1 Non-Contrastive Learning-based Behavioural Biometrics for Smart IoT Devices

2025-04-30 0 0 1.05MB 12 页 10玖币

侵权投诉

Non-Contrastive Learning-based Behavioural

Biometrics for Smart IoT Devices

Oshan Jayawardana, Fariza Rashid, and Suranga Seneviratne

Abstract—Behaviour biometrics are being explored as a viable

alternative to overcome the limitations of traditional authenti-

cation methods such as passwords and static biometrics. Also,

they are being considered as a viable authentication method for

IoT devices such as smart headsets with AR/VR capabilities,

wearables, and erables, that do not have a large form factor

or the ability to seamlessly interact with the user. Recent

behavioural biometric solutions use deep learning models that

require large amounts of annotated training data. Collecting such

volumes of behaviour biometrics data raises privacy and usability

concerns. To this end, we propose using SimSiam-based non-

contrastive self-supervised learning to improve the label efﬁciency

of behavioural biometric systems. The key idea is to use large

volumes of unlabelled (and anonymised) data to build good

feature extractors that can be subsequently used in supervised

settings. Using two EEG datasets, we show that at lower amounts

of labelled data, non-contrastive learning performs 4%–11%

more than conventional methods such as supervised learning and

data augmentation. We also show that, in general, self-supervised

learning methods perform better than other baselines. Finally,

through careful experimentation, we show various modiﬁcations

that can be incorporated into the non-contrastive learning process

to archive high performance.

Index Terms—Behavioural Biometrics, Smart Sensing, EEG,

Authentication, IoT

This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be

accessible

I. INTRODUCTION

The pervasive use of smart devices and the vast amounts

of sensitive information stored in those devices exacerbate the

problem of user authentication on smart devices. Traditional

methods such as passwords, PINs, and security tokens have

usability issues [1] and static biometrics such as ﬁngerprinting

and face ID are vulnerable to spooﬁng attacks. As a result,

behavioural biometrics has been explored by many works

as a user-friendlier (i.e., implicit by nature and no extra

effort is required from the user) and secure (i.e., difﬁcult

to spoof and allows continuous authentication) alternative for

user authentication in smart devices. Example behavioural bio-

metric modalities include gait patterns [2]), typing patterns [3],

[4], breathing acoustics [5], [6], and EEG patterns [7], [8].

Behavioural biometrics also ﬁnds applications in Smart IoT

devices that either do not have enough form factor or limited

interactive components [7], [6]

Compared to static biometrics, behaviour biometrics needs

a signiﬁcant number of training samples to be collected

from users at registration time and in most cases at different

O. Jayawardana is with The School of Computer Science, The University

Sydney, Australia and The University of Moratuwa, Sri Lanka.

F. Rashid and S. Seneviratne are with The School of Computer Science,

The University Sydney, Australia.

contextual settings [5], [7]. Moreover, the majority of recent

behavioural biometrics solutions use deep learning models that

are known to require higher amounts of training data [9],

[10]. Speciﬁcally, many solutions used Convolutional Neural

Networks (CNNs) [11], [6] or Recurrent Neural Networks

(RNNs) [8], [12] that require massive amounts of labelled data

for better generalisation.

Collecting such volumes of labelled behavioural biometric

data is not practical in many real-world scenarios. For instance,

collecting a signiﬁcant amount of training data at registration

time will inconvenience users, reduce usability, and raise

privacy concerns. As a result, it is important to build learning

methods that enable building deep learning models using less

labelled data.

While collecting large volumes of labelled data for be-

havioural biometrics is challenging and inconvenient, col-

lecting large volumes of unlabelled data is relatively easy.

Unlabelled data can be collected while the device is in use by

the user without any supervision and anonymously so that the

data does not contain any personally identiﬁable information,

eliminating threats to the user’s privacy. For example, a mobile

platform provider planning to build a gait-based behavioural

biometric can collect unlabelled data from the motion sensors

of their platform users. Therefore, it is necessary to develop

learning methods that can leverage large volumes of unlabelled

data to reduce the labelled data requirement of behaviours

biometrics. To this end, in this paper, we propose to use

non-contrastive self-supervised learning. More speciﬁcally, we

make the following contributions.

•We propose a SimSiam [13]-based non-contrastive learn-

ing approach and associated modiﬁcations such as shal-

low feature extractors and weight decay to develop label-

efﬁcient classiﬁers for behavioural biometrics data.

•Using two EEG-based behavioural biometrics datasets in

three authentication system development scenarios, we

show that the proposed non-contrastive learning approach

outperforms conventional supervised learning approaches

by 4%–11% at lower amounts of labelled data. We also

show that non-contrastive learning performs comparably

to a state-of-the-art multi-task learning-based baseline.

