Privacy-Protecting Synthetic Trajectories

Updated 9 May 2026

Privacy-protecting synthetic trajectory generation is a method that creates synthetic mobility datasets preserving key spatiotemporal patterns while preventing individual re-identification.
It leverages statistical models such as Markov chains, prefix trees, and adaptive noise allocation based on differential privacy to maintain both utility and privacy guarantees.
The approach requires careful tuning of parameters like grid granularity and Markov order to balance the trade-off between noise-induced privacy and accurate data representation.

Privacy-protecting synthetic trajectory generation addresses the need to share human mobility data for research and practical analyses while maintaining strict safeguards against individual privacy leakage. This challenge has led to the emergence of diverse algorithmic frameworks—rooted in differential privacy, local differential privacy, and generative modeling—designed to produce synthetic trajectory datasets that maintain essential spatiotemporal characteristics without incurring re-identification risks or regulatory violations.

1. Foundations: Differential Privacy and Local Differential Privacy

Differential privacy (DP) provides a formal mathematical guarantee that the inclusion or exclusion of any individual's trajectory in a dataset has a bounded effect on the output. For trajectory synthesis, this is typically formalized at the trajectory-instance level: a randomized mechanism $\mathcal{M}$ satisfies $\varepsilon$ -DP if for any pair of datasets differing in one user’s trajectory and any measurable output set $O$ ,

$\Pr[\mathcal M(D)\in O]\;\le\;e^\varepsilon\;\Pr[\mathcal M(D')\in O].$

Local differential privacy (LDP) is a stronger model where data are perturbed on the user’s device before release. An LDP mechanism $\Psi$ on user trajectory domains $X$ satisfies $\varepsilon$ -LDP if for all $x, x'\in X$ , $S\subset Y$ ,

$\Pr[\Psi(x)\in S] \le e^\varepsilon \Pr[\Psi(x')\in S}.$

LDP has the advantage that the aggregator (curator) never observes true trajectories, mitigating centralized attack vectors (Du et al., 2023, Hu et al., 2024).

2. Statistical Modeling: Markov Models, Prefix Trees, and Grid Discretization

Many DP trajectory generators discretize space (and sometimes time) into a regular grid, translating each GPS trace into a sequence of grid cells or “anchor points” (Wang et al., 2024, Ge et al., 2024). Trajectories are then modeled as Markov chains, typically of order one or two, capturing either immediate or short-range sequential dependencies.

In DPTraj-PM (Wang et al., 2024), a height- $\varepsilon$ 0 prefix tree captures initial trajectory segments (global directionality), while a noisy $\varepsilon$ 1-gram Markov frequency matrix encodes next-move probabilities. Both representations are differentially privatized via Laplace noise added to cell or transition counts, with total $\varepsilon$ 2 split adaptively between the prefix and transition mechanisms. Consistency constraints (e.g., bottom-up node count enforcement) and parallel composition of DP mechanisms guarantee overall $\varepsilon$ 3-DP.

Grid granularity and Markov order are critical parameters: finer grids and higher Markov order capture finer spatial and temporal behaviors but increase the model’s domain size, thus amplifying noise under fixed $\varepsilon$ 4 (Wang et al., 2024, Wang et al., 2022). Practical grid sizes are recommended to maintain an average of at least 10 points per cell (Wang et al., 2024).

3. Local Differential Privacy: User-side Perturbation and Streaming

LDPTrace (Du et al., 2023) and RetraSyn (Hu et al., 2024) exemplify the local model. In LDPTrace, each user encodes three key aspects: trajectory length, intra-trajectory transitions (first-order Markov), and origin/destination events, as one-hot vectors and perturbs them via Optimized Unary Encoding (OUE). The server aggregates noisy vectors to infer global trajectory-length, transition, and OD distributions. Synthetic trajectories are sampled via the aggregated Markov model and length/OD marginals.

RetraSyn addresses real-time streaming and employs w-event LDP (protecting any window of w consecutive timesteps) with adaptive, error-minimizing budget allocation. The system updates a global Markov mobility model only for transitions exhibiting significant drift, reducing unnecessary budget expenditure and improving stream utility (Hu et al., 2024). The synthesis process accounts for entering and quitting probabilities to generate lifelike streaming behaviors.

