CANTraj: Recovering vehicle trajectories from naturalistic driving CAN signals
CANTraj: Recovering Vehicle Trajectories from CAN Signals in Naturalistic Driving
The authors release the CANTraj dataset and benchmark, uniformly reproducing six CAN trajectory recovery attacks; the authors state that, given a known start point and heading, the online map filter using speed plus steering achieves an 87% route recovery rate on five-minute routes.
- Open driving data and usage-based insurance often claim to protect location privacy by removing GPS coordinates, but CAN bus signals still carry location information. Existing route inference attacks have inconsistent assumptions and metrics and small datasets, making results difficult to compare and their real effectiveness uncertain.
- The authors release the CANTraj dataset (3,940 drives, 454 drivers, CAN signals paired with time-aligned GNSS trajectories) and the CANTraj-Bench benchmark, reproducing six published attacks under a unified information access protocol, paired comparisons, and metrics that count failures, comparing them with baselines such as dead reckoning, map particle filtering, and offline beam, and additionally testing a Qwen3-14B agent that selects among candidate routes.
- Given a known start point and heading, the online map filter using speed plus steering achieves a median ADE of 43 m and 87% route recovery on the test driver five-minute window, while offline beam achieves 88% recovery; map-free dead reckoning has a median ADE of 921 m. Steering is the signal with the largest gain; after quantizing steering to 10° or speed to 1 m/s, the tested map attacks remain largely effective.
- The authors note that the data are predominantly North American roads and vehicles, that region and driver factors are confounded, and that method rankings change in the controlled supplementary experiments; the task mainly measures recovery after one location is known, and the region variant is not unrestricted unknown-start tracking; four of the six attacks lack verification against the original authors' samples; continuous pedal and CAN yaw rate are available only on a smaller subset.
- Threat model
- The attacker obtains CAN bus signals with GNSS coordinates removed (speed, steering angle, pedals, etc.), knows the start location (some protocols include initial heading), has access to the OpenStreetMap road network, and aims to recover the vehicle trajectory from naturalistic driving data; the victim systems are open driving datasets and usage-based insurance programs that share or collect vehicle data.
- Evaluated models
- Learned GRULearned physics residualSignal-set GRUQwen3-14B
- Benchmarks
- CANTrajCANTraj-BenchOpenStreetMap (Geofabrik extracts)
- Metrics
- ADEc100route recovered (path recall@50 ≥ 0.8)tail-lostterminal excursion
The paper releases the CANTraj naturalistic driving dataset and the CANTraj-Bench benchmark, showing that after GPS coordinates are removed from open driving data and Usage-Based Insurance, vehicle trajectories can still be recovered from CAN signals. The dataset pairs decoded CAN events with time-aligned GNSS trajectories, containing 3,940 drives and 454 drivers; the benchmark covers 3,675 drives, 1,380 hours, and 85,919 kilometers, comparing six reproduced attacks on 776 evaluation drives from 94 unseen drivers. Data and code are open-sourced at huggingface.co/datasets/HenryYHW/CANTraj and github.com/WangYuHang-cmd/GPS--CAN.
Paper analysis
What problem does this paper attempt to solve?
CAN data with GPS removed still leaks trajectories, but existing attacks have inconsistent assumptions and metrics and are difficult to compare; the paper provides a unified dataset and benchmark.
Core question: How accurately can the driving trajectory be recovered from vehicle CAN signals with GNSS coordinates removed, and how can various route inference attacks be compared fairly under a unified protocol.
- Scenario and importance: Modern cars broadcast wheel speed, steering angle, pedal position, and yaw rate over the CAN bus, which can be obtained via the OBD port; usage-based insurance, fleet telematics, and open driving datasets all collect these signals. The authors argue that removing GNSS coordinates appears to protect location privacy, but the remaining vehicle signals can still leak the vehicle's whereabouts.
- Shortcomings of existing methods: The authors point out that existing route inference attacks rely on inconsistent prior assumptions (known start point and heading, contained region, or additional traffic information) and evaluation metrics (destination ranking, trajectory overlap, union coverage of candidate routes), and are tested only on small-scale, restricted datasets (e.g., SPy uses 58 trips, OPD-II uses 59 trajectories), making results difficult to compare with each other and their real effectiveness uncertain.
- What the paper proposes: Releases the CANTraj naturalistic driving dataset (3,940 drives, 454 drivers, decoded CAN events paired with time-aligned GNSS trajectories) and the CANTraj-Bench benchmark (3,675 drives, 431 drivers, 1,380 hours, 85,919 km), reproduces six published attacks and compares them with baseline methods under an explicit information access protocol, paired comparisons, and metrics that count failures, and additionally evaluates an LLM tool that selects among candidate routes.