•We conduct further experiments and provide insights into

the effectiveness of different types of augmentations on

the non-contrastive learning process. We also provide

empirical evidence of how our modiﬁcations to SimSiam

models help in the context of behavioural biometrics.

The rest of the paper is organised as follows. In Section II,

we present the related work and in Section III, we describe

arXiv:2210.12964v1 [cs.CR] 24 Oct 2022

the overall methodology. Next we explain the datasets and

model details in Section IV. Section V presents the results

and Section VI presents further analysis of various model pa-

rameters’ effect on performance. Finally, Section VII discusses

limitations of our work and possible extensions, and concludes

the paper.

II. RELATED WORK

A. Behavioural Biometrics

There is a vast body of work proposing various behavioural

biometric modalities. Early work involved using typing pat-

terns and touch gestures [3], [14], [15] while later modal-

ities leveraged human physiology [5], [12], [7], [8], [16],

[17]. The authentication solutions generally involve building

machine learning classiﬁers or signature similarity-based ap-

proaches [9]. More recent works use deep learning methods,

given their broader success in other domains [18].

Other works in behavioural biometrics aimed to increase

the training efﬁciency with class incremental learning [19]

or improved label efﬁciency using few-shot learning [20] and

transfer learning [21]. Similar efforts were also made in human

activity recognition [22], [23].

In contrast, we propose to improve the label efﬁciency by us-

ing non-contrastive self-supervised learning. Non-contrastive

learning leverages large volumes of unlabelled data to build

label-efﬁcient classiﬁers. To the best of our knowledge, our

work is the ﬁrst to use non-contrastive learning for be-

havioural biometrics.

B. Self-supervised Learning (SSL)

Self-supervised learning (SSL) refers to a broader family

of methods in which a model learns representations from

unlabelled data using pretext tasks. The pretext task acts as

a feature extractor for supervised learning tasks reducing the

labelled data requirement. For example, in computer vision,

a pretext task learning may train a model to predict whether

an image is an original or an augmentation. In this way, the

model learns the distinguishing features of the original image.

The pretext model is then ﬁne-tuned for a downstream task in

a supervised setting with labelled data. Jing et al. [24] provide

a survey of SSL methods.

Early work closely resembling modern SSL includes Brom-

ley et al. [25], where the authors proposed the ”Siamese”

neural network architecture for signature veriﬁcation. How-

ever, due to excessive resource requirements, SSL didn’t

receive much attention until their success in natural language

processing. In 2013, Mikolov et al. [26] used self-supervised

learning to introduce word2vec, which paved the way to

powerful generative language models such as BERT [27],

RoBERTa [28] and XLM-R [29].

Nonetheless, neither generative methods [30], [31], [32]

nor discriminative approaches [33], [34], [35], [36] were

successful in other domains such as computer vision due

to high computational complexity [37]. In contrast, Siamese

networks-based comparative methods have shown promising

results in computer vision [37], [38], [39], [13].

The basic form of Siamese networks consists of two iden-

tical neural networks which take two views of the same input

(i.e., a positive pair) and outputs embeddings that have a

low energy (or high similarity) between them. To increase

the similarity of the two views, the networks learn spatial or

temporal transformation invariant embeddings. Despite many

successful applications of Siamese Networks, collapsing net-

works (where the network converges to a trivial solution) limit

their performance.

To overcome these limitations, contrastive learning methods

[37], [40], [41], [42], [43] used negatives to avoid collapsing

by not only pulling positives towards each other but also by

pushing apart negatives in the embedding space. An example

is the SimCLR model [37]. However, contrastive learning

requires large batch sizes [37], [43], support sets [41], or

memory queues [42], [44], [40].

As a result, non-contrastive learning methods, and in partic-

ular the SimSiam model [13], emerged as a viable alternative.

Non-contrastive learning generally involves clustering [39],

[45], momentum encoders [38], and using a cross-correlation

matrix between the outputs of two identical networks as the

objective function [46], to address collapsing networks. These

methods avoid the use of negatives to overcome the limitation

of contrastive learning whereby two positive pair samples

can get pushed apart in the embedding space, consequently

becoming a negative pair and harming the performance of the

end task [47]. However, the SimSiam [13] outperforms other

non-contrastive approaches without using complex training

approaches such as momentum encoders. It emphasises the

importance of stop-gradient to present an efﬁcient and a simple

solution to the collapsing networks problem.

C. SSL in Sensing and Behavioural Biometrics

While SSL has majorly contributed to natural language pro-

cessing, computer vision, and speech processing, its feasibility

has been explored in sensing and mobile computing [48].

