Policylint: Investigating internal privacy policy contradictions on google play

Abstract

Privacy policies are the primary mechanism by which companies inform users about data collection and sharing practices. To help users better understand these long and complex legal documents, recent research has proposed tools that summarize collection and sharing. However, these tools have a significant oversight: they do not account for contradictions that may occur within an individual policy. In this paper, we present PolicyLint, a privacy policy analysis tool that identifies such contradictions by simultaneously considering negation and varying semantic levels of data objects and entities. To do so, PolicyLint automatically generates ontologies from a large corpus of privacy policies and uses sentence-level natural language processing to capture both positive and negative statements of data collection and sharing. We use PolicyLint to analyze the policies of 11,430 apps and find that 14.2% of these policies contain contradictions that may be indicative of misleading statements. We manually verify 510 contradictions, identifying concerning trends that include the use of misleading presentation, attempted redefinition of common understandings of terms, conflicts in regulatory definitions (e.g., US and EU), and “laundering” of tracking information facilitated by sharing or collecting data that can be used to derive sensitive information. In doing so, PolicyLint significantly advances automated analysis of privacy policies.