Privacy-Aware Distributed Mobility Choice Modelling over Blockchain

TL;DR

This paper introduces a privacy-preserving distributed mobility choice modeling framework utilizing blockchain technology, enabling secure, local computations without sharing raw data, demonstrated through a mode choice case study.

Contribution

It presents a novel blockchain-based system for distributed mobility modeling that maintains privacy and data security during parameter estimation.

Findings

Model parameters are consistent across distributed computations.

The system ensures privacy without sacrificing accuracy.

Distributed simulated annealing effectively estimates model parameters.

Abstract

A generalized distributed tool for mobility choice modelling is presented, where participants do not share personal raw data, while all computations are done locally. Participants use Blockchain based Smart Mobility Data-market (BSMD), where all transactions are secure and private. Nodes in blockchain can transact information with other participants as long as both parties agree to the transaction rules issued by the owner of the data. A case study is presented where a mode choice model is distributed and estimated over BSMD. As an example, the parameter estimation problem is solved on a distributed version of simulated annealing. It is demonstrated that the estimated model parameters are consistent and reproducible.

Figures

Tables

Table 1. Table 1: Alternative Specific Variables and Constants

Parameters	Value	std. err.
${\beta }_{tr}$	0 (ref.)	0 (ref.)
${\beta }_a$	0.3444	0.0012
${\beta }_c$	-0.0062	0.000016
${\beta }_t$	-0.0008	0.000003
Model statistics
Null likehood	-257.5839
Final likehood	-166.3163
rho square	0.3543

Table 2. Table 2: Average latency of sending and getting a message from the BSMD

Nodes	Avg. latency of sending message	Avg. latency of getting message
chief	0.149 ($\sigma$ = 0.031)	0.043 ($\sigma$ = 0.039)
worker 1	0.036 ($\sigma$ = 0.020)	0.041 ($\sigma$ = 0.028)
worker 2	0.037 ($\sigma$ = 0.029)	0.042 ($\sigma$ = 0.025)
worker 3	0.036 ($\sigma$ = 0.033)	0.043 ($\sigma$ = 0.021)
worker 4	0.037 ($\sigma$ = 0.026)	0.041 ($\sigma$ = 0.022)

Equations

$$\underset{\vec{\beta }}{\max }\sum _{n=1}^N\ln ((y_{a,n}P(a_n;\vec{\beta })+y_{tr,n}P(t{r}_n;\vec{\beta })))$$

$$y_{a,n}$$

$$y_{tr,n}$$

$$\vec{\beta }$$

$$P(a_n;\vec{\beta })$$

$$P(t{r}_n;\vec{\beta })$$

$$V_{\vec{\beta },a_n}$$

Full text

1 Introduction

With the advances in ubiquitous networks, nowadays it is possible to obtain detailed information on individual person’s behaviour by using smartphones, cellphone towers, Wi-Fi hot-spots or traffic sensors [1]. All this information could potentially boost the accuracy of prediction tools—especially in the field of mobility choices modelling. Nowadays, data sources are vast and literally every individual with a smartphone is a potential data source.

Protecting the privacy of individuals is becoming more and more important. Privacy-preserving is a major challenge when sharing information to third parties. While some researchers or agencies may use private information in a responsible way others could misuse the private information for shady purposes [2]. On the other hand analyzing private information from a large scale population can bring great benefits to humankind as it can help providing better understanding on health care [3], mobility [4], and other fields [5].

With the use of distributed ledger technologies like blockchain, it is possible to create secure networks. In broad terms, blockchain is distributed ledger where participants can write or query their contents in a transparent manner. The first iteration of blockchain was specifically designed to make the participant anonymous [6]. However, when dealing with people sharing information it is desirable that not all participants of the blockchain are anonymous. In order to build trust with people, they may need to know with whom the information is been shared. Furthermore, there has to be some level of control over the nodes in terms of who can participate in consensus mechanics, so they cannot accumulate enough computational power to tamper the blockchain [7]. With these issues in mind a multi-layered blockchain framework over the public closed Blockchain for Smart Mobility Data-market (BSMD) was developed by [8], where participants own and shares their mobility data while their privacy is preserved. The extensive details of the framework can be found in [9].

Choice (e.g. mode, route, start-time) modelling is an important aspect of mobility demand analysis where detailed disaggregate data is collected from individuals and used to develop parametric models inspired from microeconomics and econometric. The individuals have no control on the data they provided and anyone having access to the collected data can use their information without their knowledge. Using the basic principles of distributed ledger technologies and BSMD framework, we propose an alternative for mobility choice modelling, which is able to estimate parameters on distributed data from different sources where:

Personal raw information (e.g. age, gender, income, etc.) are never shared 2. 2.

