Search
Mobile menu Mobile menu
Security , Supply Chain , Data Engineering Nov 17, 2026

GPS Removal Is Not a Privacy Strategy: What CAN Bus Data Tells Attackers About Your Fleet

VECTOR Labs Team
VECTOR Labs Team
GPS Removal Is Not a Privacy Strategy: What CAN Bus Data Tells Attackers About Your Fleet
Last updated on: Nov 17, 2026

Fleet telematics and usage-based insurance platforms have operated for years under a working assumption that removing GPS coordinates from driving logs constitutes meaningful de-identification. It does not. The signals that remain after stripping location fields, specifically wheel speed, steering angle, pedal inputs, and yaw rate, carry enough geometric and contextual information to reconstruct vehicle routes at scale. Understanding why this is true, and what a defensible architecture actually requires, is now a practical obligation for any engineering or data leader responsible for naturalistic driving data.

The Signal Layer Is the Privacy Surface

Modern vehicles broadcast dozens of signals over the controller area network at high frequency. These signals are accessible through the OBD-II port and are routinely collected by aftermarket dongles, fleet telematics hardware, and insurance telemetry platforms. The industry's instinct has been to treat GPS as the sensitive field and everything else as operational telemetry. That framing is wrong.

Speed and steering sequences encode the geometry of a journey. When those sequences are matched against a road network, the candidate set of plausible routes collapses quickly. The privacy surface is not the GPS field; it is the time-series of vehicle dynamics that persists after that field is removed.

What Route Recovery Actually Looks Like

Recent research makes the attack surface concrete. Wang et al. (Wang et al., HuggingFace 2026) introduce CANTraj, a naturalistic driving dataset pairing decoded CAN events with time-aligned GNSS trajectories across 3,940 drives from 454 drivers. Their benchmark evaluates six re-implemented trajectory recovery methods under controlled information-access protocols.

The results are operationally significant. When steering signals are added to speed and matched against a road network using an online map filter, five-minute route recovery rises from 56% to 87%, with a median average displacement error of 43 metres. An offline beam search approach recovers 88% of test routes when a starting heading is known, and 81% without it. An LLM-based route selection harness recovers 79% of routes on held-out test drives.

These are not laboratory edge cases. The dataset covers 1,380 hours and 85,919 kilometres of real driving from 431 drivers, and the strongest methods generalise to 94 drivers the models had never seen. The practical implication is that any pipeline exporting CAN signal sequences without GPS is still exporting route-reconstructible data.

Why Existing De-identification Practices Fall Short

The standard playbook for telematics data sharing involves dropping GPS columns, truncating timestamps, and occasionally coarsening speed to binned categories. Each of these steps addresses the metadata layer while leaving the signal layer intact. Coarsening speed reduces precision but does not eliminate the directional signature encoded in the sequence of acceleration and deceleration events relative to road geometry.

Steering angle is particularly problematic. Turns are sparse, distinctive, and tightly constrained by road network topology. A sequence of left-right-left events at specific time intervals maps to a small number of real-world intersections. Removing GPS does not remove that constraint; it just forces an attacker to do slightly more work.

The regulatory consequence is direct. Data minimisation obligations under frameworks such as GDPR require that data be processed in a form that cannot reasonably be used to identify individuals. A dataset that enables 81-88% route recovery from unseen drivers does not meet that standard, regardless of whether it contains a latitude column.

Architectural Controls That Actually Reduce Risk

Defensible data minimisation at the signal layer requires intervention earlier in the pipeline than most teams currently apply it. The following controls address the mechanism rather than the symptom.

Signal-Level Suppression

Collecting only the signals necessary for the stated analytical purpose, and suppressing steering angle and yaw rate where they are not required, removes the primary geometric reconstruction input. This requires explicit signal-selection policies at the collection layer, not post-hoc column dropping.

Temporal Perturbation and Aggregation

Introducing calibrated jitter into signal timestamps and aggregating high-frequency samples into event summaries disrupts the sequence alignment that map-matching depends on. The goal is to degrade the temporal resolution below the threshold at which road-network constraints become binding, while preserving the aggregate statistics needed for insurance scoring or fleet analytics.

Differential Privacy on Derived Features

Where raw signal streams must be retained for model training, applying differential privacy mechanisms to derived features (trip duration, hard-braking count, average cornering force) rather than sharing raw sequences limits re-identification risk without eliminating analytical utility. This is a materially different architecture from stripping GPS and exporting the remainder.

Auditing What Your Pipeline Actually Exposes

The practical question for engineering leaders is not whether their pipeline contains GPS fields. It is whether the signals their pipeline exports are sufficient inputs for a map-matching or dead-reckoning reconstruction. That audit requires running, or commissioning, a recovery attempt against a representative sample of exported data under the weakest plausible attacker assumption: road network access, no GPS, no prior knowledge of the driver.

If recovery succeeds at rates comparable to those documented by Wang et al. (Wang et al., HuggingFace 2026), the pipeline is not de-identified in any legally or technically meaningful sense. Documenting that finding, and the architectural response to it, is the kind of evidence that regulators and insurers will increasingly expect to see before a data-sharing arrangement is approved.

The signal layer has always been the privacy surface in vehicle telematics. The research now exists to demonstrate that at scale, and the burden of proof has shifted accordingly.

FAQs

Does stripping GPS from telematics data satisfy GDPR data minimisation requirements?

Not if the remaining signals are sufficient to reconstruct routes. GDPR data minimisation requires that data cannot reasonably be used to re-identify individuals. Research demonstrates that CAN signal sequences enable 81–88% route recovery from unseen drivers, which is a strong indicator that GPS-stripped exports do not meet that threshold without additional signal-level controls.

Which CAN signals carry the most re-identification risk?

Steering angle is the highest-risk signal because turns are geometrically distinctive and tightly constrained by road topology. Wheel speed combined with steering is sufficient to drive map-matching recovery to high accuracy. Yaw rate provides additional constraint. Pedal inputs carry lower geometric information but contribute to event-level fingerprinting when combined with the others.

Can usage-based insurance scoring work without retaining raw CAN sequences?

Yes. Most UBI scoring models consume derived features: hard-braking events, cornering force, time-of-day distribution, trip distance. These can be computed on-device or at the edge and transmitted as aggregates, with raw sequences discarded after feature extraction. That architecture retains scoring utility while materially reducing the signal data available for route reconstruction.

How should we assess whether our current data exports are reconstructible?

The most direct method is to run a map-matching recovery attempt against a representative sample of exported data using publicly available road network data and one of the established dead-reckoning or online filter approaches documented in the literature. If recovery rates on your sample approach the benchmarks reported in recent research, the export is not de-identified. That test should be part of any data-sharing due diligence process.

What does a defensible data minimisation posture look like for an open driving dataset release?

It requires signal-selection policies that suppress geometric signals not needed for the stated research purpose, temporal perturbation sufficient to degrade map-matching alignment, and a documented recovery audit before release. Publishing a dataset with GPS removed but full steering and speed sequences intact is not a defensible posture under current research conditions, and the regulatory expectation around demonstrable de-identification is tightening.

A team that understands you
With 20+ years of experience in the world's leading consultancy companies, implementing AI and ML projects in industry-specific contexts, we are ready to hear your challenges.
Subscribe to our newsletter for insights and updates on AI and industry trends.
By clicking "Sign me up", you agree to our Privacy Policy.
By clicking the Accept button, you are giving your consent to the use of cookies when accessing this website and utilizing our services. To learn more about how cookies are used and managed, please refer to our Privacy Policy and Cookies Declaration