PRISMA generates realistic private in-car energy traces using low-dimensional latent space
The claim
PRISMA achieves better differentially private synthesis performance at ε=0.5 than prior tabular and marginal-based DP baselines
Published as: “PRISMA: Differentially Private Synthetic In-Car Telemetry via Differentiable Simulation”
So what
PRISMA enables the release of high-fidelity synthetic in-car energy telemetry with strong differential-privacy guarantees, facilitating data sharing for research and industry without exposing individual driver behavior.
The five-minute version
PRISMA retains fidelity under a strong privacy budget (ε = 0.5), achieving the highest recall (0.999) and an unchanged FID (0.96×) compared to non-private baselines, making it the most faithful model when privacy is enforced. It works by encoding trips into a low-dimensional descriptor via a differentiable physics simulation, applying DP synthesis only to this descriptor, and reconstructing full traces with public data. The evidence is based on experiments with 10,000 training trips and 2,000 held-out trips, showing PRISMA’s fidelity is largely unchanged under privacy while other models degrade substantially.
What they found 3
- PRISMA’s FID remains essentially unchanged under ε = 0.5, making it the most faithful model under privacy.
“PRISMA’s FID is unchanged within noise (0.96×).”
- PRISMA achieves the highest recall (0.999) and recall is unchanged under privacy.
“PRISMA achieves the highest recall in both conditions and remains unchanged under privacy (0.999).”
- Overall fidelity of PRISMA is largely unaffected by the privacy mechanism.
“PRISMA’s fidelity is largely unchanged.”
What undercuts it the paper’s own results
- PRISMA is slightly worse than the best baseline but still competitive.
Stated limits 3
- PRISMA represents each trip using only three fitted quantities (aggressiveness, detour, and vehicle) beyond its endpoints. This enables privatisation in a low-dimensional space rather than over raw telemetry, but necessarily discards variation not captured by these parameters. For example, weather, auxiliary loads, and fine-grained driving behavior cannot be recovered by the decoder, irrespective of the privacy budget.
- Reconstruction fidelity is constrained by the coverage and quality of the public road network, elevation data, and the EPA vehicle catalog used by the decoder. Trips involving vehicles poorly represented in the candidate pool, or routes that differ substantially from those selected by the public router, may therefore be difficult to reconstruct faithfully even when the underlying descriptor is accurately estimated.
- Our evaluation is conducted on a dataset with a specific geographic and temporal scope (Section 6). This may limit the generality of our findings, particularly where road topology, vehicle composition, driving behavior, or environmental conditions differ compared to the eVED dataset.
What’s new here 4
- Encodes high-dimensional telemetry into a compact, physically-grounded latent descriptor using differentiable simulation.
- Applies a standard DP tabular synthesis algorithm (AIM) only to the low-dimensional descriptor, avoiding the dimensionality penalty of DP on raw time series.
- Uses the simulation as a privacy-free post-processing step to reconstruct full traces, a combination not explored before.
- Demonstrates that this approach preserves both fidelity and diversity under ε = 0.5, outperforming prior DP time-series generators.
Artifacts & methods 4
- PRISMAsystem
- AIM (DP tabular synthesis algorithm)primitive
- Differentiable physics simulationprimitive
- eVED datasetdataset