PaperlayerView exact revision on arXiv

Source-linked AI summary

TEMPEST: Temporal Embeddings for Scalable Driver Identification via Angular Margin Learning

Kyle Musgrove, Dylan B. Lewis, Sarah Powers, Emma J. Reid, Hector Santos-Villalobos

arXiv:2610.06855v1cs.LGcs.CV

TL;DR

Driver-identification embeddings must remain discriminative as the enrolled driver pool grows, while existing triplet-loss approaches provide local constraints and face temporal-evaluation challenges. TEMPEST uses a TCN-inspired embedding model with ArcFace training; under identical temporal evaluation, replacing triplet loss with ArcFace yields a 58.4 pp Rank-1 gain over DriverRep.

  • Problem

    Embedding methods support dynamic driver enrollment, but triplet and contrastive objectives impose local pairwise constraints, and prior results can be affected by temporal or session-specific leakage.

    Source: §4.2 · p. 3

  • Method

    TEMPEST is a Temporal Convolutional Network-inspired embedding model that replaces triplet-loss training with an additive angular margin (ArcFace) objective.

    Source: §1 · p. 1

  • Results

    58.4 pp Rank-1 accuracy gain over DriverRep (33.35% →91.71%) under identical temporal evaluation conditions.

    Source: §6 · p. 8

  • Takeaways & Limitations

    Cosine-similarity matching supports open enrollment by updating the embedding gallery without retraining or classifier refitting.

    Source: §1 · p. 2

    Dataset scope and evaluation coverage are limited: DriverID has one session per driver, while KIA has four sessions but only 10 drivers.

    Source: §5.3 · p. 8

Abstract

from arXiv · show

Scalable driver identification requires embedding models that maintain discriminative performance as fleet size grows, yet existing triplet-loss formulations degrade rapidly with driver pool size and overfit to session-specific patterns under rigorous temporal evaluation. We introduce TEMPEST, a Temporal Convolutional Network embedding model trained with an additive angular margin (ArcFace) loss that enforces global class-level separation in a normalized angular space. TEMPEST maps 60-second multimodal driving windows to compact 96-dimensional embeddings, supporting truly dynamic enrollment without any retraining or classifier refitting. Under rigorous temporal evaluation on a 45-driver dataset, TEMPEST achieves 91.71% Rank-1 accuracy, outperforming the best classical model by 17.9 pp and the strongest triplet-loss baseline by 58.4 pp. TEMPEST degrades by only 4.3 pp when growing the subject pool from 10 to 45 drivers, compared to 22 pp and 32.5 pp for supervised and unsupervised triplet-loss baselines, and its cross-session advantage is corroborated on the public KIA Soul dataset, where it outperforms the best classical model by 7.3 pp within-session and 14.3 pp cross-session. With 720K parameters, a 2.80 MB footprint, and 50-epoch convergence, TEMPEST establishes a rigorous, reproducible baseline for scalable behavioral driver biometric identification.

1. Introduction

Behavioral driving signals can identify drivers, but existing embedding objectives and evaluation practices leave scalability and temporal generalization unresolved. TEMPEST combines ArcFace training with embedding-based identification and is reported to retain accuracy as the driver pool grows.

1. Introduction

  • Driving behavior, including throttle use, braking, steering, and physiological responses, can jointly characterize individual drivers.The introduction frames this behavioral biometric as useful for continuous driver identification during operation.Source: §1 · p. 1
  • The paper introduces a Temporal Convolutional Network-inspired embedding model that replaces triplet-loss training with an additive angular margin (ArcFace) objective.The objective is adapted from face recognition.Source: §1 · p. 1
  • 91.71% Rank-1 accuracy on 45 drivers under rigorous temporal evaluation exceeded the best classical baseline by 17.4 pp and the strongest triplet-loss baseline by 33.8 pp.The contribution describes an additive angular margin loss that enforces global class-level separation in normalized angular space.Source: §1 · p. 2
  • TEMPEST lost only 4.3 pp accuracy when the pool grew from 10 to 45 drivers, versus 22 pp for supervised and 32.5 pp for unsupervised triplet-loss baselines.The paper presents this as evidence of more scalable angular-margin training for behavioral biometric identification.Source: §1 · p. 2
  • Random cross-validation inflated accuracy by up to 80 pp relative to chronological and cross-session evaluation across the two-dataset benchmarking study.The study reports this difference to quantify temporal data leakage and establish reproducible baselines.Source: §1 · p. 2
  • TEMPEST supports open enrollment by matching embeddings with cosine similarity, allowing gallery updates without retraining or classifier refitting.The reported scheme achieves 91.71% Rank-1 accuracy without a pool-dependent classifier.Source: §1 · p. 2

2. Related Work

Driver identification developed from classical models to deep time-series and embedding methods, which can accommodate changing driver pools. Prior embedding work relies on local triplet constraints, motivating the paper's application of angular-margin learning to driving signals.

2. Related Work

  • Early classical models included Hidden Markov Models, Gaussian Mixture Models, and decision trees, but assumed fixed subject pools and required retraining for new drivers.Later CAN and OBD-II signals and public datasets such as HCRL KIA Soul enabled systematic comparisons,.Source: §2 · p. 2
  • Deep models applied convolutional and recurrent architectures to multivariate time series [12], while embeddings such as Driver2Vec and Drive2Vec used triplet or contrastive objectives [13] [31] [14].Embeddings allow gallery identities to be added or removed on demand, but reported accuracy often falls as driver counts rise or evaluation crosses sessions,,[16].Source: §2 · p. 2
  • Driver2Vec [31] and Drive2Vec [14] use triplet-loss formulations that constrain local pairs rather than global class-level structure.The paper contrasts these objectives with ArcFace's fixed angular margin between class centers on a hypersphere.Source: §2 · p. 2
  • The authors describe TEMPEST as the first application of angular-margin learning to multimodal driving-signal behavioral identification.The approach brings an objective previously used in facial recognition to this setting.Source: §2 · p. 2

3. Datasets

The evaluation uses two complementary datasets that differ in size, session coverage, vehicle type, and sensor configuration. DriverID provides a larger real-world cohort, while HCRL KIA Soul supplies repeated sessions for cross-session testing.

3.1. ORNL Driver Identification Dataset

  • The ORNL source cohort comprised 50 volunteer truck drivers assigned to three groups by cyberattack warning, each completing a 40–60-minute fixed-route drive in the same instrumented vehicle.CAN, GPS/inertial, and heart-rate streams were timestamp-aligned into a multimodal time series with 816 raw features.Source: §3.1 · p. 3
  • The processed DriverID dataset contains 45 drivers and 19 retained behavioral features after filtering, resampling to 1 Hz, and removing vehicle-state signals.The three warning groups showed no systematic biometric differences in the reported analysis.Source: §3.1 · p. 3

3.2. HCRL KIA Soul Dataset

  • HCRL KIA Soul contains 10 drivers with four labeled sessions each, enabling leave-one-drive-out evaluation that holds out an entire session.The dataset has 54 features, of which 15 were considered important by prior work [18].Source: §3.2 · p. 3

4. Modeling Approach

The section contrasts fixed-feature classification and triplet-based embeddings with TEMPEST’s class-level angular-margin training. It describes the encoder and the inference configurations used to match its embeddings to driver identities.

4. Modeling Approach

  • Embedding-based methods represent drivers in a space that decouples identity modeling from the enrolled-driver count, while classical classifiers use hand-engineered features and assume a fixed, closed driver set.The section notes that classical classifiers can be vulnerable to temporal leakage and session-specific artifacts under chronological or cross-session evaluation.Source: §4 · p. 3

4.1. DriverRep Baseline

  • DriverRep maps fixed-length driving windows to 128-dimensional embeddings with a dilated-causal TCN and triplet loss, then matches identities by similarity or downstream classifiers.Its triplet objective encourages same-driver embeddings to be more similar than different-driver embeddings.Source: §4.1 · p. 3

4.2. TEMPEST

  • TEMPEST replaces triplet loss with ArcFace’s additive angular margin, imposing global class-level separation in normalized angular space without explicit sample mining.The section motivates this change by noting triplet loss’s reliance on informative triplet selection and lack of explicit compact class-level structure.Source: §4.2 · pp. 3–4

