A comparative analysis use of recurrent auto-encoders as a component of the proposed automated decision support system for biometric verification of the behavioral type user from the indicators of the chest accelerometer sensor is carried out. The purpose of the proposed system is to perform continuous, implicit user verification based on the accelerometer, to improve the security usage of application or device, as well as improve user interaction with the device. The goal of research is to build a continuous-based user biometric verification system based on accelerometer data using unsupervised deep learning recurrent algorithms. Auto-encoders enable one-class classification, i.e. unsupervised learning, as well as additional layers in the network architecture allow to automatically generate features and encode a data input into a feature vector, which greatly facilitates data processing and makes the system more automated. The purpose of the model is to define the boundaries of the positive class, namely to distinguish a specific user from others, which in fact is to solve the problem of detecting anomalies. The comparative analysis compared three types of recurrent auto-encoders with recurrent units of long-short term memory type with two methods of classical machine learning (one-class support vector machines and “isolation” forest) which required manual feature generation. Proof of concepts of using recurrent autoencoders for biometric verification was implemented and tested on the open-source dataset. Usage of recurrent autoencoders for user behavioral-based verification has shown robust and high accuracy results and the ability of implementing such algorithms in modern security systems.
Nowadays our phones and other big variety of electronic devices become a great part of our life. We have a lot of information on our phones, even such important info like the number of credit cards, credentials from some payment service, etc. Also, now it is very important to make analytics on your devices, for further development of different data-driven decisions, which will improve the level of user satisfaction from using the app and provide a smart decision-making system on the user device. So, it is very important to provide accurate user identification.
The problem of user verification, that this is impossible to create a supervised task because we can’t compare one particular user with all other users. So, this becomes the unsupervised task and the aim is to detect the boundaries of one good class and all the samples that don’t belong to it, will be considered as anomalies. In the context of user verification, this will be other users. The data from the accelerometer is in a timeseries format. The objective is to solve the anomaly detection problem on time series data, which is the task of higher complexity.
Reference to [
An object of the study is the user biometric verification based on behavioral patterns. The accurate and speed behavioral-based verification allows making it continuous and implicit, which is very important to create secure and protected services for different purposes.
The subject of study is unsupervised machine and deep learning-based algorithms for one-class classification.
The purpose of the work is the comparative analysis of using different unsupervised deep learning autoencoder-based approaches for solving biometric verification problem which is, in fact, an anomaly detection problem. 2
Given the set of time series windows of length l and overlapping percent p with particular activity a { x ai }- i-th data instance with activity a, where i =1,2,...N (N – number of instances in train set).
The problem of user verification can be formulated as finding such Topt, which is optimal for next target functions:
N M { (E(xi ) T ) → max, (E(z j ) T ) → min} ,
i=1 j=1 where T - error threshold, which detects abnormality of given data instance; Topt optimal threshold; δ - function which get the logical expression and returns 1 if it is true, otherwise 0; E(x) - model error on given x data instance; M - number of anomaly data instances; zj - j-th anomaly data instance. 3
In [
Also, it is possible to use autoencoders not for authentication but for the inverse
problem - anonymization, to prevent deanonymization when sending data to untrusted
resources [
Not only accelerometer data can be used for biometric identification, but also ECG or
EEG data. In [
The idea of implicit authentication based on behavioral patterns such as a
combination of all data sources which your smartphone can propose, like application usage,
phone call patterns, etc. proposed in [
Another set of algorithms that can be used as anomaly detection method are
artificial immune systems (method of positive and negative clonal selection for example),
and in [
So, for working with sensor time-series data, we need to use recurrent architectures but as well we want to use unsupervised algorithms both for feature engineering and for detecting non-self users. In this work, we combine all these parts as one model to check the ability of such algorithms to solve such complex, multistage problems. 4
For biometric user verification were used autoencoders.
Here we look deeper into autoencoder internals.
Autoencoder architecture contains two main parts: encoder and decoder.
