LLM-PrivCheck : Detecting Privacy Policy Violations in Android Applications with LLMs

With the widespread use of Android applications, frequent access to personal user data (such as location information and device identifiers) has become the norm. However, inconsistencies between privacy policy statements and actual app behavior often lead to significant privacy risks. This paper proposes a consistency detection method, LLM-PrivCheck, based on large language models (LLMs), which automatically checks whether the data collection statements in privacy policies align with the actual behavior of applications. LLM-PrivCheck utilizes LLMs for privacy policy parsing and combines taint analysis techniques to track data flows, ultimately achieving consistency detection through ontology mapping. Experimental results show that in tests on 960 commercial applications from Google Play, LLM-PrivCheck detected that 96.25% of the apps had inconsistencies between their privacy policies and actual behavior.