- Threat model:
- Attacker goal: Recover the vehicle's trajectory or route within a given time window from CAN signals.
- Attacker knowledge: The start location is known (some protocols give initial heading; the region variant gives a 5 km square centered on the start point); OpenStreetMap road network is accessible (the protocol restricts it to a disk region centered on the start point).
- Attacker capability: Obtains the signal columns permitted by the protocol (speed, steering angle, pedals, etc.); static vehicle parameters come from that drive's own records; GPS position, speed, heading, accuracy fields, and absolute time are all excluded.
- Victim systems: Open driving releases and usage-based insurance programs that share or collect vehicle data.
What related research exists?
Related work falls into four groups: route inference attacks based on speed, events and steering; vehicle localization and learned motion models; location protection methods; and public driving datasets.
Route inference from vehicle signals
- Speed-based attacks: Early work searches for destinations or paths consistent with stops, travel distance, and feasible turns [3], [4], [14]; Wahlström et al. [6] use map particle filtering with lateral force and speed limit constraints; Zhou et al. [5], [15] combine hidden Markov search with speed sequence matching. The paper notes that these studies include GPS-derived speed trajectories and do not isolate the effect of signal source or decoder on a common vehicle population.
- Event-based attacks: BBLT and its journal extension [8], [9] infer maneuvers from brake signals; CAN-Trace [10] uses speed and accelerator pedal events; Waltereit et al. [7] match distance and turn sequences, but evaluate with GPS-derived events with injected error.
- Steering-based attacks: SPy [11] matches steering profiles to road curvature; MacroTracking and OPD-II [12], [13] combine speed, steering, and map correction. The paper notes that their evaluations answer different questions (proximity to reference path, trajectory similarity, multi-candidate coverage), and this paper retains route and location metrics and counts failed searches.
Vehicle localization and learned motion models
- Map matching [16], sparse trajectory recovery [17], arc-length map matching with known lane centerlines [18]; learned navigation includes AI-IMU [19], FormerReckoning, RoTrax [20], [21], S-TRAVERSE [22], GHNet [23], etc. The paper notes that the additional sensors or GNSS observations of these methods must be distinguished from CAN-only input, and this paper explicitly distinguishes them with signal source labels.
Location protection and evaluation data
- RAoPT and VehiTrack [26], [27] reconstruct trajectories from perturbed locations; TransProtect, Olympus, and adversarial mobility prediction [27], [28], [29] learn representations that balance authorized tasks and sensitive inference. The paper states that these works motivate testing adaptive attacks and reporting both accuracy and computational cost.
- Public datasets: comma2k19 [31] provides CAN and GNSS for highway commuting; DriveDNA [32] provides naturalistic driving data without GPS coordinates. The paper states that CANTraj adds CAN-GNSS paired data across hundreds of drivers and vehicle platforms and a unified evaluation of six attacks.
Baselines and benchmark: The six attacks reproduced by the paper (MacroTracking, OPD-II, Elastic Pathing, Dewri et al., CAN-Trace, SPy) are the main comparison objects, and baseline methods include dead reckoning, map particle filtering, and offline beam search; evaluation is on the test and few-shot splits of CANTraj-Bench.
The authors position this paper as: a cross-method comparison of six attacks under a unified protocol, paired comparisons, and metrics that count failures, isolating the contribution of each signal.
How does the paper address this problem?
Release a paired CAN-GNSS dataset and benchmark, uniformly reproducing six attacks under eight information access protocols, with paired comparisons using metrics that count failures.
Overall approach: build a naturalistic driving dataset CANTraj pairing CAN signals with GNSS trajectories, establish the CANTraj-Bench benchmark on it, uniformly define the task, priors, signal sets, metrics, and failure scoring rules, reproduce six published attacks, and compare them with baseline methods.
Task and protocol
- Task: Given the location p0 at time t0, each method estimates the trajectory within [t0, t0+D] using the permitted vehicle signals, outputting positions at GNSS reference timestamps (at most one per second); methods that output only routes are localized by integrating their own speed.
- Eight protocols: Three priors—online/offline (O/F), with/without map (N/M), with/without initial heading (H+/H−)—combined independently, denoted O/F-N/M-H+/H−; this paper reports five of them (map-free only O-N-H+, map-based all four).