Autoencoders have learned to create a reconstruction of the original input. The goal is to minimize reconstruction error:
E = 1. Undercomplete (when the dimension of inner layers is smaller than input), encoder solve dimensionality reduction problem. 2. Sparse (when the dimension of inner layers is bigger than input), adding some sparsity penalty to reconstruction error. 3. Denoising (adding some noise to input and put it into reconstruction error). It minimizes not (1) but (2)
Where xi is x with some noise.
4. Contractive (add the penalty Ω(h) is the squared Frobenius norm (sum of squared
elements) of the Jacobian matrix of partial derivatives associated with the encoder
function)[
Variational autoencoders are directed probabilistic graphical model. Variational autoencoders don’t minimize reconstruction error but optimize reconstruction probability.
It is absolutely different approach, we still have the encoder and decoder, but
decoder doesn’t decode, but sample data examples from some probabilistic distribution
with some parameters and encoder maps data samples into latent probabilistic space,
so in core of loss function for variational autoencoder we have Kullback–Leibler
divergence, which shows the difference between different probabilistic distributions
[
In [
So based on previous consideration the purpose is that for user personalization problem solving based on chest accelerometer sequential data it is important to use recurrent architectures to think about previous values, and because of high complexity of solving classification problem of distinguishing users, the best approach is to define boundaries of “self” class (the user in this case), and for this we need unsupervised machine learning techniques.
Using recurrency and unsupervised learning together the answer is recurrent autoencoders.
Let's propose a continuous verification system (fig. 1). 6. Collecting accelerometer data (or other sensor data, that describe motion patterns) for different activities type. 7. Recognize human activities and divide data into parts by activities. 8. Train models and fine-tune parameters. 9. Choose the most appropriate model for a particular user based on a low falsepositive rate.
Dataset was taken from UCI repository - archive of public datasets for machine learning purposes.
The dataset contains accelerometer data from 15 volunteers. Data contains from x, y, z axis values [14]. The sampling frequency of the accelerometer is 52 Hz, so we have 52 rows of values each second. Also, data contains 7 activities labels (standing up, walking and going up and downstairs, standing, walking, going up and downstairs, walking and talking with someone, talking while standing). As in [15] user verification goes after solving human activity recognition (HAR) tasks on walking patterns, we also use the user verification algorithm only on some particular motion pattern. So, actually we detect in which way some person walking or walking and talking with someone etc. The most rows from volunteers were with working at a computer label, so we decide to train models with working at computer activity.
For deep learning models, we split data in overlapping on 50 percent windows with a length of 52.
The number of train and test samples shown in table 1.
We build autoencoders with python on Keras library and Tensorflow backend [16].
We compare three types of autoencoders: ─ variational long-short term memory (LSTM) autoencoder; ─ contractive LSTM autoencoder; ─ undercomplete LSTM autoencoder.
As loss was used mean absolute error. Train in 10 epochs with 32 batch size and Adam optimizer.
The threshold formula:
T = N MAEi i=1
n + std (MAEi ) where MAE is mean absolute error between a sample and predicted sample, std standard deviation and n is the number of samples in the train dataset.
The architecture of autoencoders is shown below at fig. 2.
As model evaluation metric “recall” for positive (“self”) class were chosen and also area under curve (AUC) value for correctly compare models with each other, not depending on classification threshold. Also equal error rate was calculated.
Autoencoder-based models were compared with one-class support vector machines (SVM) and isolation forest algorithms. For one-class SVM and isolation forest, not raw data, but different types of features were extracted (in time and frequency domain). The examples of features were taken as in [17]. Fig. 2. Architecture of a) contractive autoencoder, b) variational autoencoder, c) undercomplete autoencoder
Positive class recall score for every user and comparing with one-class SVM and isolation forest are shown in table 1-3 below.
Also because of not very high difference in recall score in autoencoder models, we compare them with area under curve value. ROC curve and AUC value is shown in fig. 3 and fig. 4.
In table 2 we can see that the best result showed recurrent contractional autoencoder with 0.82 average recall score for positive class.
Based on EER value in table 3 contractive and variational autoencoder show almost the same performance.
On the fig. 3 and fig. 4 we can see ROC curves and AUC values for every type of autoencoders that was used in experiment. AUC value allows us to compare the models in general, while “recall” is calculated on some fixed threshold. Thus, AUC metric is better for model comparison, while “recall” score should be considered more as final key performance indicator (KPI) for specific business usage. Based on the results, AUC value is different depending on user, so we can say that the model should be adaptively selected based on performance, error rate and amount of data for particular user. However, contractive autoencoder is still a leader, as it has highest result in 8 users from 15, the second is classic autoencoder. 7
Following from previous results as we can see recurrent autoencoders show robust results with a high value of chosen metric for evaluation. Also, depending on data size and particular user result can be different. For every user, the best way will be to choose the model with the highest recall and AUC value. Comparing to a one-class SVM and isolation forest where feature where manually engineered, automatically feature extraction as part of the autoencoder model doesn’t underperform, which is important for ubiquitous usage, its versatility, and portability for different types of sensors and data sources.
But overall most robust results were shown by recurrent contractive autoencoder.
Also, should be considered the data separability degradation with increasing the numbers of user. In our case we have only 15 users, which is a pretty small number compare to the real-world tasks. In this case, we have to understand how properly train and evaluate the model.
For example, we can cluster users in different group and evaluate the model only inside specific group of most similar user. Thus, we will choose the model which distinguish “self” user from others, even the most similar one.
We need to consider possible issues, that may appear in using such systems in real world problem, as: a big number of similar users, noisiness of accelerometer signal, high model latency, etc. Another loss should be considered as well in order to customize the problem for specific biometric purposes.
Our research shows that using deep unsupervised learning can be successfully used as part of biometric continuous personalization decision support system. 8
Here we compare different autoencoders for biometric user personalization based on motion patterns. The purpose was to define the boundaries of positive class to distinguish particular users from the rest. Undercomplete LSTM autoencoder, LSTM variational autoencoder, and contractive LSTM autoencoder were compared with one-class SVM and isolation forest. Recurrent autoencoders show robust and high accuracy results. The advantages of autoencoders are the unsupervised approach and automatic feature extraction, engineering, and selection, the disadvantage is their requirements for computational resources.
The contractive autoencoder show the best results with 0.82 recall score for positive class and highest AUC value for 8 users from 15. Contractive autoencoder show the lowest equal error rate as well (0.106). The other type of autoencoders, as classic autoencoder and variational show the good result with 0.81 recall score for positive class. Compare with classic machine learning algorithms as isolation forest and oneclass SVM with 0.77 and 0.47 recall score respectively autoencoders show much better results.
In the future, we can extend the list of autoencoder models, and use more sophisticated cascades of autoencoders, like stacking autoencoders of different types in one big model or adding convolutional and subsampling layers before encoder to improve feature extracting part of the model, but there may be an issue with computational efficiency. Also, there has to be research about optimization using deep learning models in different decision support systems. As well there is a need to propose the more complex evaluation metric for such types of models. 13. Dua, D. and Graff, C. UCI Machine Learning Repository. Irvine, CA: University of
California, School of Information and Computer Science (2019). 14. UCI machine learning repository, https://archive.ics.uci.edu/ml/datasets/Activity+Recognition+from+Single+ChestMounted+Accelerometer, last accessed 2020/15/10. 15. Casale, P., Pujol, O., Radeva, P.: Personalization and user verification in wearable systems using biometric walking patterns. Personal and Ubiquitous Computing. 16, 563–580 (2011). 16. Chollet, F.: The Keras Blog, https://blog.keras.io/building-autoencoders-in-keras.html, last accessed 2020/15/10 17. Thingom, B. S., Rajsekhar, K. N., Narsimhadhan, A. V.: Person Recognition using
Smartphones' Accelerometer Data. CoRR abs/1711.04689 (2017). 18. Nguyen, Q.P., Lim, K.W., Divakaran, D.M., Low, K.H., Chan, M.C.: GEE: A Gradientbased Explainable Variational Autoencoder for Network Anomaly Detection. 2019 IEEE Conference on Communications and Network Security (CNS) (2019).