Mental-health studies increasingly use phone location data as an objective check on what people say about themselves. If a participant reports low stress and good mood but has stopped leaving home, the mismatch between the sensor and the self-report is the signal of interest. Shrinking movement radius and rising time at home often appear exactly when questionnaires are distorted by social desirability (people answering the way they think they should) or by recall bias (memory reshaping past states). That check requires sending a person's movements to a server, which reveals where they live, work, and seek care. The obvious fix is to blur or decoy the locations first, using a location privacy mechanism. But privacy noise and self-report error then become tangled. A pipeline that blurs locations will see larger gaps between sensors and self-reports, and will wrongly read those gaps as people misreporting.
We propose to treat the privacy mechanism as what it is: a measurement error that the study designs and publishes, unlike self-report error, which is unknown. Because its distribution is known, its effect on mobility features can be estimated and removed when the analyst calibrates self-reports against sensors. Because it is chosen by design, it can be tuned to preserve the features that calibration needs rather than geometric accuracy. The two-sentence version: location privacy should be built and analyzed as a known noise channel feeding a self-report calibration model. Done this way, one can have strong protection against tracking without inventing self-report bias that is not there.
Core Claim
Existing location privacy mechanisms measure utility as distance between true and reported points. Privacy Games Along Location Traces poses mechanism design as a Stackelberg game: the user commits to an obfuscation strategy first and an adversary performing Bayesian inference responds. The user's constraint is on expected distance error. Privacy Preservation in Location-Based Services: A Novel Metric and Attack Model introduces the Viterbi tracking attack and the Robust Dummy Generation (RDG) algorithm, which picks fake locations to maximize transition entropy. Neither measures what happens to the behavioral features used in digital phenotyping, meaning the inference of mental state from phone data. Conversely, the phenotyping papers (Monitoring Behavioral Changes Using Spatiotemporal Graphs, Personalized Student Stress Prediction with Deep Multitask Network, Personalised depression forecasting using mobile sensor data and ecological momentary assessment) use raw traces and offer no protection.
The obvious intersection would be to run existing mechanisms on these datasets and report how much prediction accuracy drops. That study is useful but shallow. It treats privacy noise as damage to be measured and does not ask whether the damage can be corrected. Our claim is sharper, and it has three parts. First, privacy noise inflates the apparent self-report discrepancy in a predictable way. Second, because the mechanism is known, that inflation can be estimated and corrected. Third, a mechanism tuned to the calibration task beats a distance-tuned mechanism at equal resistance to tracking.
The core relation is the classical errors-in-variables model. Let $f_{it}$ be person $i$'s true mobility feature on day $t$ (for example, fraction of time at home), let $\tilde f_{it}$ be the same feature computed from the privatized trace, and let $y_{it}$ be the self-report from an ecological momentary assessment (EMA), a brief in-the-moment phone survey. The calibration model is
$$y_{it} = \alpha_i + \beta f_{it} + b_{it} + \epsilon_{it},$$
where $\beta$ links behavior to report, $b_{it}$ is the self-report distortion we want to estimate, and $\epsilon_{it}$ is residual noise. If we fit on $\tilde f$ instead of $f$, the slope shrinks by $\lambda = \mathrm{Var}(f)/(\mathrm{Var}(f)+\mathrm{Var}(\tilde f - f))$, which is the standard attenuation result for an additive, independent error. The lost signal then reappears in the residual as spurious "bias."
Real mechanisms act on coordinates, not features, so the induced feature error is neither additive nor independent. We therefore estimate the full transfer distribution $p(\tilde f \mid f)$. We do this by running the published mechanism many times on a public trajectory corpus, never on participants' raw data, and then using that distribution for correction, through regression calibration or simulation extrapolation. The design objective follows. Among mechanisms that hold an adversary's tracking success below a target, choose the one that minimizes the variance of the corrected estimate of $b_{it}$, not the one that minimizes distance error.
Research Program
The work has three components.
Feature-error transfer functions. For each mechanism family, we estimate how the main phenotyping features are distorted. The mechanism families are geometric perturbation, k-dummy generation (via RDG and the dummy query sequences of Constructing dummy query sequences to protect location privacy and query privacy in location-based services), and Stackelberg-optimal obfuscation. The features are home dwell time, location variance, total distance, location entropy, and circadian regularity, as used in the MAIKI and StudentLife papers. We estimate these transfer functions on GeoLife and then test whether they hold up on StudentLife's Dartmouth traces. That test matters because Beijing commuters and college students move differently, and a mismatched transfer function would bias the correction.
Correction. We fit the calibration model with the correction in place. Then we check whether the per-person discrepancy estimates match those obtained from raw data.
Calibration-aware mechanism design. We modify the Stackelberg linear program, which already allows an arbitrary per-location-pair cost. We replace the distance cost with a semantic cost that penalizes moving a point across categories such as home, campus, and other. These are the categories that drive dwell-time and entropy features. For trajectory-level features such as movement radius, which a per-point cost cannot express, we extend RDG's greedy dummy selection with a constraint on feature-error variance.
Network-level protections are out of scope as mechanisms, but we treat them as a constraint we must not ignore. Phantom routing sends packets on random detours, and cover traffic adds dummy transmissions to hide when real ones occur. The timing of EMA submissions leaks routine and compliance, as the compliance-decay analysis of Assessing Real-Time Moderation makes clear. We will report that leak by adding a timing-based adversary alongside the location adversary.
