Source-linked AI summary
IDNet: Smartphone-based Gait Recognition with Convolutional Neural Networks
Matteo Gadaleta, Michele Rossi
TL;DR
Wearable-data analysis and authenticating mobile users are identified as major problems in assessing whether data sources are genuine. IDNet is proposed as a user-authentication framework combining automatic feature engineering with classification techniques and one-class support vector machines. Results report misclassification rates smaller than 0.15% in fewer than five walking cycles, with superiority against prior work.
Problem
Wearable-data analysis and authenticating mobile users are identified as major problems in assessing whether data sources are genuine.
Method
IDNet combines automatic feature engineering with classification techniques and one-class support vector machines for user authentication.
Results
Misclassification rates smaller than 0.15% were reported in fewer than five walking cycles, with superiority against prior work.
Takeaways & Limitations
IDNet is presented as a user-authentication framework whose results show superiority against prior work.
Takeaways & Limitations
Comparisons are affected by differences in datasets and acquisition settings, including subjects, walking time, and smartphone or sensor location.
Abstract
from arXiv · showhide
Here, we present IDNet, a user authentication framework from smartphone-acquired motion signals. Its goal is to recognize a target user from their way of walking, using the accelerometer and gyroscope (inertial) signals provided by a commercial smartphone worn in the front pocket of the user's trousers. IDNet features several innovations including: i) a robust and smartphone-orientation-independent walking cycle extraction block, ii) a novel feature extractor based on convolutional neural networks, iii) a one-class support vector machine to classify walking cycles, and the coherent integration of these into iv) a multi-stage authentication technique. IDNet is the first system that exploits a deep learning approach as universal feature extractors for gait recognition, and that combines classification results from subsequent walking cycles into a multi-stage decision making framework. Experimental results show the superiority of our approach against state-of-the-art techniques, leading to misclassification rates (either false negatives or positives) smaller than 0.15% with fewer than five walking cycles. Design choices are discussed and motivated throughout, assessing their impact on the user authentication performance.
1. Introduction
IDNet authenticates mobile users from smartphone inertial motion data by combining orientation-independent gait processing, CNN feature extraction, and multi-stage classification. The framework achieves high authentication accuracy and reports misclassification rates below 0.15% with fewer than five walking cycles.
- IDNet authenticates mobile users from smartphone-acquired motion data using accelerometer and gyroscope signals.
- The preprocessing stage extracts walking cycles and transforms smartphone motion signals into an orientation-invariant reference system.This design addresses varying smartphone orientations during use.
- CNNs are trained once on representative users and used at runtime as universal feature extractors for gait signatures.The CNNs automatically extract statistical features rather than relying on predefined features.
- CNN-extracted features are combined with a one-class SVM trained solely on the target subject, while scores accumulate across walking cycles.This combination forms the paper’s multi-stage identification framework.
- Gyroscope integration further improves performance, and IDNet outperforms state-of-the-art gait authentication solutions.The paper reports authentication errors below 0.15% using fewer than five walking cycles.
2. Related Work
Prior smartphone gait-recognition studies used varied sensors, segmentation methods, handcrafted features, and classifiers, often under controlled device orientations. This work adopts an orientation-invariant signal transformation and combines automatic CNN feature extraction with one-class SVM scoring and multi-step decisions.
- Smartphone gait recognition: Earlier studies analyzed gait using smartphone accelerometers or inertial sensors placed on the belt, hip, or waist, commonly with constrained orientations.Ailisto et al. used a triaxial accelerometer with fixed axis directions, while other studies used hip- or waist-worn phones and single-axis analysis.
- Signal processing and classification: Related systems extracted gait cycles through peak detection, fixed windows, or other segmentation schemes before comparing templates or classifying segments.Template matching, DTW, HMMs, and machine-learning classifiers were among the reported approaches.
- Limitations of prior work: 30 seconds of authentication were required in one HMM-based approach despite good identification results.This illustrates the authentication-duration cost reported for that system.
- Features and classifiers: Handcrafted temporal, statistical, frequency, and cepstral features were widely used with classifiers including k-NN, SVMs, MLPs, and classification trees.The reviewed features included mean, standard deviation, root mean square, zero-crossing rate, histogram counts, Fourier or wavelet features, and cepstral coefficients.
- Orientation invariance: Most related work assumed controlled sensor orientation, whereas this paper adopts an orientation-invariant three-dimensional reference system extracted directly from inertial data.The paper contrasts rotation-invariant features with signal transformation and chooses the latter approach.
- Proposed pipeline: The proposed pipeline uses automatic CNN feature extraction together with one-class SVM scoring and multi-step decision analysis.These components are presented as distinctive features of the processing pipeline.
3. Signal Processing Framework
IDNet processes smartphone inertial signals through preprocessing, walking-cycle extraction, orientation-independent transformation, normalization, CNN feature extraction, and one-class classification. The framework addresses limitations of hand-designed statistical features and existing gait datasets while targeting user recognition from walking patterns.
- IDNet targets user recognition from walking patterns acquired through smartphone accelerometer and gyroscope signals.
- The signal-processing workflow includes pre-filtering, walking-cycle extraction, orientation-independent transformation, and cycle normalization.Cycles are represented as fixed-length, zero-mean, unit-variance vectors before identification.
- CNNs automatically assess statistical features during supervised training instead of relying on designer-selected features refined through trial and error.The CNN is trained as a universal feature extractor, followed by a one-class SVM as the final classifier.
- IDNet assumes the CNN produces discriminant features for unseen users while the one-class SVM is trained only on the legitimate user’s walking data.Impostor walks are not used for training.
- The authors collected motion data from 50 subjects over six months using Android smartphones worn in users’ right front pockets under variable, realistic conditions.Sessions used multiple smartphone models, different shoes and clothes, and comfortable walking behavior.