4.2.1 Encoder Architecture

  • A single block of five exponentially dilated causal-convolution layers spans the 60-second window, followed by adaptive average pooling and projection to a 96-dimensional embedding.The 96-dimensional representation replaces DriverRep’s 128-dimensional embedding; preliminary experiments found only marginal accuracy gains from the larger embedding at increased model size.Source: §4.2.1 · p. 4, Figure 2 · p. 4

4.2.2 ArcFace Loss Formulation

  • ArcFace adds margin m to the target angle between a normalized window embedding and its true class center, promoting tighter same-driver clusters and wider inter-class boundaries.The selected hyperparameters are m = 0.21 rad and s = 20.25, tuned on a held-out split from the training portion without test data.Source: §4.2.2 · p. 4
  • TEMPEST-Raw uses cosine-distance nearest-neighbor retrieval against the enrolled-driver gallery, while TEMPEST-SVM applies an RBF-kernel SVM to the same frozen embeddings.The two configurations share the encoder; the SVM uses C = 1.0.Source: §4.2.2 · p. 4

5. Results and Discussion

The evaluation uses repeated runs for deep-learning models and compares triplet-loss training configurations on a fixed TCN encoder. It defines chronological and cross-session protocols alongside random cross-validation to illustrate leakage effects.

5. Results and Discussion

  • DriverID deep-learning experiments and MLP were run 10 times with different seeds, with mean accuracy and standard deviation reported for DriverID unless otherwise noted.KIA results report best accuracy; classical models were run once because of negligible variance.Source: §5 · p. 5
  • TEMPEST, supervised triplet loss, and original unsupervised DriverRep form a controlled ablation over training objectives on a fixed TCN encoder.The supervised variant samples anchor, positive, and negative examples using ground-truth driver labels.Source: §5 · p. 5
  • The leakage-aware protocols are a within-session split using the first 80% of each drive for training and a cross-session leave-one-drive-out split holding out entire sessions.Random cross-validation is also reported to illustrate data-leakage effects.Source: §5 · p. 5

5.1. Evaluation Protocol Validation and Baseline Reproduction

Strict temporal and cross-session evaluations exposed differences that random splits concealed. TEMPEST also combined leading accuracy with a compact footprint and faster convergence than triplet-loss baselines.

5.1. Evaluation Protocol Validation and Baseline Reproduction

  • 81.50% average accuracy reproduced DriverRep’s reported 81.70% under its original protocol, but strict temporal evaluation with session-specific features removed reduced it to approximately 25% on the KIA 10-driver task.Source: §5.1 · p. 5
  • Random 10-fold cross-validation yielded ≈99% across models, but chronological DriverID evaluation separated TEMPEST at 91.21% from DriverRep at 33.35%; KIA cross-session scores were 40.93% vs. 17.36%.Table 1 reports results from the models’ best performance.Source: Table 1 · p. 5, §5.1 · p. 5

5.1.1 Efficiency–Accuracy Trade-offs

  • 91.71% Raw and 91.21% SVM Rank-1 accuracy led at 2.80 MB; similarly sized DriverRep and Sup-TL scored 33.35% and 57.39%, respectively.MLP reached 73.89% at 574 KB and XGB 70.76% at 16.6 MB; Random Forest and kNN used 159 MB and 88 MB for far lower accuracy.Source: Figure 3 · p. 5, §5.1.1 · p. 5

5.1.2 Supervised Triplet Loss vs. Unsupervised Triplet Loss

  • TEMPEST reached 81% within 5 epochs and 91% by epoch 50, while Sup-TL and DriverRep plateaued at 57% and 33%, respectively, after 400 epochs.ArcFace provides gradients for all 45 driver classes at each step, whereas triplet losses update only the anchor and negative driver identities per triplet.Source: Figure 4 · p. 6, §5.1.2 · p. 6

5.2. Deep Learning Model Analysis

In chronological evaluation, TEMPEST maintains strong identification performance as the driver pool grows, forms well-separated embeddings, and retains an advantage over triplet-loss models in cross-session KIA evaluation.