4. Generative and Hybrid Methods: Beyond First-Order Sequences

Recent work moves beyond Markov chains to hybrid and deep-learning approaches. DP-STTS (Ge et al., 2024) discretizes both space and time into spatiotemporal cubes and uses sanitized start-cube and transition frequencies (first-order Markov in the cube space) to sample synthetic traces, enforcing $\varepsilon$ 5-DP via Laplace mechanism. PrivTrace (Wang et al., 2022) adaptively combines first- and second-order Markov models, uses two-layer spatial discretization, and applies convex optimization to reconstruct unbiased trip distributions.

DPE (Liu et al., 2023) diverges from cell-count statistics, instead privatizing pairwise trajectory distances (vector-translation invariant aggregation) by adding Laplace noise to the sum of distances for each trajectory. This permits constrained gradient-based synthesis (allowing explicit traffic or geographical constraints) and achieves $\varepsilon$ 6-DP on distances with no further budget needed for the optimization post-processing.

5. Empirical Utility and Privacy Evaluation

Evaluation metrics span spatial fidelity (Jensen–Shannon divergence (JSD) on grid or trip distributions, range-query absolute error, density error), sequential pattern preservation (F1 measures on frequent sub-trajectories, transition pattern AvRE), and rank-based utility (Kendall’s tau on hotspot orderings). Tables of key results indicate the state-of-the-art models such as LDPTrace and DPTraj-PM achieve density and trip JSD an order of magnitude lower than prior LDP n-gram and hierarchical DP methods; LDPTrace achieves JSD $\varepsilon$ 7 on Porto taxi data versus $\varepsilon$ 8 for n-gram (Du et al., 2023).

Privacy analysis requires more than formal guarantees; attack resilience (trajectory user-linking, membership inference) is empirically tested. Models satisfying LDP or trajectory-level DP (with the correct unit of privacy) withstand trajectory linking and reconstruction attacks, while weaker or improperly configured mechanisms (e.g., location-level DP, GANs with no formal DP) are vulnerable (Buchholz et al., 2024, Cherigui et al., 21 Apr 2026).

6. Trade-offs, Deployment Recommendations, and Limitations

There exists a fundamental privacy–utility trade-off: smaller $\varepsilon$ 9 yields higher noise and degrades utility, especially for high-dimensional or fine-grained models. Approaches such as adaptive Markov order selection, grid-size tuning (e.g., $O$ 0 in LDPTrace), and proportionate budget splits help identify the practical sweet-spot (Du et al., 2023).

Key deployment recommendations:

Choose grid/cube resolution to maintain per-cell event frequencies sufficient to overcome DP noise; avoid over-discretization on sparse data (Wang et al., 2024, Ge et al., 2024).
Prefer higher Markov order (2–3) for utility if data density permits (Wang et al., 2024).
Budget should be allocated preferentially (e.g., $O$ 1 to prefix component in DPTraj-PM) (Wang et al., 2024).
For streaming data, adaptive error-aware updating and budget reallocation are critical for sustaining privacy over indefinite horizons (Hu et al., 2024).

Limitations include:

Vulnerability to utility collapse on sparse or long-range behaviors due to noise amplification.
Static uniform grids may not align with heterogeneous spatial density—extensions to adaptive/hierarchical grids or functional data spaces are proposed.
While LDP models avoid central attacks, per-user overheads may rise if frequent reporting is mandated.
Most synthetic generators cannot yet simultaneously satisfy rigorous DP, complete empirical privacy, point-sequence fidelity, and real-world scalability (Buchholz et al., 2024).

7. Directions for Future Research

Several open problems are highlighted in recent systematization of knowledge (SoK) works:

Attaining formal ( $O$ 2, $O$ 3)-DP at instance-level for variable-length sequence models that preserve high spatial and sequential fidelity remains open (Buchholz et al., 2024).
Hybridization with deep generative architectures (e.g., diffusion models, imitation learning frameworks) under DP constraints is promising but not fully mature (Guo et al., 8 Jul 2025, Wang et al., 2024).
Membership-inference-resistant evaluation should be standard; most current GAN- and VAE-based synthetic traces fail this test unless trained with DP-SGD (Cherigui et al., 21 Apr 2026).
Extensions to road-network, semantic, or functional-curve domains call for new privacy-preserving model architectures (rotation-invariant metrics, fine-grained constraint enforcement) (Burzacchi et al., 2024).

The field continues to advance towards frameworks that can generate city-scale synthetic mobility datasets with quantifiable privacy (at the trajectory/agent level), verifiable utility for downstream analysis, and operational efficiency for real-time or offline applications (Du et al., 2023, Wang et al., 2024, Hu et al., 2024, Wang et al., 2022).