At the beginning of the year, researchers from Leipzig University and ScaDS.AI Dresden/Leipzig were able to apply for funding through the “ScaDS.AI Early Career Innovation Projects on the Future of AI” initiative. As part of the program, selected early-career researchers took part in joint brainstorming, writing, and pitch-practice sessions. The program helped them prepare competitive research funding applications and further develop their project ideas. A jury of ScaDS.AI Principal Investigators selected the seven best projects from 71 applications. These projects will receive funding for up to six months in 2026.

Among the funded projects is Maja Schneider’s PriME-LLM: Privacy-Preserving Masking Engine for LLM Interaction. It focuses on a growing challenge in everyday AI use: people often share sensitive information with chatbots without fully realizing the privacy risks. The project develops a privacy-preserving masking engine for LLM interactions on Android smartphones. It automatically detects, sanitizes, and explains sensitive information before prompts are sent to external chatbot providers.

The project goes beyond traditional rule-based methods and named entity recognition and takes into account context-based detection of sensitive information. This includes broader categories of sensitive content, such as medical, financial, legal, social, and proprietary information. At its core is a multilayer masking engine that combines filtering techniques with a locally running SLM-based sensitivity checker.

Protecting privacy on device

PriME-LLM aims to protect user data directly on the device. The system uses small language models for on-device inference and detects sensitive content before it leaves the smartphone. It then masks relevant parts and reintegrates them into chatbot responses, so the interaction remains usable.

In addition to sanitizing prompts, the app helps users understand what kind of information they share. It visualizes sensitive content, explains potential risks, and keeps a cumulative privacy score. In this way, PriME-LLM combines privacy protection with privacy awareness.

Making privacy-aware AI practical

The project will deliver a prototype that can be used with popular LLM chatbots. Beyond personal use, the masking engine could also support other LLM-based applications and provide a basis for privacy-preserving AI systems in consumer and professional settings.

In the long term, PriME-LLM could help advance privacy-by-design AI applications, support compliance with emerging regulations such as the EU AI Act, and strengthen public trust in AI technologies.