- The accelerometer-magnitude cycle extraction uses correlation-distance local minima to align templates, while weighted averaging produces an increasingly robust template.The resulting extraction is intended to remain invariant to smartphone rotation and device carrying orientation.
- A PCA-based transformation makes walking data rotation invariant, allowing subject-specific gait patterns to emerge in the new coordinate system.
4. Convolutional Neural Network
IDNet uses a convolutional neural network to extract gait features from smartphone accelerometer and gyroscope signals, then classify walking cycles. The architecture combines convolutional feature extraction with a fully connected classifier and is optimized through supervised training and validation.
- Architecture: Two convolutional layers reduce dimensionality and extract features, while a fully connected layer performs classification.The first layer processes signal vectors separately; the second captures cross-vector correlations and applies pooling.
- Architecture: The CNN processes normalized accelerometer and gyroscope signals from each walking cycle as its input matrix.The matrix contains axis-specific and magnitude signals for both sensors.
- Architecture: The FL1 output contains F CNN-extracted features, while FL2 uses K softmax outputs corresponding to the trained subjects.Each softmax entry represents the probability that the input belongs to a particular user class.
- Training: The CNN is trained as a supervised K-class classifier by minimizing categorical cross-entropy with stochastic gradient descent.Training uses separate training, validation, and test sets, with validation loss determining termination and the selected network weights.
- Optimization and results: At least F = 20 neurons are needed in FL1, accuracy stabilizes around F = 40, and larger networks outperform the smallest configuration.Increasing the kernel count beyond the intermediate larger network produces no appreciable improvement.
- Optimization and results: CNN features improve downstream classification accuracy by more than 10% in typical cases, including k-NN accuracy rising from 71% to 94% at Nc = 30.The comparison uses manually extracted features as the alternative to CNN-extracted features.
5. One-Class Support Vector Machine Training
IDNet converts CNN outputs into reduced feature vectors and classifies them with a one-class SVM trained only on the target user’s motion data. The section evaluates OSVM parameter robustness, PCA selection, sensor inputs, and training-set size.
- OSVM classification: The OSVM is trained solely with target-user motion data, enabling one-class classification when data from other users are unavailable.This setting is motivated by practical authentication scenarios where target-user signals are available but other-user signals are not.
- Revised classification architecture: The CNN acts as a universal feature extractor by discarding its output layer and producing user-dependent feature vectors for new subjects.PCA further reduces the CNN feature vector from F dimensions to S dimensions before OSVM classification.
- OSVM classification: The OSVM learns a boundary around target-user feature vectors and produces a score based on a new vector’s distance from that boundary.Positive-class-only training supports extending the classifier to users not included in CNN training.
- Parameter analysis: For F = 40, retaining S < F PCA features and selecting the lowest-variance components gives better OSVM results.The finding concerns the PCA transformation from CNN-extracted features to reduced OSVM inputs.
- Parameter analysis: Training with more than 1,000 target-user cycles yields little improvement, corresponding to about 15 minutes of walking distributed across sessions.Multiple acquisition sessions are recommended to account for variation caused by wearing conditions.
6. Multi-Stage Authentication
IDNet aggregates OSVM scores from successive walking cycles and applies a sequential probability-ratio test to decide whether the sequence belongs to the target user. The multi-stage framework achieves low error rates while limiting the number of cycles needed for authentication.
- Sequential decision process: The framework resolves uncertain individual OSVM scores by jointly analyzing scores from successive walking cycles.A score near the intersection of the target and non-target probability densities indicates high identity uncertainty.
- Assumptions: The sequential model assumes that scores from successive cycles are generated by the same subject and are independent and identically distributed.A first-order autoregressive extension was tested but increased complexity without appreciable improvement because temporal score correlations were light.
- Sequential decision process: SPRT accumulates log-likelihood ratios until the result crosses threshold A or B, accepting either the non-target or target hypothesis.The test continues to the next cycle while A < Λ_n < B.
- Authentication results: False positive and false negative rates are both smaller than 0.15% with an appropriate choice of SPRT thresholds.The reported evaluation uses leave-one-session-out cross-validation, with one target user and 14 negative users in the final tests.
- Authentication results: Fewer than five walking cycles are sufficient for reliable authentication in 80% of cases.The authors characterize the framework as both accurate and fast under this cycle requirement.
- Comparison and scope: Direct comparison with prior approaches is difficult because datasets, walking durations, acquisition settings, and sensor locations differ.The paper instead points to comparisons with classical feature extraction techniques on its own dataset.
7. Conclusions
IDNet combines convolutional neural networks, one-class classification, and multi-stage decision-making for smartphone-based inertial gait authentication. Its reported misclassification rates were below 0.15% with fewer than five walking cycles.
- Contributions: IDNet uses convolutional neural networks for automatic feature engineering and universal feature extraction in gait recognition.The extracted features feed a one-class support vector machine and a multi-stage decision algorithm.
- Contributions: The paper contrasts IDNet's learned feature extraction with prior approaches based on manual feature extraction and selected-feature classification.The stated contrast is between CNN-based automatic feature engineering and recent smartphone-motion methods using manually extracted features.
- Contributions: The framework combines target-only training with score accumulation across subsequent walking cycles for authentication.The classifier scores the dissimilarity of newly acquired data from the target and accumulates scores through multi-stage decisions.
- Experimental validation: IDNet was experimentally validated against prior solutions, with the paper reporting superiority over those approaches.The conclusion describes design choices and processing-block optimization as discussed and compared against prior work.
- Experimental validation: Misclassification rates were smaller than 0.15% using fewer than five walking cycles.The reported misclassification measure includes false negatives or positives.