts, with no projection and no prior. This is what a practitioner would do first. BO4Mob reports TuRBO at a 69.5% improvement over initial estimates on the 151-dimensional Small Region with clean counts. That is the strongest optimization number in the file and our ceiling's anchor.
Direct private origin–destination release. The Gaussian-noised matrix, clipped at zero, is used as the estimate with no simulator. It is the simplest serious competitor and likely the winner in dense regimes.
Private synthetic trajectories. PrivTrace and DPTraj-PM are trained on the SUMO trips at the same total budget. Their outputs are aggregated into an origin–destination matrix, which is used directly and also as the starting point for count-based calibration. Grid size spans 10×10 to 40×40 and Markov order 1–3, and the best configuration is selected on a held-out validation city slice. No reported result found on Lacuna for either method on simulator calibration error. Their only reported comparisons are fidelity metrics, such as PrivTrace's over-50% divergence reduction versus AdaTrace and DPT.
Fairness controls. Every arm uses the same budget $\epsilon \in \{0.1, 0.5, 1, 2, 5, 10\}$, the same number of simulator evaluations per run, the same optimizer (TuRBO, with SPSA as a check), equal tuning trials, and ten seeds over both privacy noise and simulator noise.
Ablations. Three ablations isolate the key ingredient: replacing the variance-derived weights in $J$ with a single tuned scalar, removing the flow-conservation projection, and dropping either data term.
What counts as beating. The hybrid beats a baseline if its origin–destination error relative to $\theta^\*$ is lower at a given $\epsilon$, with seed-level confidence intervals not overlapping, in the sparse regime. It must also be no worse in the dense regime.
Why not just…?
Why not just add noise to the counts and run TuRBO? Squared error against a target with constant-variance additive noise has the same expected minimizer as against the clean target. Plain calibration may therefore already be nearly optimal when noise is uniform and sensors are numerous, and this is a real threat to the count-side contribution. The plain-calibration baseline and the projection ablation settle it. If they match the projected version, we report that the count-side machinery is unnecessary. The claimed gain would then rest only on adding the origin–destination prior for sensor-invisible pairs, which plain calibration cannot see by construction.
Why not just release a private origin–destination matrix and skip the simulator? In dense regimes this probably works as well as anything, and we expect it to. Its weakness is the same sparse-cell noise problem PrivTrace and DPTraj-PM describe for fine grids: when most cells hold a handful of trips, noise dominates them. The density sweep is the decisive experiment. If the direct release wins even at the sparsest setting, claim 1 is false, and the recommendation to data holders becomes "release the matrix."
Why not just publish private synthetic trajectories, which serve many uses at once? Synthetic trajectories are more general. However, each budget unit is spread over a full transition model with many more cells than the calibration task needs. In addition, PrivTrace reports that long, multi-hop trips, which load freeway sensors most heavily, are reproduced worst. The synthesizer arm measures this directly. A rank correlation between fidelity scores and calibration error across grid and order settings tests whether fidelity scores are a usable proxy.
Proposed Experiments and Decisive Checks
The main experiment crosses release type, budget, and trip density on the 21-, 44-, and 151-dimensional scenarios. The 10,100-dimensional scenario is excluded because a single simulation exceeds 11 hours and BO4Mob shows no optimizer improves on the initial guess there even without noise. The primary metric is origin–destination error relative to $\theta^\*$.
Two secondary metrics are counts at sensors withheld from every release, measured against clean simulated counts, and NRMSE on a held-out time window. The withheld-sensor metric is the direct test of whether the private prior closes the visibility gap.
Three outcomes would be decisive.
- Crossover. The density at which the direct matrix release stops beating calibration should fall where the scaling argument predicts. The hybrid should sit at or below both curves everywhere.
- Fidelity proxy. The Kendall rank correlation between synthesizer fidelity scores and calibration error should fall toward zero in sparse regimes. That would carry the synthetic-to-real lesson from optimization benchmarks over to privacy evaluation.
- Ranking stability. The Kendall correlation between optimizer rankings on clean and privatized objectives should either stay high, confirming that fixed-target noise mainly shifts the minimizer, or drop, which would mean BO4Mob's clean-data conclusions do not carry over to privacy-protected deployments.
As a final sanity check, we run the assignment-based recovery attack of Xu et al. on the SUMO trips. We run it against both raw and privatized aggregate counts, to show the privacy margin is not merely theoretical.
References
- HPO ˟ ELA: Investigating Hyperparameter Optimization Landscapes by Means of Exploratory Landscape Analysis
- BO4Mob: Bayesian Optimization Benchmarks for High-Dimensional Urban Mobility Problem
- Trajectory Recovery From Ash: User Privacy Is NOT Preserved in Aggregated Mobility Data
- Privacy-friendly mobility analytics using aggregate location data
- PrivTrace: Differentially Private Trajectory Synthesis by Adaptive Markov Model
- DPTraj-PM: Differentially Private Trajectory Synthesis Using Prefix Tree and Markov Process
- PateGAIL++: Utility Optimized Private Trajectory Generation with Imitation Learning
- Differentially Private Multivariate Time Series Forecasting of Aggregated Human Mobility With Deep Learning: Input or Gradient Perturbation?
- On the effect of normalization in MOEA/D for multi-objective and many-objective optimization