- Map access: Map methods can only see the OSM disk centered on p0 with radius 5 km + 40 m/s·D (at most 60 km); OPD-II's Meili matcher searches its Valhalla region tiles (within 200 m of the dead-reckoned trajectory) and does not clip to this disk.
Signals and data
- Signal source labels: C0 read from the vehicle bus (speed, steering angle, pedals, yaw rate), C1 exported by in-vehicle software from CAN, I0 recorder raw IMU, F driving software localizer output (used GPS in this data version, for reference only, never counted as CAN-only).
- Data scale: CANTraj totals 3,940 drives and 454 drivers (1,659 hours of GNSS trajectories); CANTraj-Bench uses 3,675 of these drives, 431 drivers, 266 vehicle platforms (27 brands), 1,380 hours, 85,919 km; drivers are split into disjoint training, validation, test, and few-shot folds, and results are reported on test (43 drivers, 302 drives) and few-shot (51 drivers, 474 drives).
- Quality control: Raw logs are re-decoded into timestamped message tables (full logs 100 Hz, compressed logs 10 Hz), and the first and last two minutes of each drive are trimmed; GNSS and CAN speed receiver delays are estimated by cross-correlation on the training fold (modem 0.7 s, u-blox 0.2 s); map tiles are rebuilt through an audit, with a median distance of 2.6 m from the start point to the nearest drivable road.
Methods
- Six reproduced attacks: MacroTracking (speed+steering, online), OPD-II (speed+steering, output delayed up to 3 s), Elastic Pathing (speed only, offline), Dewri et al. (speed only, offline), CAN-Trace (speed+accelerator pedal, offline), SPy (steering only, offline); implementations are verified against the original authors' sample data (MacroTracking reproduces within a median deviation of 7.1 m on their sample trajectory, OPD-II within a median deviation of at most 0.33 m from the original script on 8 sample logs).
- Baseline methods: Map-free baselines (hold-start, straight line, steering dead reckoning, start prior, route retrieval), learned motion models (causal GRU and physics residual model, 64 units, 10 Hz, three training seeds), map particle filtering (2,000 particles), and offline beam search (B=200, K=10, σh=15°, σturn=20°).
- Signal combination study: With fixed learner, decoder, and drive, compare six signal sets S1 (speed), S2c/S2b (accelerator pedal continuous value/pressed flag), S3 (speed+steering), S4c/S4b (S3 plus pedal or flag).
Metrics and statistics
- Metrics: Average displacement error ADE, coverage c100 (proportion of seconds with error ≤100 m), 50 m path recall/precision, route recovery (path recall@50 ≥ 0.8), and two failure definitions (tail-lost and terminal excursion).
- Failure scoring: Expected output that is missing (timeout, no match, error) is counted as failure, with ADE counted as +∞ and coverage and route flags counted as zero; means are reported only on episodes where all compared methods have output.
- Statistics: Confidence intervals are percentile bootstrap by driver resampling (1,000 resamples); comparisons between methods are paired on common episodes; five comparisons are primary and the rest exploratory; all statistics for learned methods are the mean of three training seeds.
What experiments did the paper conduct?
The main evaluation covers 776 drives and 94 unseen drivers; steering plus map provides the largest gain, with up to 88% five-minute route recovery, and attacks remain largely effective after signal quantization.
Experimental setup
- Methods tested: Six reproduced attacks (MacroTracking, OPD-II, Elastic Pathing, Dewri et al., CAN-Trace, SPy) plus baseline methods (hold-start, straight line, three steering dead reckoning variants, start prior, route retrieval, learned GRU, learned physics residual, map particle filtering, offline beam); additionally evaluates a Qwen3-14B agent selecting among candidate routes.
- Data: CANTraj-Bench test fold (43 drivers, 302 drives) and few-shot fold (51 drivers, 474 drives); the main evaluation comprises 1,917 test and 1,963 few-shot five-minute episodes (1,959 under the map protocol), totaling 776 drives and 94 drivers.
- Metrics: Median ADE (with 95% driver bootstrap CI), c100, route recovery rate (path recall@50 ≥ 0.8); failures are counted in the denominator.
- Protocol: Unless stated otherwise, start point and heading are both given; the region variant gives a 5 km square centered on the start point.
- Signal study: 407 five-minute windows, 31 evaluated drivers (with continuous accelerator signal), three seeds.
- Repetition: Learned methods use three training seeds; the LLM agent uses three seeds.
