Human Biometrics

Prediction of Human Biometrics

Human biometrics recognition has been of wide interest recently due to its benefits in various applications such as health care and recommender systems. We have developed systems for the prediction and analysis of three biometrics:

Gender recognition.
Age estimation.
Person identification.

In the following we will briefly give some details of each.

We have developed an system, called BioDeep for age estimation and gender recognition based on gait data acquired from Inertial Measurement Units (IMUs). BioDeep consists of two phases: (1) applying a statistical method for feature modeling, the autocorrelation function, and (2) building a Convolutional Neural Network (CNN) for age regression and gender classification. We also used random forest as a baseline model to compare the results achieved by both methods. We have experimented as well with the use of transfer learning across different dataset datasets: we train a CNN on one dataset and reuse its feature maps over the other datasets for solving both age estimation and gender recognition. Transfer learning achieved 20−30x speedup in the training time in addition to keeping the acceptable prediction accuracy.

BioDeep System Architecture

Age regression results for each dataset evaluated in mean absolute error.

RF stands for the use of the random forest model. It is apparent that CNN performs better across all datasets used for the empirical studies as well as for different sensors' locations. All other abbreviations refer to the sensors' locations on the human body (see the reference for details). The following figure shows the corresponding results for gender classification (using the accuracy metric). Almost all the results give high accuracy in the order of >90%, across almost all sensor locations. This indicates that the gait style is very discriminative between males and females and such differentials are manifested across the whole of the human body. The OU-ISIR dataset (from Osaka University) always gave the worst results. This observation is also consistent across all the literature that worked on this dataset. One explanation might be, though this dataset has a huge number of subjects the recording time for each subject is so small, that the walking dynamics may not have been captured well.

Gender classification results for each dataset evaluated in percentage accuracy.

One of the interesting aspect about this work is the transfer learning results which has proven to be effective in the current tasks of age estimation and gender classification. Though still much more empirical work need to be done over larger number of datasets in order to verify that possibility. The following two tables show samples of cross-testing results where a model is pretrained on some dataset and tested on another. The diagonal elements in these tables represent the base case, where the training and testing datasets are the same.

Results of cross-testing in age regression evaluated in mean absolute error.

Results of cross-testing in gender classification evaluated in percentage accuracy.

References:

Abeer Mostafa., Samir Elsagheer., and Walid Gomaa. Biodeep: A deep learning system for imu-based human biometrics recognition. In Proceedings of the 18th International Conference on Informatics in Control, Automation and Robotics - ICINCO,, pages 620–629. INSTICC, SciTePress, 2021.
Abeer Mostafa., Toka Ossama Barghash., Asmaa Al-Sayed Assaf., and Walid Gomaa. Multi-sensor gait analysis for gender recognition. In Proceedings of the 17th International Conference on Informatics in Control, Automation and Robotics - Volume 1: ICINCO,, pages 629–636. INSTICC, SciTePress, 2020.

The next main problem we tackled in human biometrics is person identification through gait signals. Gait-based person authentication using smartphones and wearable devices has attracted a lot of attention in recent years due to its unobtrusive nature and the widespread proliferation of inertial sensors embedded in such devices. While traditional methods of authentication relied on hand-crafted feature extraction tailored for specific conditions, recent deep learning algorithms have been used to automatically discover hidden features and patterns in the raw data. However, current deep learning methods of gait authentication are not portable. They train models to classify a specific user as legitimate and the rest as imposters without being able to change that user to a different one without retraining the model from scratch. Furthermore, they are impractical, since they require a large volumes of data for each class to train properly. In our work, we proposed using a Siamese Network based model which (1) can be trained using only few samples per subject. (2) automatically extract features from raw inertial data., and(3) once trained, can authenticate any new user given only one gait instance of that new user. Our experimental results show that our model achieves 3.42% EER (equal error rate) on the world’s largest inertial gait dataset (OU-ISIR) after training on raw data of 592 subjects and testing on the raw data of completely different 78 subjects, where each subject provided an average of 5.9 seconds of gait data only. We also showed the ability of our model to generalize to new unseen datasets using transfer learning. The proposed framework is depicted in the following figure.

Proposed framework including Siamese Network architecture for authentication.

Projection of the three datasets’ difference vectors D over 2-D plane using PCA. Red points correspond to similar vectors and blue points correspond to dissimilar vectors. The first row is the model’s predictions and the second row is the ground truth.

The empirical work has produced promising results, samples of these results are shown below.

Average, standard deviation, min and max EER (equal error rate) on experiments with different CNN architectures for the Siamese network.

It is apparent that the performance improves across all datasets with increasing the size of the difference vector. In the left table below we train on the absolute difference vectors generated from the training subjects, then test on the absolute difference vectors generated from the test subjects. Instead of reporting the EER, we report the accuracy over 10 runs. Of course, the input is the input to the Siamese network which are two gait samples, either from the same subject or from two different subjects. The table on the right shows transfer learning results, where the model is trained in some dataset and tested on another with quite different hardware and software configurations. Of course, the diagonal elements act as the baseline against which the transfer learning results can be assessed. It is apparent that the dataset EJUST-GINR-1 is the best regarding the transfer learning results, that is, it induces the lowest degradation in performance when testing on another dataset. However, it is the worst as a baseline.

Accuracy of using different on top of the absolute difference vectors (that result from the Siamese network.

EER (mean±std)% of transfer learning results

References:

Osama Adel, Mostafa Soliman, and Walid Gomaa. Inertial gait-based person authentication using siamese networks. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1–7, 2021.
Osama Adel., Yousef Nafea., Ahmed Hesham., and Walid Gomaa. Gait-based person identification using multiple inertial sensors. In Proceedings of the 17th International Conference on Informatics in Control, Automation and Robotics - Volume 1: ICINCO,, pages 621–628. INSTICC, SciTePress, 2020.

Page updated

Report abuse