Computations are always done locally 3. 3.

No central cloud servers are required to store the information

Users are always in control of their information as no raw personal information is shared and all the information transactions are recorded on the blockchain. Thus the users can track where their information is used.

In this paper a choice modelling tool distributed over the BSMD is proposed. The main objective is to develop an environment where participant can run choice models using personal information, but where personal raw data is never shared. The rest of the paper is organized as follows: it is first introduced the background on how blockchain has been used for distributing process and databases while maintaining the privacy of individuals. We develop a methodology for creating distributed models and share results over the blockchain. A case study is presented on a stated preference travel survey with mode choice information. Privacy and security aspects are discussed. In the end, concluding remarks and future work are presented.

2 Related work

Mobility choice modelling is been an active field in the past decades and it has evolved from standard logit or probit models [10] to more detailed models that can account for heterogeneity and dynamics [11]. To mention a few recent examples, mode choice was studied from the perspective cultural context [12], social influence [13] and crime perception [14].

The common denominator of choice modelling is that the data that feeds the models is centralized which raises privacy concerns. Also, accessing big amounts of data from heterogeneous groups is always desirable for creating richer models, but this is often expensive and out of reach for entities like non-for-profit organizations. As of today, to the best of our knowledge the privacy and cost problems have not been tackled specifically in the field of choice modelling.

In recent years, blockchain networks have gained attention because of their ability to transact assets securely and privately. This has resulted in blockchain applications that are outside the world of cryptocurrencies. In recent developments, blockchain has been used for distributed databases [15] and crowdsourcing [16], where participants can query or write the ledger. But also blockchain have been used to distribute process while preserving the privacy and security of individuals, distributed applications over the blockchain have been created for IoT applications [17], smart grids management [18] and artificial intelligence  [19].

In this paper we explore how choice modelling can be distributed over a blockchain, while privacy is maintained and raw personal data do not necessarily have to be shared. In particular, we tested our proposed methodology for the maximum likelihood based parameter estimation of choice model using a distributed version of simulated annealing on BSMD framework. The objective of the paper is to show how to distribute a process over the BSMD where the personal information used to feed the models is always in control of the individual who generate such information.

3 Distributed behavioural choice modelling over the BSMD

This study developed a distributed behavioural choice modelling framework over BSMD proposed in [8, 9]. The BSMD is a public blockchain meaning that the blockchain ledgers is open for consultation but a special permission is need it participation in consensus mechanisms [20].

Nodes in the BSMD owns a digital identification (id) where participant locally store a wide range of personal information such as sensing mobile data, state preferences or even mathematical models. Node may opt to add in id and identity key which issued by trusted agency (e.g. government or non-profit organization). This key is used by other nodes to verify claims, for instance, if a node claims that they are over 18years old, other nodes can use their identity key and asks the issuer if the node is telling the truth.

To self-enforce fair trade in transactions of information smart contracts are activated before all transactions. In these the participants set the terms for sharing information. In a distributed process the participants need to set at least the following terms: (1) temporality (how long the participant are willing to participate in the distributed process, (2) idle (nodes may opt to participate only when his device is idle) and (3) permissions (nodes could select to publicly share information in the blockchain).

When a node wants to start a distribute process for a behavioural choice model, it must first create a domain and announce it over the blockchain so that participant nodes can join. In the announcement the node preset itself as the chief and clarifies the type of observations it needs to estimate the choice model and the incentives for participating in the model estimation process. Once a node joins the domain it is registered as a worker node and the chief node can solicit its services for the period established in the terms of the smart contract.

The distributed process for choice modelling is depicted in Fig. 1 and it starts when the chief sends a connection request to each worker. The request includes terms defined by the chief and worker that will be used in the smart contract. If the terms of all involved parties do not contradict each other, a peer-to-peer channel of communication is opened between the chief and each worker for sharing information. The chief node then publishes in the ledger a state preference (SP) survey and an objective function, e.g. log-likelihood function in case of max-likelihood method (see step 1 in Fig. 1). The workers answer the survey and keep their responses on their devices. In the following step, worker nodes uses their responses to evaluate the objective function and send back their results to the chief (see step 2 in Fig. 1).

It is important to note that workers nodes send the evaluated function to the chief using a peer-to-peer connection where the information is encrypted. Although the transaction details such as sender, receiver, date, etc are written in the ledger the evaluation value of the objective function is not–publishing such values in the ledger may result in the possibility to reverse-engineer the evaluation value and infer the SP survey response on an individual.

Once the chief collects and merge all the objective function values, if necessary, it may send the new model to the workers for another round and the procedure starts again (see step 3 in Fig. 1) until the results are optimized. The chief sends models using the peer-to-peer connection and it may write the model in the ledger if it wants to do so (see step 4 in Fig. 1). Since objective functions from single participants are merged it is harder to extract personal information from the model.

The use of peer-to-peer connections for data transactions in the BSMD ensures the protection of the privacy of individuals, since some information is privacy-sensitive and it cannot be published on the ledger. Furthermore, the amount of information that can be contained in a single transactions are limited by the size of the blocks forming the blockchain. Hence, if the models or evaluations are too complex and are of large size, it is better to directly send the information. If nodes want to publish complex models they can use cloud serves and publish in the ledger the address of their optimized models. In the next section, we demonstrate a distributed version of simulated annealing (a parameter estimation method) over the BSMD to evaluate behavioural choice modelling.

4 Case Study

The distributed behavioural choice modelling over BSMD is demonstrated using the maximum likelihood method to estimate the parameters for a mode choice model over a stated preference (SP) choice survey by [21]. For model evaluation a distributed version of simulated annealing [22] is used, however any other parameter estimation algorithm can be implemented with minimum effort. The SP survey was collected in 2017 consisting of commuters along the rail corridor between Montréal in Canada and NYC, Maine, and Boston in USA. In the SP choice survey, each respondent was presented with up to 6 alternatives: automobile, automobile rental, bus, plane, train, and train-hotel, where each mode alternative is characterized with trip duration, trip reliability and trip cost. For testing the distributed choice model we use a subset of the SP survey consisting in $N=246$ observations, only including automobile and train as the two mode choices.

The choice model of user preferences between train ($tr$) and automobile ($a$) is computed with the likelihood function associated with the binary logit model as shown in Equation 1. The parameters $\beta_{a}$, $\beta_{c}$ and $\beta_{t}$ are associated with average automobile-preference, cost and travel time respectively. In Equation 4 the variables $c_{a,n}$ and $c_{tr,n}$ are associated to automobile and train costs of the observation $n$ while $t_{a,n}$ and $t_{tr,n}$ are associated to automobile and train travel times of the observation $n$.

[TABLE]

where,

[TABLE]

[TABLE]

The behavioural choice model of Equation 1 is obtained by using a distributed simulated annealing (dSA) algorithm over one chief and four worker nodes. Algorithm 1 and 2 show the pseudo code of the dSA hosted in the chief and workers nodes respectively. The dSA algorithm starts when the chief distributes the $\vec{\beta}$ parameters among the worker nodes (see lines 2–4 in Algorithm 1). Then each worker node use his personal observations and $\vec{\beta}$ to compute his personal choice, $l_{w}$, and send $l_{w}$ to the chief node (see Algorithm 2). Next, the chief collects and summarize all $l_{w}$’s from each worker to compute a global choice $l$ (see line 5 in Algorithm 1).

Once the initial parameters are set the chief node starts to change the “temperature” for obtaining optimal values in the choice model (lines 10–25 in Algorithm 1). First, it creates a new parameter $\vec{\beta}_{new}$ which is computed as the sum of $\vec{\beta}$ plus a small random error (line 13). Then, it sends $\vec{\beta}_{new}$ to all workers so each one can compute and send back the chief their own $l_{w}$ (lines 14–16). In the following step, the chief collects and summarizes all $l_{w}$’s and “anneals” the new cost $l_{new}$ with the previous cost $l$ (lines 17–18). If the new cost passes the annealing conditions $\vec{\beta}$, $l$ are updated (lines 20–21), else the parameters are not updated. After a 1000 rounds of “annealing”, the chief change the “temperature” (line 25). When the temperature stabilizes, i.e., $temp\leq temp_{min}$ the optimum cost $l$ and $\vec{\beta}$ parameters are obtained.

4.1 System setup

We propose the following setup for testing a maximum likelihood approach based choice model over BSMD framework. The setup has one chief, four worker nodes and four additional nodes to run the blockchain and participate in the consensus mechanism. We assume that the chief node has enough computing power to run complex computations, while worker nodes are individuals that run evaluation of the model on their mobile devices using their personal observations only. Complete instructions for installing the setup can be found at https://github.com/LITrans/saBSMD. For the case study we use the following configuration:

1. a.

5 t3.medium111Amazon cloud EC2 virtual machines with 2cores at 3.1GHz and 4GB of RAM cloud servers. One server simulates the chief node and the remaining four are nodes running the blockchain 2. b.

4 Raspberry Pi (RPi) model 3B2224cores at 1.2GHz and 1GB RAM using WiFi and running one worker node each. The RPIs represent users participating in the distributed choice model using their smartphones.

The chief node runs a script similar to Algorithm 1 while each RPI runs a script similar to Algorithm 1. For the sake of the study case we randomly divide the $246$ observations of [21] as follows; three worker nodes have $61$ observations and the last one has $64$.

4.2 Simulation results

The dSA algorithm over the BSMD end in about 2hrs and the results are shown in Table 1. For verification purposes we solve the likelihood based choice model using a non-distributed version of simulated annealing, obtaining the same results as in Table 1.

Table 2 shows the average latency of sending and receiving messages on BSMD for the dSA based estimation process. The latency of worker nodes remains under $0.05$ secs, thus the response times are promising for implementing more complex distributed choice models over BSMD. In this particular case, the computational resources needed for running the model on worker nodes are minimal. Hence, in theory mobile phone usage is not affected when running the model. However, considering that a node will send and receive approximately $200,000$ messages in the distributed annealing process and that the size of each message is around $400$ Bytes, a total of $80MB$ of information worker nodes will send. It is thus recommended that devices use an unmetered network when participating in the distributed annealing process. Fine tuning of the algorithms (e.g. batch evaluation of a set of possible parameters) are also required to minimize the number of messages communicated and iterations needed for conversion.

The average latency of sending a massage from chief node is $0.149$ secs, a bit higher than the workers. The increment is due to chief sending the information to the four workers in batches. This could be a problem in a real-life scenario where thousands of workers nodes participate, so further exploration is needed regarding this issue. One solution that can be explored is to use multiple chief nodes coordinating the distribution estimation process.

5 Discussion

In blockchain developments for cryptocurrencies, incentives are used to convince users into participating in the consensus mechanisms, host the network, and trade assets. Hence, blockchain is a perfect platform for the users to control and sell their information, and at the same time companies will benefit from the blockchain as they may be able to exploit a massive market for buying information. In particular the BSMD platform is public so everyone can query the ledger and find with whom they are sharing information. Ability to tracking personal data usage brings transparency and a level of confidence that user’s data is used for the purpose they agree upon. In the case where a node misuse or leak information, the public characteristics of the BSMD will help authorities to easily find all affected parties by the untrusted node and take legal actions.

Identity keys in BSMD create a system of trust, when users participate in the distributed estimation of mobility models. With the identity keys, chief nodes can verify the age, gender or other valuable information of workers nodes. At the same time workers nodes can verify if the chief is a reputable or a real organization. Verification minimizes the risk of sharing data with untrusted nodes that could lead to develop poor mobility models or prevent misbehaving nodes scam individuals to participate in fake distributed models. As we mention in Section 3 the smart contracts are scripts that self-enforce a fair trade in all transactions, these scripts are embedded within the blockchain so it is virtually impossible to alter its contents without other nodes noticing the change. Identity keys and smart contracts are tools for transparency that work on both sides and are meant to secure all transactions and protect the privacy of individuals.

By not sharing personal observations our framework protects the identity of participants, while preserving the advantages large-scale data can bring into mobility models. Even if someone intercepts the evaluated objective function value and the model for each node, to the best of our knowledge, the best the attacker could do is to only approximate the personal observations. Also, all transactions are peer-to-peer and protected with asymmetric encryption so if an attacker would want to steal evaluated objective functions from various nodes they would have to tamper all connections and decrypt all keys from all users, which may not be worth the effort.

6 Conclusions and future work

Protecting individual’s privacy is an important issue that has to be tackled if researchers, governments and companies want to fully exploit the potential of large-scale personal data for developing mobility behaviour models. Distributing models and algorithms over BSMD framework is one solution for preserving the identity of individuals in such models. In this paper we present a distributed tool over BSMD for choice modelling where personal raw information is never shared, instead users shares evaluated objective functions given some parameters and model structure.

The proposed framework can work on mobile phones, but more advanced mobility models may consume significant computation power or bandwidth. If the market share of smart phones and/or unmetered networks is low, there exists a risk of creating biased models with the wealthiest sub-populations only. Even though in some developed cities the distributed choice modelling over the BSMD is feasible in terms of population heterogeneity for some other cities this technology is not ready to be implemented yet.

Chief nodes knows the model and have access to the decrypted worker evaluations, hence if the chiefs are honest but curious (in the best case) they may be able to infer personal observations of workers by using the evaluations. In future, we will consider the use differential privacy techniques for adding noise to the data the workers sends and in this manner preventing the chief node to infer the data of the workers.

In future developments we will implement the distributions of model over the BSMD using more powerful data-driven prediction tools like Federated Learning [23] and also adapt advanced Machine Learning algorithms for behavioural models like the one presented in [24]. Machine Learning is gaining momentum and in combination with blockchain there is a strong potential to develop powerful distributed intelligence, where users can privately and securely train models to create comprehensive prediction tools.