Main results
Five-minute window results (per Table II, ADE is median, in m):
Swipe horizontally to read the table
Method Test ADE Test route Few-shot ADE Few-shot route Steering dead reckoning (no map) 921 0.04 1,669 0.02 Learned GRU (no map) 543 0.07 393 0.05 Map PF, speed only 286 0.56 618 0.48 Map PF, steering 43 0.87 27 0.90 Map beam, steering 37 0.88 21 0.92 OPD-II (online readout) 42 0.79 24 0.78 MacroTracking 84 0.62 52 0.65 Elastic Pathing (no heading) 129 0.53 177 0.50 - The authors state that map and steering provide the largest gains: with known start point and heading, speed plus steering raises online map filtering's five-minute route recovery from 56% (speed only) to 87%, with a median ADE of 43 m, whereas map-free dead reckoning is 921 m.
- Offline beam is less sensitive to missing initial heading than online filtering: beam's median is 37 m with heading and 41 m without heading, but route recovery drops from 88% to 81%; filtering rises from 43 m to 138 m.
- Among published methods, the paired median difference between OPD-II and filtering includes zero, but the paired mean is 116 m higher; SPy recovers only 20% of test routes with known start point and heading; CAN-Trace matches only 2 of 245 test windows with valid pedal, and the authors note this evaluates this paper's event detector and region adaptation, not CAN-Trace's original task.
Signal contribution and controlled checks
- Signal shuffling test: Shuffling steering on the validation block makes the fixed learner's loss 3.4 to 3.8 times worse; shuffling continuous pedal changes results by −0.5% to +9.2% relative to speed only, and the authors state the pedal effect is small and inconsistent across seeds relative to speed only.
- Controlled signal study: With fixed learner, decoder, and drive, adding steering changes the paired median error by −120 to −234 m (median from 339 to 49 m); the change intervals for adding continuous pedal all include zero; the pressed flag makes speed-only results 5.2 m worse.
- Controlled supplementary drives: 9 methods are rerun on 46 five-minute windows (12 other drivers, one vehicle, one metropolitan area); MacroTracking has the lowest median (19 m, 85% recovery), OPD-II, beam, and filtering are 25, 46, and 75 m respectively, with overlapping intervals; the authors state this suggests rankings do not generalize to arbitrary vehicles or regions.
Failures, cost, and robustness (RQ3)
- Failure complementarity: Taking the best of ten map methods per window recovers 98% of test routes (beam is 88%), and 2.3% of windows are not recovered by any method; common failures rise from 0.1% of windows with at most one turn to 8.7% of windows with seven or more turns.
- Computational cost: Median time on five-minute windows is 0.03 s for steering dead reckoning, 0.08 s for MacroTracking, 0.08 s for OPD-II, 0.24 s for beam, and 0.57 s for filtering; SPy's region search has a median of 4.93 s.
- Signal degradation: Quantizing steering to 10°, offsetting by 5°, or keeping 1 Hz changes the route recovery of the two map attacks by at most 1.3 percentage points; quantizing speed to 1 m/s shows no measurable change; the authors state that even at the strictest level the attacks still recover 82.6% or more of routes, and note that these are sensitivities to fixed attacks and do not set a protection threshold.
LLM candidate selection
- Setup: The first qualifying five-minute window of each evaluated drive (268 test, 391 few-shot); the baseline is beam (A), and the supervised logistic selector (B) and Qwen3-14B agent (C) select among beam, filtering's best lineage, and Elastic Pathing; the extended agent (C+) has ten reconstructor presets.
- Results: On test drivers, route recovery is 69% for beam, 75% for the selector, 75% for the base agent, and 78% for the extended agent (per Table IV); the difference between agent and selector is 0.1 percentage points [−1.6, 3.4], and the authors state the base agent has no clear advantage over the selector, while its serial cost is 36.4 s versus 1.1 s.
- Exceptions: The extended agent is 4.1 percentage points higher than its selector on test [1.3, 7.6], but this does not carry over to the few-shot group (81% versus 82%); oracle-style selection can recover 83% (three tools) to 87% (all presets), and the authors state there is still room for improvement in selection, and the results involve only one model and two configurations.
Other ablations and analyses
- Duration (same cohort of 205 test drives): steering beam's median ADE at 1, 5, and 15 minutes is 17, 44, and 303 m, and filtering is 19, 56, and 672 m (per Figure 5).
- Full drives: steering beam's median ADE is 1,179 m on test and 94 m on few-shot; 18/282 test and 36/454 few-shot drives end beyond the 60 km map disk.
- Road density: speed-only filtering's five-minute median from sparse to dense is 73, 677, and 843 m, and steering filtering is 29, 39, and 36 m; the authors note that density is confounded with region, driver, etc., and this is association, not causation.
