BERT-ADLOC: A secure crowdsourced indoor localization system based on BLE fingerprints

Abstract

Crowdsourced indoor localization methods have grasped much attention in recent years as a method of reducing the cost of constructing the fingerprint database. In a crowdsourcing environment, however, the localization system is vulnerable to malicious attacks, which possibly lead to serious localization errors. In this paper, we conclude the potential attacks during fingerprint database updates and online inference phases and propose a secure indoor crowdsourced localization system, BERT-ADLOC, based on BLE fingerprints. Our system consists of two main parts: adversarial sample discriminator BERT-AD and indoor localization model BERT-LOC. Our proposed BERT-AD recognizes fake fingerprints during the database update phase, while BERT-LOC defends against attacks online, in which valid beacons are moved or malicious beacons are deployed. A tailored BERT model is introduced to extract deep hidden features through the self-attention mechanism. Our experiments show that BERT-ADLOC achieves a good localization performance against adversaries both in the fingerprint database update phase and online inference phase.

Introduction

Indoor positioning has been an active research front for almost two decades. Many signal modalities – radio signal in particular – have been extensively studied and fruitful results have been published. Certain pioneering business efforts [1] have been practiced, but large-scale deployment of indoor positioning systems are not yet to come.

The challenge of indoor positioning comes from the multi-path propagation of wireless signal and the complexity of indoor environments.

There are two trends in the research field. One trend is to develop more advanced analytical solutions [2], [3], [4], [5], [6], [7], [8], [9], [10]. They either seek to mitigate the multi-path effect by aggregating more bandwidths to better identify the direct path [5], [6], [7] or leverage the multi-path by forming multiple triangulations and resort to AoA principle for positioning [8]. Another trend is along the statistical way, where deep neural network (DNN) [11], convolutional neural network (CNN) [12], deep long short-term memory (DLSTM) [13] have been attempted in the past few years. Due to the superiority in extracting the deep hidden features, these deep learning models further improved the localization accuracy. Data required by these systems can be readily collected using commodity smartphones and are thus crowd-sensing friendly.

It is well known that building up a location database is extremely laborious and expensive. The two most labor-demanding efforts are data collection and labeling, which dominate the practicality of an IPS. In [14], the Gauss Process Regression (GPR) was applied to obtain the radio map of the entire space. As advocated in [15], [16], [17], [18], the most promising way is through crowdsourcing. The efficiency of crowdsourcing is evidently related to the number and also the involvement of participants. To this end, for efficient crowdsourcing, besides good incentive mechanism design, device requirement should be low, data labeling should be easy. Ideally, little or no labeling should be required.

Although many scholars have studied indoor localization systems based on crowdsourced environments, most of these researches only focus on indoor localization solution, ignoring the attacks that may be encountered in the realistic indoor positioning scene. As a large number of machine learning-based methods are used in indoor localization, the vulnerability of these approaches should not be ignored.

We have investigated recent researches on indoor positioning security. Wang et al. performed a comprehensive investigation of the security properties of ML algorithms under adversarial settings [19]. Tiku et al. [20] analyzed the possible attacks during the maintenance and update of the fingerprint database in a crowdsourcing environment and proposed the strategy for identifying malicious fingerprints. In [21], the malicious beacons attack during the online inference phase of the localization model was studied and a solution was designed. As far as we know, we are the first to design a complete system to resist attacks in the crowdsourcing scenario. We not only propose a reliable fingerprint database update scheme, but also provide a robust localization model that can resist online attacks in the crowdsourcing environment.

Our contributions are summarized as follows:

• We introduce tailored BERT as a powerful extractor to explore the underlying hidden deeper relation among BLE beacons through self-attention mechanism.

• We design BERT-AD to recognize malicious fingerprints in the fingerprint database update phase under crowdsourcing environment.

• We present robust indoor localization model BERT-LOCwhich can resist attacks in the online positioning phase and maintain good positioning results. These attacks include two aspects where valid beacons are moved and malicious beacons are deployed.

• For the first time, We present a secure crowdsourced indoor localization system BERT-ADLOC, which provides a reliable solution for updating the fingerprint database and ensures the robustness of the online location model in the presence of online attacks. We conduct multiple experiments to evaluate our proposed BERT-ADLOC.

The paper is organized in this order: Section 2 introduces the relevant work of indoor localization methods and indoor localization security. Section 3 analyzes multiple type of attacks during fingerprint database update and inference phase. Section 4 describes our proposed safe BERT-ADLOC localization system in detail. Section 5 shows experimental results and our discussions. Section 6 summarizes this paper.