Baselines
The naive reference point is the raw, unprotected pipeline. On StudentLife this is the spatiotemporal graph network of Monitoring Behavioral Changes Using Spatiotemporal Graphs, a model that combines graph convolutions over similar users with recurrent layers over time. It reports a depression (PHQ-9) RMSE of 3.5 ± 3.1 with all features, 3.5 ± 3.4 with WiFi mobility alone, and a stress F1 of 65.4 ± 3.1. The second raw reference is CALM-Net from Personalized Student Stress Prediction with Deep Multitask Network, an LSTM autoencoder with per-user prediction heads that reports an average stress F1 of 0.594, against 0.408 for Location-MLP. On MAIKI, the reference is subject-dependent standardization from Personalised depression forecasting, which reports a PHQ-2 MAE of 0.801 against 1.062 for the non-personalized model. These raw numbers set the ceiling, and the raw-data discrepancy estimates serve as the reference against which every privatized pipeline is scored. We will re-run these models ourselves on identical splits rather than compare against published numbers, since our feature pipeline and splits will differ.
The strongest existing privacy competitor is RDG feeding the same downstream models without correction. Its reported strength is protecting at least 50% of queried locations against the Viterbi attack when the dummy count k is 15 or more, on GeoLife. The second is the Stackelberg-optimal mechanism with a distance constraint, for which no numeric result on behavioral tasks was found on Lacuna. Every mechanism will be run at matched adversary success under the same Viterbi and home-inference attacks. All mechanisms use the same downstream models with an identical tuning budget for hyperparameters, chosen on the same validation folds.
Two ablations isolate our contributions. "Correction without redesign" applies the transfer-function correction to RDG and to distance-optimal outputs, which isolates the value of knowing the noise. "Redesign without correction" uses the semantic-cost mechanism with a naive fit, which isolates the value of choosing the noise. Our method counts as winning if, at matched tracking success, the corrected discrepancy estimates correlate with raw-data estimates significantly better than the uncorrected RDG pipeline does across bootstrap resamples. It must also cut the rate of false "misreporting" flags, meaning people flagged on privatized data but not on raw data.
Why not just…?
Why not just compute features on the phone and send only summaries? This is the most serious alternative, and it may dominate raw-trace obfuscation. Our response is to include it rather than argue against it. Daily feature series are not harmless: a series of home dwell times and visit counts is a behavioral fingerprint. Co-location and background knowledge already break naive protections, as shown in Quantifying the Effect of Co-location Information on Location Privacy. We therefore run an on-device arm that adds calibrated noise to feature summaries, and we attack it with re-identification and routine inference. The known-channel correction applies directly to that arm, since added feature noise is also a known channel. If this arm wins at matched privacy, the result is still a positive finding for the core claim: privacy should be designed as a correctable channel. Only the best release point changes.
Why not just retrain the downstream model on privatized data and let it absorb the noise? Retraining can recover predictive accuracy, but prediction is not the goal. The models are trained on self-reports, which is the "no gold standard" problem raised by the CALM-Net authors. A model fit end-to-end on noisy features against biased labels will reduce its error partly by learning the bias, because the noisier the features, the more the labels' own patterns dominate. We will test this directly with semi-synthetic labels in which a known bias is injected, and measure whether the retrained models recover the injected $b_{it}$.
Why not just accept the accuracy loss from a standard mechanism? The loss is not just lower accuracy; it is systematically biased toward "this person misreports." For a clinical study, falsely concluding that patients under-report is worse than wide error bars. The uncorrected baselines quantify how large this artifact is. If it turns out to be negligible at useful privacy levels, our correction is unnecessary and we will report that.
Proposed Experiments
The decisive experiments use StudentLife for GPS, WiFi, daily stress and mood EMA, and pre/post PHQ-9. MAIKI supplies GPS-derived features, daily EMA, and weekly PHQ-9, subject to data access. GeoLife is used to estimate transfer functions and to benchmark attacks. The Intern Health Study has no location data, so it is used only to quantify the timing leak in EMA submissions.
Privacy is measured by Viterbi tracking success, home-cell inference accuracy, and re-identification accuracy from feature series. Utility is measured by PHQ-9 RMSE, stress F1, and PHQ-2 MAE. Calibration fidelity is measured by the correlation between privatized and raw per-person discrepancy estimates, the false-flag rate, and, in the semi-synthetic arm, recovery error for injected social-desirability shifts and recall-smoothing biases.
The first decisive check is whether uncorrected privatized pipelines inflate estimated self-report bias in proportion to the predicted attenuation, and whether the correction restores raw-data estimates within their confidence intervals. The second is whether the semantic-cost mechanism beats the distance-optimal and RDG mechanisms on calibration fidelity at equal Viterbi success. The third is whether GeoLife-estimated transfer functions remain valid on StudentLife. If they fail, correction requires in-population estimation, and we will measure what that costs in privacy.
References
- Privacy Games Along Location Traces: A Game-Theoretic Framework for Optimizing Location Privacy
- Privacy Preservation in Location-Based Services: A Novel Metric and Attack Model
- Monitoring Behavioral Changes Using Spatiotemporal Graphs: A Case Study on the StudentLife Dataset
- Personalized Student Stress Prediction with Deep Multitask Network
- Personalised depression forecasting using mobile sensor data and ecological momentary assessment
- Constructing dummy query sequences to protect location privacy and query privacy in location-based services
- Assessing Real-Time Moderation for Developing Adaptive Mobile Health Interventions for Medical Interns: Micro-Randomized Trial
- Quantifying the Effect of Co-location Information on Location Privacy