- Fixing an OPD-II code defect gives a paired mean difference of −21.2 m [−54.2, 12.6] on fifteen-minute test; fixing a MacroTracking defect leaves the paired median unchanged (0.0 m) but the paired mean is −301 m.
What are some points that can be further explored?
The authors note that the data are predominantly North American, the task is limited to a known start point, and four attacks lack verification; the experiments cover 776 drives and one LLM.
Limitations and future directions noted by the authors
- Data coverage: North American roads and vehicles supported by the logging system predominate; drivers are independent across splits, but vehicle platforms and regions are not; the comparison of road density and number of turns is correlational rather than causal, because road conditions, drivers and vehicles vary together (Section VII-b).
- Generalizability of rankings: Method rankings change in controlled supplementary experiments, and the authors state that rankings should not be regarded as universal (Section VII-b).
- Task scope: The task mainly measures recovery after one location is known; the region variant receives a square centered on that location and is not a test of unrestricted tracking from an unknown start point, and many routes leave that square (Section VII-b).
- Implementation checks: Checks of MacroTracking and OPD-II support implementation correctness, but not every experiment in the original papers was reproduced; the other four attacks lack the same verification evidence (Section VII-b).
- Failure is not a privacy guarantee: Failure of these methods does not provide a privacy guarantee; the per-window oracle is a diagnostic result, not a deployable selector (Section VII-b).
- Directions for protection evaluation: The authors state that evaluating protection methods also requires measuring their impact on intended downstream tasks and testing attackers adapted to the modified signals (Section VII-a).
Scope of experiments
- The methods tested are six reproduced attacks plus about eleven baseline and learned variants; the main evaluation covers 776 driving trips and 94 drivers (test and few-shot combined).
- The signal degradation evaluation includes steering quantization/offset/downsampling and speed quantization/downsampling, one seed each; attacks were not retuned or retrained on degraded signals.
- LLM candidate selection evaluates only one model, Qwen3-14B, in two configurations (base and extended), with three seeds.
- Continuous accelerator pedal and CAN yaw rate are available only on a smaller subset (pedal: 48 of 302 test driving trips).
- The paper does not report an evaluation of attackers adapted to degraded signals (retrained or recalibrated); the authors explicitly state that the robustness results only bound the publisher's gains against these fixed attacks.
Summarize the main content of the paper
The CANTraj benchmark uniformly compares six CAN trajectory recovery attacks: steering plus map provides the largest gain, with up to 88% five-minute route recovery, and removing GPS is insufficient to protect location privacy.
A benchmark paper on vehicle data location privacy: the authors release the CANTraj dataset (3,940 drives, 454 drivers, CAN signals paired with time-aligned GNSS trajectories) and CANTraj-Bench, reproducing six published route inference attacks under a unified information access protocol, paired comparisons, and metrics that count failures, answering the question "how much trajectory can CAN signals still leak after GPS is removed?"
- Method highlights: The task is defined as estimating the trajectory within a time window from CAN signals given a known start point (some protocols include initial heading); eight protocol combinations of online/offline, with/without map, with/without heading; reproduces MacroTracking, OPD-II, Elastic Pathing, Dewri et al., CAN-Trace, SPy, and compares with dead reckoning, map particle filtering, offline beam, and learned motion models; failed searches are counted in the metrics.
- Main results: With known start point and heading, online map filtering with speed plus steering achieves a median ADE of 43 m and 87% route recovery on five-minute windows for test drivers, and offline beam achieves 37 m and 88%; map-free steering dead reckoning has a median ADE of 921 m (Table II).
- Signal contribution: In the controlled signal study, steering is the signal with the largest gain; adding steering reduces the paired median error from 339 to 49 m; continuous accelerator pedal contributes little and inconsistently (Section VI-A1).
- Robustness: After quantizing steering to 10° or speed to 1 m/s, the route recovery of the tested map attacks changes by at most 1.3 percentage points, and the authors state that even at the strictest level more than 82.6% of routes are still recovered (Section VI-C1).
- LLM selection: The Qwen3-14B extended agent recovers 78% of routes on test drivers, with no stable advantage over the supervised selector, at a serial cost of 36.4 s versus 1.1 s (Table IV).
- Authors' conclusions: Removing GPS coordinates or initial heading does not remove route information; the authors recommend matching methods to available signals and information, and believe that evaluating protection methods requires measuring both downstream utility and adaptive attacks; the authors also note that the data are predominantly North American, the task is limited to a known start point, and four attacks lack verification against the original authors.