cs.LGSep 26, 2026

Handwritten Digit Leakage from Smartphone Motion Sensors Across Unseen Users and Phone Models

Authors: Constantino Álvarez Casado, Erkka Rantahalvari, Matteo Pedone, Matti Matilainen, Manuel Lage Cañellas, Le Nguyen, Simo Hosio, Olli Silvén, 

Organizations: University of Oulu, Oulu, Finland · Candour Ltd, Oulu, Finland · University of Tokyo, Tokyo, Japan

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Smartphone motion sensors support interactive applications, but their readings may also reveal touchscreen input beyond their intended use. Assuming known drawing intervals, we study whether handwritten digits remain predictable across users and devices, as a 10-class problem on 19,628 HuMIdb recordings from 481 participants. We compare handcrafted features with classical machine learning algorithms, MiniRocket kernels, and a compact sensor patch transformer on accelerometer, linear acceleration, gyroscope, and gravity signals. The transformer achieves 57.74% accuracy and 82.64% top-3 accuracy on 75 unseen participants, and 58.77±\pm0.95% over 3 seeds for unseen participants on 9 unseen phone models. Low motion recordings remain informative, accuracy is not monotonic in motion level, and the tested contrastive pretraining, augmentation, and derived signals give no consistent gains. Digits are thus predictable beyond familiar users and phone models under assumed segmentation, while acquisition-order shortcuts limit conclusions about practical privacy exposure. Code available at: https://github.com/Arritmic/motion-digit-leakage.

Figures & tables

Figure 1: Overall view of the evaluation pipeline, illustrated with a correctly classified test recording of the digit 5. (a) Finger trajectory, dot at first contact. Touch supplies only the label and the drawing interval. (b) Raw motion after mean removal, xx , yy , zz from dark to light, drawing interval in gray. (c) Aligned 14 channel input, standardised per channel for display. (d) Classifier families, fitted on training participants. (e) Frozen transformer probabilities.
 New at test     
ProtocolpeoplephonesTrainTestPeopleAcc.95% CI
P✓\checkmark×\times13,7943,0367557.74[54.31, 61.18]
S✓\checkmark×\times8,6022,5936454.80[50.57, 58.69]
M×\times×\times8,6022,59332356.69[54.15, 59.22]
U✓\checkmark✓\checkmark8,6472,6085957.71[53.10, 61.59]
U, 3 seeds✓\checkmark✓\checkmark8,6472,6085958.77SD 0.95
Table 1: Protocols and transformer test accuracy (%). Counts are recordings. 95% bootstrap intervals over participants (P, S, M) or phone models (U). Last row: mean and SD over three seeds.
Figure 2: Schematic of the four protocols. Columns are phone models and each cell is one participant, who uses a single phone (A3). In M, each participant contributes recordings to every partition. Proportions are illustrative.
ComponentSetting
ConditioningShared overlap, median step grid, sixth order Butterworth at 0.4/Δ0.4/\Delta , T=128T=128 , C=14C=14
Handcrafted29 statistics per channel at 50 Hz, 406 features, median imputation, standardization
RBF-SVMC=10C=10 , γ=1/(d Var)\gamma=1/(d\,\mathrm{Var}) , Platt probabilities
Random forest500 trees, d\sqrt{d} features per split, minimum leaf size 1
MiniRocket9,996 features, at most 32 dilations per kernel, PCA to 256, RBF-SVM
TransformerP=8P=8 , S=4S=4 , D=64D=64 , 2 layers, 4 heads, feed forward 256, GELU, dropout 0.15, 108,426 parameters
Table 2: Implementation settings.
Representation and trainingValidationTest
Handcrafted, RBF-SVM39.6739.72
Handcrafted, random forest35.9537.85
MiniRocket, PCA 256, RBF-SVM †54.0452.64
Transformer †56.7555.67
without temporal positions †35.6735.97
Transformer (selected) ∗60.2957.74
Table 3: Accuracy (%) on P, seed 2026 above, mean ±\pm SD over three seeds below. † Independent sensor time axes, no antialiasing. Contrastive models are fine tuned with augmentation. ∗ Selected on validation before the test set was opened, the only prospective test result. Other test values are retrospective.
Figure 3: (a) Test accuracy of the 75 P participants. (b) U accuracy by motion level (training on all recordings) and outside the low group (64.7% of test) for three training sets, mean and SD over three seeds.
Figure 4: Test confusion matrices normalized by true digit (%). (a) P, frozen transformer, 3,036 recordings, 75 unseen participants. (b) U, three seeds pooled, 2,608 recordings, 59 unseen participants, nine unseen phone models. Off diagonal values below 10% omitted.
More in this topic

Handwritten Digit Recognition

See all 8 papers

Explore similar work

CardsList

  1. DigitCode: Symbolic Tokenization of Hand Motion by Anatomical Units

    Aug 4, 2026Haoyu Gu, Haotian Lu, Jingrun Du +1Human Motion Generation

    School of Future Technology, South China University of Technology · Shenzhen Key Laboratory of Ubiquitous Data Enabling, Tsinghua Shenzhen International Graduate School, Tsinghua University

  2. Secrets Everywhere: Auditing Memorization in Mobility Prediction Models

    Aug 3, 2026Anne Josiane Kouam, Hristo Boyadzhiev, Konrad RieckPrivacy AuditingHuman Behavior Prediction

    Inria & TU Berlin Palaiseau, France · BIFOLD & TU Berlin Berlin, Germany

  3. MultiGait: A Multi-Sensor Multi-Perspective Multi-Session Biometric Inference Benchmark and its Dataset

    Sep 1, 2026Julian Todt, Felix Morsbach, Philip Dissert +1Privacy AuditingBiometric Identification

    KASTEL Security Research Labs · Karlsruhe Institute of Technology