CANTraj: Recovering Vehicle Trajectories from CAN Signals in Naturalistic Driving

CANTraj: Recovering Vehicle Trajectories from CAN Signals in Naturalistic Driving

Removing GPS coordinates from vehicle data sounds like a straightforward privacy safeguard. The authors of CANTraj argue that it is not enough. Modern cars expose speed, steering angle, pedal positions, yaw rate and other signals through the controller area network (CAN). When those signals are combined with a road map, they can reveal where a vehicle travelled.

The paper’s central contribution is not simply another route-inference attack. It is a larger dataset and a more controlled comparison of existing approaches. The authors release CANTraj, containing 3,940 drives from 454 drivers, paired with decoded CAN events and time-aligned GNSS trajectories. Its evaluation subset, CANTraj-Bench, contains 3,675 drives from 431 drivers, covering 1,380 hours and 85,919 kilometres. The benchmark evaluates methods on unseen drivers while making their information access explicit.

The privacy problem is in the combination of signals

Speed alone can reveal stops, distances and possible turns, but it leaves many candidate roads. Steering adds a different kind of information: it reflects the vehicle’s heading changes and therefore helps distinguish road geometries. A map-based tracker can use these signals to maintain plausible positions on a road network, even when the original GPS trace is absent.

The benchmark gives a method the starting position and, in some protocols, the initial heading. It then tests online and offline recovery over one-, five- and fifteen-minute windows. Missing outputs and failed searches count as failures rather than disappearing from the results. The authors report average displacement error, the fraction of time within 100 metres, route recovery, and failure measures that expose large excursions.

Maps and steering drive the biggest improvement

The strongest demonstrated result is the gap between map-free dead reckoning and map-constrained tracking. On five-minute test windows, nominal steering dead reckoning produced a median average displacement error of 921 metres. With speed and steering supplied to an online map particle filter, the median fell to 43 metres, and 87% of routes were recovered under the benchmark’s route criterion. An offline beam search reached 37 metres and recovered 88% of routes.

These results are not evidence that every drive can be reconstructed precisely. They are results for the paper’s defined task: a known starting position, a restricted road network and the tested CAN signals. Still, they show why deleting GPS fields does not necessarily remove location leakage.

Steering was much more valuable than the additional pedal signals tested by the authors. In a controlled signal study using the same learner, decoder and drives, adding steering to speed reduced the median error from 339 metres to 49 metres. Adding continuous accelerator-pedal data produced changes ranging from small improvements to small degradations, with confidence intervals often including no difference. The paper therefore identifies steering, rather than pedal data, as the most important tested signal for this task.

Initial heading and offline access matter differently

With steering available, withholding the starting heading increased the online filter’s five-minute median error from 43 to 138 metres. Route recovery fell from 87% to 79%. The offline beam was more resilient: its median changed from 37 to 41 metres without the heading, although route recovery still fell from 88% to 81%.

The authors interpret this difference as a consequence of access to future observations. An offline method can compare later turns against competing initial directions, while an online filter must maintain uncertainty until the route provides enough evidence. Over longer horizons, the offline beam also avoided some of the online filter’s largest failures.

Existing attacks do not all generalize equally

The unified evaluation exposes substantial variation among published methods. OPD-II, which combines speed and steering with map matching, reached a 42-metre median error on five-minute test windows and recovered 79% of routes, but its information access and delayed output differ from the benchmark’s standard filter. MacroTracking reached 84 metres. Speed-focused methods were less consistent, and CAN-Trace produced very few matches in the tested configuration.

These comparisons require care. The authors re-implemented six attacks, but they note that checking against released samples does not reproduce every original experiment. Some methods also require different priors, signal availability or map access. The paper’s point is therefore less about declaring one universal winner than about showing why published scores cannot be compared without aligning assumptions and failure rules.

Why failures remain important

Different methods failed on different windows. Selecting the best of ten tested map-based methods recovered 98% of five-minute test routes, compared with 88% for the fixed beam. That is a diagnostic upper bound for this method pool, not a deployable system. The authors also found that common failures increased on windows with many turns, and that steering-based tracking often began to drift shortly after a turn.

Signal degradation offered limited protection against the tested attacks. Quantizing steering to 10 degrees or speed to 1 metre per second changed route recovery only modestly. More aggressive speed quantization had a larger effect on steering-based attacks, but the authors caution that these were fixed, untuned attacks. An adaptive adversary could respond differently.

Practical implications and limitations

For data publishers, the practical lesson is straightforward: treating CAN data as harmless after GPS removal is unsafe. A privacy review should consider combinations of signals, road-map access, starting information and attacker computation. Quantization alone may not be a reliable safeguard, and any protection should be tested against adapted attacks while measuring its impact on the intended automotive application.

For localization engineers, CANTraj suggests that a road map and steering signal can compensate for substantial dead-reckoning drift. The benchmark and released implementations may also help evaluate GPS-denied tracking under controlled assumptions.

The authors emphasize important boundaries. North American roads and supported vehicle platforms dominate the corpus. The task usually assumes one known starting position, and area-based variants are not unrestricted unknown-start tracking. Vehicle, regional and road-density effects are not fully separable, while smaller signal subsets limit conclusions about pedals and yaw rate. Results from a small controlled supplement also changed the method ranking. CANTraj broadens the evidence for CAN-based trajectory recovery, but it does not establish universal performance—or a universal privacy threshold.


Sources

AI-assisted research summary. Check the original paper for complete methods, results, and limitations.

Avatar photo

Tech News

Jeff is a passionate blog writer who shares clear, practical insights on technology, digital trends and AI industries. With a focus on simplicity and real-world experience, his writing helps readers understand complex topics in an accessible way. Through his blog, Jeff aims to inform, educate, and inspire curiosity, always valuing clarity, reliability, and continuous learning.

↑ ↓ navigate Enter open Esc close