[Paper Review] On the Unicity of Smartphone Applications

Jagdish Prasad Achara, Gergely Ács+1 more|arXiv (Cornell University)|Jul 28, 2015

Privacy, Security, and Data Protection12 references4 citations

TL;DR

This paper investigates the re-identifiability risk posed by smartphone app lists, demonstrating that as few as four installed apps can uniquely identify 95% of users in a dataset of 54,893 Android users. Using Markov Chain Monte Carlo sampling and non-linear regression, the authors show that app lists are highly unique, challenging anonymization and highlighting urgent privacy risks in data sharing, especially when combined with third-party tracking libraries.

ABSTRACT

Prior works have shown that the list of apps installed by a user reveal a lot about user interests and behavior. These works rely on the semantics of the installed apps and show that various user traits could be learnt automatically using off-the-shelf machine-learning techniques. In this work, we focus on the re-identifiability issue and thoroughly study the unicity of smartphone apps on a dataset containing 54,893 Android users collected over a period of 7 months. Our study finds that any 4 apps installed by a user are enough (more than 95% times) for the re-identification of the user in our dataset. As the complete list of installed apps is unique for 99% of the users in our dataset, it can be easily used to track/profile the users by a service such as Twitter that has access to the whole list of installed apps of users. As our analyzed dataset is small as compared to the total population of Android users, we also study how unicity would vary with larger datasets. This work emphasizes the need of better privacy guards against collection, use and release of the list of installed apps.

Motivation & Objective

• To assess the unicity of smartphone app lists and their re-identifiability risk in user datasets.
• To evaluate the privacy implications of releasing app lists, even in pseudonymized form, by measuring uniqueness across subsets of apps.
• To develop a method for unbiased estimation of unicity for arbitrary subsets of apps using MCMC sampling.
• To predict unicity in larger populations using non-linear regression, based on limited dataset samples.
• To evaluate the generalizability of findings to larger-scale datasets, such as global Android user populations.

Proposed method

• The authors use a dataset of 54,893 Android users with app lists collected over 7 months, with apps identified by hashed names to preserve privacy.
• They apply a Markov Chain Monte Carlo (MCMC) method to uniformly sample subsets of apps and estimate the probability that a subset of size K is unique across the dataset.
• The MCMC chain is proven to be fast-mixing, ensuring efficient and accurate estimation of unicity with time complexity roughly linear in dataset size and K.
• The authors use non-linear regression to model the relationship between dataset size and unicity, learning a predictive function for unicity in larger populations.
• They validate the model on a separate mobility dataset of 1 million users, training on progressively larger subsets to assess predictive accuracy.

(a) $K=5$

Experimental results

Research questions

• RQ1 What is the minimum number of installed apps required to uniquely identify a user in a smartphone app dataset?
• RQ2 How does the uniqueness of app lists scale with dataset size, and can we predict unicity in larger populations from smaller samples?
• RQ3 To what extent does pseudonymization of app lists fail to protect user privacy due to inherent unicity?
• RQ4 Can MCMC sampling provide an unbiased estimate of unicity for arbitrary subsets of apps in a dataset?
• RQ5 How well can a learned model based on small datasets predict unicity in much larger real-world populations?

Key findings

• 99% of users in the dataset have a unique list of installed apps, indicating that the full app list is highly distinctive.
• Knowing just four apps allows re-identification of a user with 95% probability, demonstrating strong re-identifiability risk.
• With only two apps, an adversary can identify a user with 75% probability, highlighting the risk even with minimal information.
• The MCMC-based method provides an unbiased and efficient estimate of unicity for any subset size K, with fast-mixing properties ensuring accuracy.
• The non-linear regression model predicts unicity well in larger datasets when trained on sufficiently large samples—e.g., 100,000 users—achieving an average error of 0.05 on a 1 million-user mobility dataset.
• Training the model on only 50,000 users leads to poor generalization, suggesting that small datasets may not be representative for predicting unicity in global-scale populations.

(b) $K=6$