Privacy Violation Detection for Instant Messaging Platforms
A KFAS-funded project developing machine learning techniques to dynamically learn user privacy preferences and automatically detect when sensitive image content is at risk of inappropriate disclosure in instant messaging platforms.
Motivation
Instant messaging platforms such as WhatsApp and Signal offer users a convenient way to share text, images, and video — yet users regularly send sensitive content to wrong recipients, share images that are more sensitive than intended, or fail to appreciate the privacy implications of shared media. Current platform privacy mechanisms rely on coarse-grained permissions and manual user decisions, providing limited protection against inadvertent disclosure of sensitive content.
This project develops a machine learning-based framework that dynamically learns and updates a user's privacy preferences for images. Before any image is transmitted, the system analyses it, assigns a set of semantic labels, and determines a sensitivity level. This knowledge, combined with contextual factors about the intended recipient, informs a decision framework that identifies potential privacy violations before they occur. A proof-of-concept Android implementation is developed as a system-level service that any messaging application can invoke with user permission.
A distinctive contribution is the regional focus: privacy preferences in the Gulf region and the Middle East are unique and frequently underrepresented in the privacy research literature. The project extends the image classification model with culturally specific labels relevant to users in Kuwait, enabling more accurate sensitivity assessment in regional contexts. The project is funded by the Kuwait Foundation for the Advancement of Sciences (KFAS, grant PN23-15EO-1851) for three years.