5.2.1 Pool Size Scaling

  • TEMPEST stays above 90% accuracy at every pool size, degrading 4.3 pp (σ = 0.70 at 45 drivers), versus DriverRep 32.5 pp and Sup-TL 22 pp.Figure 5 reports mean Rank-1 SVM test accuracy over 10 trials, with bars showing ±1σ.Source: Figure 5 · p. 6, §5.2.1 · p. 6

5.2.2 Rank-Based Evaluation

  • TEMPEST reaches 96% by Rank-3 and 99% by Rank-5; Sup-TL requires 5 ranks to reach 90%, while DriverRep does not surpass 90% until Rank-15.The CMC curves measure the probability that the correct identity appears within the top-k ranked drivers.Source: §5.2.2 · p. 6, Figure 6 · p. 7

5.2.3 Embedding Similarity Distributions

  • Genuine TEMPEST pairs concentrate above 0.9 while impostor pairs peak near 0.53 with negligible overlap, unlike the substantial overlap in triplet-based models.The distributions compare cosine similarities for same-driver and different-driver embedding pairs.Source: Figure 8 · p. 7, §5.2.3 · p. 7

5.2.4 Per-Driver Identification Analysis

  • TEMPEST achieves a 91% mean per-class F1, with 32 of 45 drivers above 90% F1, an AUC of 99.52%, and an EER of 2.98%.The normalized confusion matrix also shows a strong TEMPEST diagonal with minimal off-diagonal confusion, unlike DriverRep’s diffuse pattern.Source: §5.2.4 · p. 7, Figure 9 · p. 7, Table 3 · p. 8

5.2.5 Cross-Dataset Evaluation on KIA

  • Cross-session KIA AUC is 0.8758 for TEMPEST, versus 0.7119 for Sup-TL and 0.5948 for DriverRep; AUC degradation is 0.11, 0.28, and 0.39, respectively.The cross-session evaluation trains on Drives 1, 2, and 4 and tests on Drive 3.Source: Figure 10 · p. 8, §5.2.5 · p. 8

5.3. Limitations and Practical Considerations

The evaluation is bounded by dataset coverage, sensing choices, and closed-set enrollment, while the model’s compact footprint and embedding-only inference are practical considerations.

Dataset scope and evaluation coverage

  • DriverID has one session per driver on a fixed route, while KIA has four sessions but only 10 drivers, leaving real-world deployment coverage incomplete.DriverID therefore cannot directly assess cross-session generalization on that cohort.Source: §5.3 · p. 8

Sensing and enrollment

  • Physiological sensing uses only one wrist-worn heart-rate monitor, and the experiments assume closed-set enrollment rather than rejecting unknown drivers.Open-set rejection remains an open problem for practical deployment.Source: §5.3 · p. 8

Practical considerations

  • TEMPEST has 720K parameters, a 2.80 MB footprint, and converges in 50 epochs, with the reported run taking 22 minutes on an NVIDIA GB10 GPU.Its decoupled encoder–embedding inference architecture eliminates retraining.Source: §5.3 · p. 8

Privacy considerations

  • At inference, only the 96-dimensional embedding vector is stored and raw sensor signals are not retained, substantially limiting sensitive data exposure.The data were collected under IRB approval with explicit participant consent.Source: §5.3 · p. 8

6. Conclusion

The conclusion presents TEMPEST as a scalable driver-identification embedding, highlights its dynamic enrollment and leakage-aware results, and identifies data collection and model extensions for future work.

Conclusion

  • With the same encoder and inference pipeline, replacing DriverRep’s triplet loss with ArcFace yields a 58.4 pp Rank-1 accuracy gain (33.35% →91.71%) under identical temporal evaluation.The comparison isolates the training objective as the dominant factor in behavioral biometric embedding quality.Source: §6 · p. 8
  • Cosine-similarity matching supports open enrollment by adding drivers to the embedding gallery without retraining or classifier refitting.Source: §6 · p. 8
  • Random cross-validation inflates accuracy by up to 80 pp relative to chronological and cross-session evaluation, motivating realistic, reproducible baselines.Source: §6 · p. 8
  • Future work should prioritize multi-session, multi-route data, adaptive angular margins for confused driver pairs, and attention mechanisms for fusing vehicle and physiological signals.Source: §6 · p. 8

Dataset and Code Dissemination

The paper describes access to the raw DriverID dataset and a planned public release of the code and materials needed to reproduce the work.

Dataset and Code Dissemination

Related papers

Loading 2610.06855v1…