Privacy Leakage Through AI-mediated Analysis of Smartphone Data
Sarah Radway, Zoe Robert, Matthew Soto, Julianna Cimillo, Sebastian Diaz, Meg Marco, James Mickens
Rethink permission dialogs. Don't ask 'Can this app access your photos?'—ask 'Can this app analyze your photos to infer demographics, interests, and behaviors?' The consent model is broken for the ML era.
Apps request photo or calendar access, but users don't realize they're granting access to ML-powered inferences—not just raw bytes. The permission model treats data as static files, not inference fuel.
Method: Priva-See, an LLM-based inference system, analyzed data from 465 participants' phones and made privacy-invasive inferences despite accessing only a subset of user data. The system reflects how real adtech companies would leverage ML to build user profiles. After seeing what Priva-See inferred, participants significantly changed their willingness to grant app permissions.
Caveats: Study showed participants inferences from their own data but didn't measure long-term permission behavior changes.
Reflections: How should OS-level permission systems communicate inference capabilities versus raw data access? · Would showing users sample inferences before granting permissions meaningfully change consent patterns? · Can technical controls limit ML inference on permitted data without breaking legitimate app functionality?