Exploring Large Language Models’ Potential for Privacy Leakage Detection in Android App Logs: An Empirical Study
Zhiyuan Chen, Vanessa Nava-Camal, Tiash Roy, Zhe Li, Yiming Tang, Xueling Zhang, Haibo Yang · IEEE Software · 2025
Android app logs are essential resources for recording runtime information, but they may also include users’ privacy information. Existing privacy leakage detection approaches in logs are primarily based on regular expressions or keyword-based searches, which heavily rely on the comprehensiveness of the keyword list and can result in high detection inaccuracy. Therefore, we propose two research questions studying the ability of LLMs to detect privacy leakage in Android app logs. In our preliminary experiment, we used ChatGPT and found that while ChatGPT can detect privacy leakage in Android app logs, its ability can result in high inaccuracy. After improving the detection strategy by modifying the prompts to allow ChatGPT to learn the contextual knowledge of privacy leakage, it significantly improved in detecting more privacy leakage and reducing false positives. The detection results also indicate that privacy leakage in Android app logs is a non-trivial issue.