iCTGANAn Attack Mitigation Technique for Random-vector Attack on Accelerometer-based Gait Authentication Systems Jun Hyung Mo

2025-05-08 0 0 812.32KB 9 页 10玖币

侵权投诉

iCTGAN–An Attack Mitigation Technique for Random-vector Attack on

Accelerometer-based Gait Authentication Systems

Jun Hyung Mo

Haverford College, USA

junmo318@gmail.com

Rajesh Kumar

Bucknell University, USA

rajesh.kumar@bucknell.edu

Abstract

A recent study showed that commonly (vanilla)

studied implementations of accelerometer-based gait

authentication systems (vABGait) are susceptible to

random-vector attack. The same study proposed a beta

noise-assisted implementation (βABGait) to mitigate the

attack. In this paper, we assess the effectiveness of the

random-vector attack on both vABGait and βABGait using

three accelerometer-based gait datasets. In addition, we

propose iABGait, an alternative implementation of ABGait,

which uses a Conditional Tabular Generative Adversarial

Network. Then we evaluate iABGait’s resilience against

the traditional zero-effort and random-vector attacks.

The results show that iABGait mitigates the impact of

the random-vector attack to a reasonable extent and

outperforms βABGait in most experimental settings. 1

1. Introduction

1.1. Gait

Gait refers to how a person walks. If captured properly,

it can be used to identify or verify one’s identity [1,

2]. While Aristotle studied it in 350 BC, human gait

became popular in the biometrics community in the 1990s

primarily due to the advancements in the means (e.g.,

cameras, force plates, 3D motion detection) that could

capture gait more precisely than ever before. Later, with the

advent and massive adaptation of smart devices, especially

smartphones and smartwatches, the study of gait attracted

even more researchers. A smart device consists of inertial

sensors, e.g., accelerometer, gyroscope, magnetometer, and

rotation vector, that help capture and characterize gait in

from IEEE must be obtained for all other uses, in any current or future

media, including reprinting/republishing this material for advertising

or promotional purposes, creating new collective works, for resale or

redistribution to servers or lists, or reuse of any copyrighted component of

this work in other works. To appear in IEEE International Joint Conference

on Biometrics (IJCB 2022), Oct.10-13, 2022, Abu Dhabi, United Arab

Emirates.

ways not studied before. For example, an accelerometer

captures and characterizes gait in terms of acceleration,

and a gyroscope does the same in rotation. The gait

captured via accelerometer and gyroscope is considered

the most distinguishing among individuals [2,3]. Most

gait recording technologies can be grouped into intrusive

or constrained (cameras and force plates are used) and

non-intrusive (personal devices such as smartphones and

smartwatches). Different recording mechanisms offer

different application possibilities, such as using gait for

surveillance, identiﬁcation, authentication, and healthcare

monitoring.

In this paper, we focus on Accelerometer-based Gait

Authentication Systems (ABGait) [2,3]. ABGait has

attracted the attention of defense organizations like the

Defense Advanced Research Projects Agency and The

Defense Information Systems Agency of the United States.

These agencies are in the process of implementing security

measures that will utilize a variety of behavioral biometrics,

including ABGait [4]. ABGait presumably offers immunity

to shoulder surﬁng, social engineering, or spoof-based

circumvention attempts [5] in addition to convenience due

to its passive and non-intrusive nature.

1.2. Accelerometer-based Gait Authentication

Previous studies have demonstrated that inertial

sensor-based gait patterns, regardless of whether the

sensors were embedded into a smartphone or smartwatch,

are sufﬁciently unique for individuals under different

device usage contexts [1,2]. Once the correct device-usage

context is identiﬁed using state-of-the-art human activity

recognition systems, gait recordings are segmented

into individual gait cycles or ﬁxed-length frames (often

overlapping). Gait cycles are extracted by identifying

minima, maxima, zero crossing, or conducting phase

analysis. On the other hand, ﬁxed-length frames are

created using a sliding window-based mechanism. The

frames of 8–12 seconds with sliding length half have

achieved superior results [3,6]. It is worth noting that

the frames of such lengths consist of multiple gait cycles.

arXiv:2210.00615v1 [cs.CR] 2 Oct 2022

For cycle-based segmentation, a distance-based metric

is usually applied to establish a match threshold for

classifying each cycle as genuine, or an impostor [5,7].

Recently researchers have focused on the extraction of

statistical and frequency domain features from the cycles

and the application of machine learning algorithms. On

the other hand, researchers have established a machine

learning pipeline to train the authentication models for the

ﬁxed-length segmentation case. [3,6].

Authentication models trained using the ﬁxed-length

frame with a machine learning pipeline have usually

outperformed cycle-based approaches with distance

metrics, e.g., Euclidean or Manhattan distances or distance

measures such as Dynamic Time Warping (DTW) [8,9,10].