Saeed et al. [23] introduced self-supervised learning for time-

series sensor data by introducing augmentations that are com-

patible with time-series data. The authors used a multi-task

SSL model to reduce the labelled training data requirement in

Human Activity Recognition (HAR). Using ten labelled sam-

ples per class, the authors achieved approximately 88.8% of

the highest score reached by conventional supervised learning.

SimCLR and several other contrastive and non-contrastive SSL

methods also have been assessed on HAR problems [49], [50].

Others such as Wright and Stewart [51] and Miller et al. [10]

explored the use of traditional Siamese networks to reduce the

training data requirement of behavioural biometrics-based user

authentication.

In contrast to these works, to the best of our understanding,

we are the ﬁrst to propose SimSiam [13]-based non-contrastive

learning for behavioural biometrics to reduce the labelled

data requirement. Our method neither uses negatives nor

requires complex training approaches such as momentum

encoders to avoid collapsing. We compare our approach with

baselines including traditional supervised learning, transfer

learning, data augmentation, and state-of-the-art multi-task

learning [23] and show that it can outperform supervised

learning and provide comparable performance to multi-task

learning at lower amounts of labelled data.

III. METHODOLOGY

A. Non-contrastive Learning Approach

Our approach is based on the SimSiam architecture pro-

posed by Chen et al. [13]. Its architecture is a more simpliﬁed

non-contrastive architecture that doesn’t use negative pairs

or other complex approaches to avoid collapsing. SimSiam

architecture consists of two twin networks that share weights,

as illustrated in Figure 1.

Feature

Extractor

Feature

Extractor

Projector

Encoder Encoder

Projector

Predictor Predictor

x x

Similarity

Fig. 1: SimSiam Architecture

The idea is to learn a good representation of inputs by

solving the problem of increasing the similarity between a

positive pair (xi, xj). A positive pair consists of two randomly

augmented versions of the same input sample x. That is;

xi=τi(x)

xj=τj(x)

Here τis a function that generates a random augmentation

each time it is called. Then the two versions are encoded using

the encoder network g(x;θg),

zi=g(xi)

zj=g(xj)

The encoder consists of a feature extractor gfe(x;θf e)and a

projector gp(x;θp). That is;

g(x) = gp(gfe(x))

The key idea of the projector is to convert the representation

learnt by the feature extractor to a vector that can be used

to calculate the similarity. Next, the encodings go through

another predictor network h(x, θh)before calculating the

similarity.

pi=h(zi)

pj=h(zj)

The purpose of the predictor is to predict the average of

the representation vector across all possible augmentations the

network has seen [13]. Next, the model calculates the cosine

similarity within the pairs (pi, zj)and (pj, zi).

Sim(pi, zj) = pi.zj

kpik2.kzjk2

Sim(pj, zi) = pj.zi

kpik2.kzjk2

Here, k.k2denotes the l2norm of a vector. The task

of the SimSiam model is to increase the total similarity,

Sim(pi, zj) + Sim(pj, zi). To do that, at training time the

symmetric negative cosine similarity loss as deﬁned below is

used.

L=−1

2Sim(pi, stopgrad(zj)) −1

2Sim(pj, stopgrad(zi))

Note that, applying the stopgrad operation is essential for

the SimSiam architecture to work [13]. It considers one side of

the network as constant when computing the gradients of the

other side, to prevent gradients from backpropagating in that

direction as shown in Figure 1. During the training process

the parameters, θfe,θpand θhare learnt.

After the pre-text training of the SimSiam model, we trans-

fer the trained feature extractor gfe(x;θf e)to our downstream

task of building a classiﬁer as illustrated in Figure 2.

Unlabelled

Data

A shared encoder network

Pre-text task

Pre-Trained

Feature extractor

Labelled

Data

Classifier

Network

Feature Extractor Projector

User Classifier

Pre-text

task

scores

Fig. 2: Using the pre-trained feature extractor to build a

downstream task classiﬁer

We introduce two modiﬁcations to make SimSiam ar-

chitecture work for time series behavioural biometrics data

and further improve its performance. They are based on the

hypothesis that easier self-supervision tasks lead to learning

useless features and such features do not hold any value for

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

1Non-ContrastiveLearning-basedBehaviouralBiometricsforSmartIoTDevicesOshanJayawardana,FarizaRashid,andSurangaSeneviratneAbstractBehaviourbiometricsarebeingexploredasaviablealternativetoovercomethelimitationsoftraditionalauthenti-cationmethodssuchaspasswordsandstaticbiometrics.Also,theyarebeingconsi...

展开>> 收起<<

1 Non-Contrastive Learning-based Behavioural Biometrics for Smart IoT Devices.pdf

共12页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

1 Non-Contrastive Learning-based Behavioural Biometrics for Smart IoT Devices

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: