1. Introduction
With the rapid growth of the devices, activities, services and information, the enormous amount of private and personal information that is stored has also increased. Therefore, users are becoming increasingly concerned about their personal information, how it is used, by whom and where it is stored [1]. For instance, a Consumer Report found that 92% of British and US Internet users are concerned about their privacy online [2]. When users became aware of online privacy issues, they were asked what made them most worried about their online privacy: 45% of British Internet users stated that it is personal information being shared between companies [3]. In addition, 89% of users avoided these companies because they believed that companies do not protect their privacy. It was also found that 76% of Internet users limited their online activity in the last 12 months due to these concerns [4]. This evidence indicates that users are sufficiently worried about their online privacy.
Due to the concerns of the users about privacy protection, most mobile operating systems such as Android and iOS provide some privacy safeguards for users [4]. Despite these provisions, it is unrealistic to manage and control a large volume of data. Prior study shows that users, on average, have to make over one hundred permission decisions [5]. This approach could significantly increase the burden on and frustration of the user.
A noticed focus has been given to the development of policies, procedures and tools that aid an end-user in managing and understanding their privacy-related information. However, these approaches assume that users can correctly configure all resulting settings and they have uniform privacy requirements. In reality, users do have different privacy concerns and requirements as they have heterogeneous privacy attitudes and expectations [6]. For example, some users consider personal information such as age, address and gender in their profile on a social network being more sensitive than others [7]. Furthermore, in practice it is unrealistic to assume homogeneous privacy requirements across a whole population [8].
In order to overcome the burden related to controlling a large number of privacy settings, prior study suggests that—despite users’ privacy preferences are very diverse, it is possible to assign users to the privacy profiles that most closely capture their privacy preferences [9] [10]. Despite these approaches, shows the diversity of users’ privacy preferences and cluster users into different profiles, they do not show how to assign users to the most closely profile using machine learning by asking the user a small number of questions.
Accordingly, there is a need for a solution in order to assign the user to the privacy profiles that most closely capture their privacy preferences. This study is an extension for the prior research which shows users are diverse and it is possible to divide users into small groups according to the users’ privacy preferences [11]. The result indicates the optimal number of clusters based on the k-means method is four clusters. However, the prior study does not show how to predict the user’s mobile app privacy preferences by asking the user a small number of questions. Applying privacy profiles as default settings for initial interfaces could significantly reduce the burden and frustration of the user.
This paper is organised into five sections. Section 2 presents an analysis of background literature. Section 3 shows how the data collected. Section 4 described the machine learning approach that applied in order to predict the user’s mobile app privacy preferences. The conclusions are presented in Section 5.
2. Background Literature
Numerous techniques have been proposed to monitor personal information [6] [12] [13]. The majority of existing techniques has focused upon the technical aspect to protect the privacy of users. They have shown that is possible to monitor sensitive information for users in real time. Some of the tools used a dynamic approach to monitor personal information for users. Whilst, a few studies used a network approach to detect information leakage in the mobile as shown in Table 1.
![]()
Table 1. A review of monitoring and privacy controls.
A number of research prototypes have not only monitored sensitive information for users but also provided user control over the personal information such as AntMonitor [14], ProtectMyPrivacy [6] and TISSA [15]. However, most current privacy controls support only binary and static privacy controls. A few studies such as TISSA [15] and AppFence [16] provided users with multiple levels of control. TISSA provides users with empty or bogus options for personal information that may be requested by the app. Whilst, AppFence provides users with two privacy controls to protect sensitive resources: shadowing and blocking. However, the tools do not allow users to limit the disclosure of their private information in multiple levels taking factors like the level of user’s knowledge to make the right choice in order to reduce the burden on users.
A few studies such as [20] [21] and [22] used machine learning to predict user preferences. Tsai et al. designed TurtleGuard that automatically make privacy decisions on behalf of the user. Olejnik et al. also designed a system that predicts permission decisions at runtime. These studies have shown that is possible to predict user’s preferences. In order to design initial interface, Lin et al divided users into a small number of privacy profiles, which collectively go a long way in capturing the diverse preferences of the entire population [9]. However, they do not elicit user’s privacy preferences in a context where they are not just about the permissions requested by an app but also about current knowledge and their desires to control different aspects of privacy and good usability design to maximise use.
However, the aforementioned privacy solutions also assume that users are the same in the context of how to use the privacy system and how to control the large volume of personal information. Whilst the current research and available literature have highlighted differences between the expert and novice knowledge in the context of using the system and the knowledge in the domain [23] [24]. Therefore, it is difficult for the novice user to configure a lot of settings correctly while the expert user has the ability to manage it because he has knowledge in the domain.
3. Data Collection
Participants were recruited through different platforms such as Email, social network, and some communities’ centre. Prior to displaying the survey questions, its aims and structure were briefed confirming that the respondents should be 18 years or older and they are free to withdraw up until the final submission of their responses. In total, 407 completed responses and the total responses are within the range of other surveys in the research domain and close to the expected and targeted figure.
Demographic information was collected including questions related to gender, age, education, and the level of knowledge in order to analyse the data, though the age ratio or any other demographic composition of the participants were not specifically controlled. Among these participants, 30% of them were male; 70 of them were female. Regarding the age, almost half of the participants were between 25 and 34 which represent 47% of participants. The second large age group was between 35 - 44 which represent 35%. The vast majority of the population of the participants were aged between 24 - 44 years old.
This section examines how to predict many of a user’s mobile app privacy preferences which could significantly reduce user burden with minimum questions. Accordingly, the machine learning technique was utilised to assign users to the privacy profiles that most closely capture their privacy preferences.
4.1. Interpreting the Resulting Privacy Profiles
As mentioned in the previous section, this paper is an extension of a previous study, which divided users into different profiles by utilising k-means method to determine the number of profiles. However, in order to determine the optimal number of clusters in the dataset, three methods were performed: statistical testing methods, visual exploration, and precision of predicting users’ preferences, which is already performed in our prior work [11].
· Statistical testing methods such as the Gap statistic to compare the total within intra-cluster variation for different values of k.
· Visual exploration relies on how the cluster is easy to interpret and meaningful.
· Precision of predicting users’ preferences: multi classifications were performed to determine the optimal number of clusters. When the classifier has a high prediction of determining each user with their cluster and has high accuracy, this indicates that the number of clusters is good to utilize. However, the result shows that the optimal number of clusters is four clusters.
The red chart in Figure 1 indicates a higher level of concern while the green chart indicates a lower level of concern. Cluster 2 represents the conservative group. However, it is clear from Figure 1 that cluster 2 is the largest cluster. Whilst the green chart indicates a lower level of concern who comfort to disclose their data (indicate of unconcerned) and represents 26% of users. The rest of the participants in the remaining clusters such as cluster 4 and 3 are covered in orange or yellow colures which indicates participants are moderately concerned or somewhat to share privacy-related information to the apps.
When the centroid of each cluster was computed by averaging the feature vectors of instances within the cluster, cluster 4 and 3 indicate somewhat concerned (cluster 4: μ = 2, cluster 3 μ = 3). Whilst the centroid of cluster 2 and 1 were μ = 1 and μ = 4 respectively. The characteristics of the clusters reveal a significant difference between each privacy profile, which means each cluster, has a unique profile.
4.2. Feature Selection
The machine learning approach is applied in order to assign users to these four privacy profiles. R program supports different machine algorithms such as Support Vector Machines (SVM), Random Forest (RF) and K-Nearest Neighbors (kNN). Before using classifier, it is important to determine which questions are the most important to ask users in order to minimise the questions. Therefore, the following techniques were used to minimise the 46 questions and help the models perform better and efficiently as shown in Figure 2.
Machine learning assigns a weight to each question about privacy-related information. Hence, it could be easy to minimise the number of questions, which in turn could actually help reduce user burden. Table 2 shows the features ranked according to the ranking algorithm with their weight. “Gini index” was performed to assign a score and rank the questions. These scores which are denoted as “Mean Decrease Gini” by the importance measure indicate how much each question contributes to the homogeneity in the privacy-related-information. After ranking the questions, a top ten questions were selected in order to minimise the number of questions.
![]()
Figure 1. The percentage of each cluster.
![]()
Table 2. The 10 most important questions.
4.3. Accuracy of Predictions
This section aims to create some models of the data and estimate their accuracy. However, there are different algorithms therefore; the following steps were performing in order to determine the best algorithm:
1) Set-up the test harness to use 10-fold cross-validation.
2) Build five different algorithms to predict the user’s mobile app privacy preferences.
3) Select the best algorithm.
Ten folds cross-validations were performed to estimate accuracy. This divides the data into 10 groups, train in 9 and test on 1 and release for all combinations of train-test splits. The process was repeated three times for each algorithm with different splits of the data into 10 groups; in an effort to get a more accurate estimate to predict the user’s mobile app privacy preferences. Five different machine-learning algorithms were evaluated in order to determine the best algorithm would be good on this problem. The following list shows the five algorithms: Feature Selection Methods
· Linear Discriminant Analysis (LDA).
· Classification and Regression Trees (CART).
· K-Nearest Neighbors (kNN).
· Support Vector Machines (SVM) with a linear kernel.
· Random Forest (RF).
Figure 3 shows the evaluation results of each algorithm and compare the spread and the mean accuracy of each model. It is clear from Figure 3 that the most accurate model, in this case, was SVM. SVM classifiers achieve the highest accuracy 86% with 10 questions.
Overall accuracy for 46 questions and 4 clusters is 99%. When asking users to answer ten of questions related to privacy decisions, the accuracy decreases to 86%. This result reflects the exploration of tradeoffs between accuracy and the number of questions—in other words, tradeoffs between accuracy and user burden. The result also decreases by 80% (46 to 10 questions) of the user’s effort.
5. Conclusion
The outcomes of this research show that it could significantly reduce the burden and frustration of the user while allowing users to better control information. In particular, the classifiers could be built to assign users to the privacy profiles that most closely capture their privacy preferences. In order to reduce user burden in terms of decisions, machine learning assigns a weight to each question about privacy-related information. Hence, it could be easy to minimise the number of questions, which in turn could actually help reduce user burden. Five different machine-learning algorithms were evaluated in order to determine the best algorithm. SVM algorithm achieved the highest accuracy, which is 86% for just 10 questions. The research was particularly encouraging as they offer the prospect of remarkably alleviating user burden.