An experiment on the same dataset showed a difference of

13.6% [10] with the frame-based training achieving 92.7%

accuracy and cycle-based approach achieving 79.1%.

Recent studies with a better machine learning pipeline have

achieved error rates under 5% [3].

Because authentication systems play an essential role

in securing users’ data and privacy, it is paramount

that the authentication systems can defend themselves

from adversarial attacks. Zhao et al. [11] examined the

vulnerability of ABGait from random-vector attacks. The

performance of ABGait is generally evaluated in terms

of the number of false accepts (or false positives) and

false rejects (false negatives). The lower these numbers

are, the better the ABGait is. However, Zhao et al. [11]

pointed out that the false-positive rates of ABGait could

be misleading as they are computed entirely based on the

impostor samples that are available at the time of testing.

Interestingly, until Zhao et al. [11], researchers had

overlooked the possibility of someone attacking ABGait

with random feature vectors with the assumption that the

attacker would know the length of the feature vectors that

are used to train the ABGait. Zhao et al. [11] showed that

ABGait accepts even uniform random inputs. Additionally,

they found that the probability of random vector acceptance

is much higher than the false-positive rates. They suggested

that the acceptance and rejection regions created during the

ABGait training are much bigger than the false accept and

false reject rates. Therefore the performance evaluation

of ABGait should include both measures, i.e., False

Acceptance Rates (FAR) and Acceptance Region (AR), to

assess the ability of ABGait to defend itself from active

adversarial attempts such as random-vector attacks [11].

1.3. Adversarial Scenarios

Its commonly believed that behavioral biometrics such

as ABGait require more effort to circumvent than physical

biometrics. However, some studies have pointed out that

ABGait is vulnerable to treadmill-assisted imitation attacks

[9] and random-vector attacks [11]. These two attack

paradigms are substantially different. The former requires

minimal system knowledge, is difﬁcult to launch, and

requires artifacts such as a treadmill. In contrast, the latter

merely requires the generation of random vectors. However,

it also assumes that the attacker would have access to the

authentication API and the knowledge of feature space

(length of feature vectors, normalization methods, etc.)

[11]. Considering the feasibility and ease with which a

random vector attack can be executed, we focus on the same

in this study.

1.4. Possible Countermeasures

The possible countermeasures to the random-vector

attack include training the authentication models using

synthetic data. Zhao et al. [11] used synthetic

noise generated around genuine samples to train the

authentication models. The generated noise around the

genuine samples was labeled as impostors during the

training process to increase the impostor samples’ variance.

The model showed greater resilience to random-vector

attacks than the models without synthetic data-based

training. Quantitatively, the models achieved a reduced

area of the acceptance region, representing a potentially

errant input as genuine. The idea worked on the

dataset [12] that was studied in [11] but needs to be

tested on multiple datasets, including ones collected from

smartwatches. Inspired by the previous countermeasure, we

propose using Conditional Tabular Generative Adversarial

Networks (CTGAN) [13] on impostor samples to increase

the variance of the impostor samples and test its usefulness

in mitigating the random-vector attack while maintaining

the reported performance. GAN has been successfully used

to safeguard gait-based key generation from vision-based

side-channel attacks in the past [14].

1.5. Main Contributions

First, we implemented vanilla ABGait (vABGait) with

no mitigation technique in the pipeline. Then we tested the

same on three different datasets consisting of a different

number of users, samples per user, and feature set and

computed the performance measures, i.e., False Accept

Rate (FAR), False Reject Rate (FRR), Acceptance Region

(AR), and Half Total Error Rate (HTER) under the

zero-effort and random-vector attack scenarios. Second, we

included the mitigation technique proposed in Zhao et al.

[11] which resulted in (βABGait), and evaluated the same

on all three datasets using the aforementioned metrics under

the zero-effort and random-vector attack scenarios. Third,

we introduced iCTGAN, a CTGAN-assisted impostor

samples generator, in the ABGait training pipeline,

evaluated its performance on the three datasets, and

compared its performance with βABGait and vABGait.

The pre-processed dataset and code is available at [15].

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：

iCTGANAnAttackMitigationTechniqueforRandom-vectorAttackonAccelerometer-basedGaitAuthenticationSystemsJunHyungMoHaverfordCollege,USAjunmo318@gmail.comRajeshKumarBucknellUniversity,USArajesh.kumar@bucknell.eduAbstractArecentstudyshowedthatcommonly(vanilla)studiedimplementationsofaccelerometer-basedga...

展开>> 收起<<

iCTGANAn Attack Mitigation Technique for Random-vector Attack on Accelerometer-based Gait Authentication Systems Jun Hyung Mo.pdf

共9页,预览2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！

iCTGANAn Attack Mitigation Technique for Random-vector Attack on Accelerometer-based Gait Authentication Systems Jun Hyung Mo

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: